diff --git a/docs/xai/docs/Makefile b/docs/xai/docs/Makefile
new file mode 100644
index 0000000000000000000000000000000000000000..1eff8952707bdfa503c8d60c1e9a903053170ba2
--- /dev/null
+++ b/docs/xai/docs/Makefile
@@ -0,0 +1,20 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line, and also
+# from the environment for the first two.
+SPHINXOPTS    ?=
+SPHINXBUILD   ?= sphinx-build
+SOURCEDIR     = source_zh_cn
+BUILDDIR      = build_zh_cn
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
diff --git a/docs/xai/docs/_ext/overwriteobjectiondirective.txt b/docs/xai/docs/_ext/overwriteobjectiondirective.txt
new file mode 100644
index 0000000000000000000000000000000000000000..745f53083d102d2a458c5f7d8bd7daaabd63ce1d
--- /dev/null
+++ b/docs/xai/docs/_ext/overwriteobjectiondirective.txt
@@ -0,0 +1,392 @@
+import re
+import inspect
+import importlib
+from typing import Any, Dict, List, Tuple
+from typing import cast
+
+from docutils import nodes
+from docutils.nodes import Node
+from docutils.parsers.rst import directives, roles
+
+from sphinx import addnodes
+from sphinx.addnodes import desc_signature
+from sphinx.deprecation import RemovedInSphinx40Warning, deprecated_alias
+from sphinx.util import docutils, logging
+from sphinx.util.docfields import DocFieldTransformer, Field, TypedField
+from sphinx.util.docutils import SphinxDirective
+from sphinx.util.typing import DirectiveOption
+
+if False:
+    # For type annotation
+    from sphinx.application import Sphinx
+
+
+# RE to strip backslash escapes
+nl_escape_re = re.compile(r'\\\n')
+strip_backslash_re = re.compile(r'\\(.)')
+
+logger = logging.getLogger(__name__)
+
+def optional_int(argument):
+    """
+    Check for an integer argument or None value; raise ``ValueError`` if not.
+    """
+    if argument is None:
+        return None
+    else:
+        value = int(argument)
+        if value < 0:
+            raise ValueError('negative value; must be positive or zero')
+        return value
+
+def get_api(fullname):
+    try:
+        module_name, api_name= ".".join(fullname.split('.')[:-1]), fullname.split('.')[-1]
+        module_import = importlib.import_module(module_name)
+    except ModuleNotFoundError:
+        module_name, api_name = ".".join(fullname.split('.')[:-2]), ".".join(fullname.split('.')[-2:])
+        module_import = importlib.import_module(module_name)
+    api = eval(f"module_import.{api_name}")
+    return api
+
+def get_example(name: str):
+    try:
+        api_doc = inspect.getdoc(get_api(name))
+        example_str = re.findall(r'Examples:\n([\w\W]*?)(\n\n|$)', api_doc)
+        if not example_str:
+            return []
+        example_str = re.sub(r'\n\s+', r'\n', example_str[0][0])
+        example_str = example_str.strip()
+        example_list = example_str.split('\n')
+        return ["", "**样例：**", ""] + example_list + [""]
+    except:
+        return []
+
+def get_platforms(name: str):
+    try:
+        api_doc = inspect.getdoc(get_api(name))
+        example_str = re.findall(r'Supported Platforms:\n\s+(.*?)\n\n', api_doc)
+        if not example_str:
+            example_str_leak = re.findall(r'Supported Platforms:\n\s+(.*)', api_doc)
+            if example_str_leak:
+                example_str = example_str_leak[0].strip()
+                example_list = example_str.split('\n')
+                example_list = ['    ' + example_list[0]]
+                return ["", "支持平台："] + example_list + [""]
+            return []
+        example_str = example_str[0].strip()
+        example_list = example_str.split('\n')
+        example_list = ['    ' + example_list[0]]
+        return ["", "支持平台："] + example_list + [""]
+    except:
+        return []
+
+def get_side_effect(name: str):
+    try:
+        api_doc = inspect.getdoc(get_api(name))
+        side_effect_str = re.findall(r'Side Effects:\n\s+(.*?)\n\n', api_doc)
+        side_effect_str = side_effect_str[0].strip()
+        side_effect_list = side_effect_str.split('\n')
+        return ["", "**副作用：**", ""] + side_effect_list + [""]
+    except:
+        return []
+
+class ObjectDescription(SphinxDirective):
+    """
+    Directive to describe a class, function or similar object.  Not used
+    directly, but subclassed (in domain-specific directives) to add custom
+    behavior.
+    """
+
+    has_content = True
+    required_arguments = 1
+    optional_arguments = 0
+    final_argument_whitespace = True
+    option_spec = {
+        'noindex': directives.flag,
+    }  # type: Dict[str, DirectiveOption]
+
+    # types of doc fields that this directive handles, see sphinx.util.docfields
+    doc_field_types = []    # type: List[Field]
+    domain = None           # type: str
+    objtype = None          # type: str
+    indexnode = None        # type: addnodes.index
+
+    # Warning: this might be removed in future version. Don't touch this from extensions.
+    _doc_field_type_map = {}  # type: Dict[str, Tuple[Field, bool]]
+
+    def get_field_type_map(self) -> Dict[str, Tuple[Field, bool]]:
+        if self._doc_field_type_map == {}:
+            self._doc_field_type_map = {}
+            for field in self.doc_field_types:
+                for name in field.names:
+                    self._doc_field_type_map[name] = (field, False)
+
+                if field.is_typed:
+                    typed_field = cast(TypedField, field)
+                    for name in typed_field.typenames:
+                        self._doc_field_type_map[name] = (field, True)
+
+        return self._doc_field_type_map
+
+    def get_signatures(self) -> List[str]:
+        """
+        Retrieve the signatures to document from the directive arguments.  By
+        default, signatures are given as arguments, one per line.
+
+        Backslash-escaping of newlines is supported.
+        """
+        lines = nl_escape_re.sub('', self.arguments[0]).split('\n')
+        # remove backslashes to support (dummy) escapes; helps Vim highlighting
+        return [strip_backslash_re.sub(r'\1', line.strip()) for line in lines]
+
+    def handle_signature(self, sig: str, signode: desc_signature) -> Any:
+        """
+        Parse the signature *sig* into individual nodes and append them to
+        *signode*. If ValueError is raised, parsing is aborted and the whole
+        *sig* is put into a single desc_name node.
+
+        The return value should be a value that identifies the object.  It is
+        passed to :meth:`add_target_and_index()` unchanged, and otherwise only
+        used to skip duplicates.
+        """
+        raise ValueError
+
+    def add_target_and_index(self, name: Any, sig: str, signode: desc_signature) -> None:
+        """
+        Add cross-reference IDs and entries to self.indexnode, if applicable.
+
+        *name* is whatever :meth:`handle_signature()` returned.
+        """
+        return  # do nothing by default
+
+    def before_content(self) -> None:
+        """
+        Called before parsing content. Used to set information about the current
+        directive context on the build environment.
+        """
+        pass
+
+    def after_content(self) -> None:
+        """
+        Called after parsing content. Used to reset information about the
+        current directive context on the build environment.
+        """
+        pass
+
+    def check_class_end(self, content):
+        for i in content:
+            if not i.startswith('.. include::') and i != "\n" and i != "":
+                return False
+        return True
+
+    def extend_items(self, rst_file, start_num, num):
+        ls = []
+        for i in range(1, num+1):
+            ls.append((rst_file, start_num+i))
+        return ls
+
+    def run(self) -> List[Node]:
+        """
+        Main directive entry function, called by docutils upon encountering the
+        directive.
+
+        This directive is meant to be quite easily subclassable, so it delegates
+        to several additional methods.  What it does:
+
+        * find out if called as a domain-specific directive, set self.domain
+        * create a `desc` node to fit all description inside
+        * parse standard options, currently `noindex`
+        * create an index node if needed as self.indexnode
+        * parse all given signatures (as returned by self.get_signatures())
+          using self.handle_signature(), which should either return a name
+          or raise ValueError
+        * add index entries using self.add_target_and_index()
+        * parse the content and handle doc fields in it
+        """
+        if ':' in self.name:
+            self.domain, self.objtype = self.name.split(':', 1)
+        else:
+            self.domain, self.objtype = '', self.name
+        self.indexnode = addnodes.index(entries=[])
+
+        node = addnodes.desc()
+        node.document = self.state.document
+        node['domain'] = self.domain
+        # 'desctype' is a backwards compatible attribute
+        node['objtype'] = node['desctype'] = self.objtype
+        node['noindex'] = noindex = ('noindex' in self.options)
+
+        self.names = []  # type: List[Any]
+        signatures = self.get_signatures()
+        for i, sig in enumerate(signatures):
+            # add a signature node for each signature in the current unit
+            # and add a reference target for it
+            signode = addnodes.desc_signature(sig, '')
+            signode['first'] = False
+            node.append(signode)
+            try:
+                # name can also be a tuple, e.g. (classname, objname);
+                # this is strictly domain-specific (i.e. no assumptions may
+                # be made in this base class)
+                name = self.handle_signature(sig, signode)
+            except ValueError:
+                # signature parsing failed
+                signode.clear()
+                signode += addnodes.desc_name(sig, sig)
+                continue  # we don't want an index entry here
+            if name not in self.names:
+                self.names.append(name)
+                if not noindex:
+                    # only add target and index entry if this is the first
+                    # description of the object with this name in this desc block
+                    self.add_target_and_index(name, sig, signode)
+
+        contentnode = addnodes.desc_content()
+        node.append(contentnode)
+        if self.names:
+            # needed for association of version{added,changed} directives
+            self.env.temp_data['object'] = self.names[0]
+        self.before_content()
+        try:
+            example = get_example(self.names[0][0])
+            platforms = get_platforms(self.names[0][0])
+            side_effect = get_side_effect(self.names[0][0])
+        except IndexError:
+            example = ''
+            platforms = ''
+            side_effect = ''
+            logger.warning(f'Error API names in {self.arguments[0]}.')
+        extra = side_effect + platforms + example
+        if extra:
+            if self.objtype == "method":
+                self.content.data.extend(extra)
+            else:
+                index_num = 0
+                for num, i in enumerate(self.content.data):
+                    if i.startswith('.. py:method::') or self.check_class_end(self.content.data[num:]):
+                        index_num = num
+                        break
+                if index_num:
+                    count = len(self.content.data)
+                    for i in extra:
+                        self.content.data.insert(index_num-count, i)
+                else:
+                    self.content.data.extend(extra)
+        try:
+            self.content.items.extend(self.extend_items(self.content.items[0][0], self.content.items[-1][1], len(extra)))
+        except IndexError:
+            logger.warning(f'{self.names[0][0]} has error format.')
+        self.state.nested_parse(self.content, self.content_offset, contentnode)
+        self.env.app.emit('object-description-transform',
+                          self.domain, self.objtype, contentnode)
+        DocFieldTransformer(self).transform_all(contentnode)
+        self.env.temp_data['object'] = None
+        self.after_content()
+        return [self.indexnode, node]
+
+
+class DefaultRole(SphinxDirective):
+    """
+    Set the default interpreted text role.  Overridden from docutils.
+    """
+
+    optional_arguments = 1
+    final_argument_whitespace = False
+
+    def run(self) -> List[Node]:
+        if not self.arguments:
+            docutils.unregister_role('')
+            return []
+        role_name = self.arguments[0]
+        role, messages = roles.role(role_name, self.state_machine.language,
+                                    self.lineno, self.state.reporter)
+        if role:
+            docutils.register_role('', role)
+            self.env.temp_data['default_role'] = role_name
+        else:
+            literal_block = nodes.literal_block(self.block_text, self.block_text)
+            reporter = self.state.reporter
+            error = reporter.error('Unknown interpreted text role "%s".' % role_name,
+                                   literal_block, line=self.lineno)
+            messages += [error]
+
+        return cast(List[nodes.Node], messages)
+
+
+class DefaultDomain(SphinxDirective):
+    """
+    Directive to (re-)set the default domain for this source file.
+    """
+
+    has_content = False
+    required_arguments = 1
+    optional_arguments = 0
+    final_argument_whitespace = False
+    option_spec = {}  # type: Dict
+
+    def run(self) -> List[Node]:
+        domain_name = self.arguments[0].lower()
+        # if domain_name not in env.domains:
+        #     # try searching by label
+        #     for domain in env.domains.values():
+        #         if domain.label.lower() == domain_name:
+        #             domain_name = domain.name
+        #             break
+        self.env.temp_data['default_domain'] = self.env.domains.get(domain_name)
+        return []
+
+from sphinx.directives.code import (  # noqa
+    Highlight, CodeBlock, LiteralInclude
+)
+from sphinx.directives.other import (  # noqa
+    TocTree, Author, VersionChange, SeeAlso,
+    TabularColumns, Centered, Acks, HList, Only, Include, Class
+)
+from sphinx.directives.patches import (  # noqa
+    Figure, Meta
+)
+from sphinx.domains.index import IndexDirective  # noqa
+
+deprecated_alias('sphinx.directives',
+                 {
+                     'Highlight': Highlight,
+                     'CodeBlock': CodeBlock,
+                     'LiteralInclude': LiteralInclude,
+                     'TocTree': TocTree,
+                     'Author': Author,
+                     'Index': IndexDirective,
+                     'VersionChange': VersionChange,
+                     'SeeAlso': SeeAlso,
+                     'TabularColumns': TabularColumns,
+                     'Centered': Centered,
+                     'Acks': Acks,
+                     'HList': HList,
+                     'Only': Only,
+                     'Include': Include,
+                     'Class': Class,
+                     'Figure': Figure,
+                     'Meta': Meta,
+                 },
+                 RemovedInSphinx40Warning)
+
+
+# backwards compatible old name (will be marked deprecated in 3.0)
+DescDirective = ObjectDescription
+
+
+def setup(app: "Sphinx") -> Dict[str, Any]:
+    directives.register_directive('default-role', DefaultRole)
+    directives.register_directive('default-domain', DefaultDomain)
+    directives.register_directive('describe', ObjectDescription)
+    # new, more consistent, name
+    directives.register_directive('object', ObjectDescription)
+
+    app.add_event('object-description-transform')
+
+    return {
+        'version': 'builtin',
+        'parallel_read_safe': True,
+        'parallel_write_safe': True,
+    }
+
diff --git a/docs/xai/docs/_ext/overwriteviewcode.txt b/docs/xai/docs/_ext/overwriteviewcode.txt
new file mode 100644
index 0000000000000000000000000000000000000000..130defb02fa41468e2bfda3b3843314c8cd0dcce
--- /dev/null
+++ b/docs/xai/docs/_ext/overwriteviewcode.txt
@@ -0,0 +1,274 @@
+"""
+    sphinx.ext.viewcode
+    ~~~~~~~~~~~~~~~~~~~
+
+    Add links to module code in Python object descriptions.
+
+    :copyright: Copyright 2007-2020 by the Sphinx team, see AUTHORS.
+    :license: BSD, see LICENSE for details.
+"""
+
+import traceback
+import warnings
+from typing import Any, Dict, Iterable, Iterator, Set, Tuple
+
+from docutils import nodes
+from docutils.nodes import Element, Node
+
+import sphinx
+from sphinx import addnodes
+from sphinx.application import Sphinx
+from sphinx.config import Config
+from sphinx.deprecation import RemovedInSphinx30Warning
+from sphinx.environment import BuildEnvironment
+from sphinx.locale import _, __
+from sphinx.pycode import ModuleAnalyzer
+from sphinx.util import get_full_modname, logging, status_iterator
+from sphinx.util.nodes import make_refnode
+
+
+logger = logging.getLogger(__name__)
+
+
+def _get_full_modname(app: Sphinx, modname: str, attribute: str) -> str:
+    try:
+        return get_full_modname(modname, attribute)
+    except AttributeError:
+        # sphinx.ext.viewcode can't follow class instance attribute
+        # then AttributeError logging output only verbose mode.
+        logger.verbose('Didn\'t find %s in %s', attribute, modname)
+        return None
+    except Exception as e:
+        # sphinx.ext.viewcode follow python domain directives.
+        # because of that, if there are no real modules exists that specified
+        # by py:function or other directives, viewcode emits a lot of warnings.
+        # It should be displayed only verbose mode.
+        logger.verbose(traceback.format_exc().rstrip())
+        logger.verbose('viewcode can\'t import %s, failed with error "%s"', modname, e)
+        return None
+
+
+def doctree_read(app: Sphinx, doctree: Node) -> None:
+    env = app.builder.env
+    if not hasattr(env, '_viewcode_modules'):
+        env._viewcode_modules = {}  # type: ignore
+    if app.builder.name == "singlehtml":
+        return
+    if app.builder.name.startswith("epub") and not env.config.viewcode_enable_epub:
+        return
+
+    def has_tag(modname, fullname, docname, refname):
+        entry = env._viewcode_modules.get(modname, None)  # type: ignore
+        if entry is False:
+            return
+
+        code_tags = app.emit_firstresult('viewcode-find-source', modname)
+        if code_tags is None:
+            try:
+                analyzer = ModuleAnalyzer.for_module(modname)
+                analyzer.find_tags()
+            except Exception:
+                env._viewcode_modules[modname] = False  # type: ignore
+                return
+
+            code = analyzer.code
+            tags = analyzer.tags
+        else:
+            code, tags = code_tags
+
+        if entry is None or entry[0] != code:
+            entry = code, tags, {}, refname
+            env._viewcode_modules[modname] = entry  # type: ignore
+        _, tags, used, _ = entry
+        if fullname in tags:
+            used[fullname] = docname
+            return True
+
+    for objnode in doctree.traverse(addnodes.desc):
+        if objnode.get('domain') != 'py':
+            continue
+        names = set()  # type: Set[str]
+        for signode in objnode:
+            if not isinstance(signode, addnodes.desc_signature):
+                continue
+            modname = signode.get('module')
+            fullname = signode.get('fullname')
+            if fullname and modname==None:
+                if fullname.split('.')[-1].lower() == fullname.split('.')[-1] and fullname.split('.')[-2].lower() != fullname.split('.')[-2]:
+                    modname = '.'.join(fullname.split('.')[:-2])
+                    fullname = '.'.join(fullname.split('.')[-2:])
+                else:
+                    modname = '.'.join(fullname.split('.')[:-1])
+                    fullname = fullname.split('.')[-1]
+            fullname_new = fullname
+            # logger.warning(f'modename:{modname}')
+            # logger.warning(f'fullname:{fullname}')
+            refname = modname
+            if env.config.viewcode_follow_imported_members:
+                new_modname = app.emit_firstresult(
+                    'viewcode-follow-imported', modname, fullname,
+                )
+                if not new_modname:
+                    new_modname = _get_full_modname(app, modname, fullname)
+                modname = new_modname
+            # logger.warning(f'new_modename:{modname}')
+            if not modname:
+                continue
+            # fullname = signode.get('fullname')
+            # if fullname and modname==None:
+            fullname = fullname_new
+            if not has_tag(modname, fullname, env.docname, refname):
+                continue
+            if fullname in names:
+                # only one link per name, please
+                continue
+            names.add(fullname)
+            pagename = '_modules/' + modname.replace('.', '/')
+            inline = nodes.inline('', _('[源代码]'), classes=['viewcode-link'])
+            onlynode = addnodes.only(expr='html')
+            onlynode += addnodes.pending_xref('', inline, reftype='viewcode', refdomain='std',
+                                              refexplicit=False, reftarget=pagename,
+                                              refid=fullname, refdoc=env.docname)
+            signode += onlynode
+
+
+def env_merge_info(app: Sphinx, env: BuildEnvironment, docnames: Iterable[str],
+                   other: BuildEnvironment) -> None:
+    if not hasattr(other, '_viewcode_modules'):
+        return
+    # create a _viewcode_modules dict on the main environment
+    if not hasattr(env, '_viewcode_modules'):
+        env._viewcode_modules = {}  # type: ignore
+    # now merge in the information from the subprocess
+    env._viewcode_modules.update(other._viewcode_modules)  # type: ignore
+
+
+def missing_reference(app: Sphinx, env: BuildEnvironment, node: Element, contnode: Node
+                      ) -> Node:
+    # resolve our "viewcode" reference nodes -- they need special treatment
+    if node['reftype'] == 'viewcode':
+        return make_refnode(app.builder, node['refdoc'], node['reftarget'],
+                            node['refid'], contnode)
+
+    return None
+
+
+def collect_pages(app: Sphinx) -> Iterator[Tuple[str, Dict[str, Any], str]]:
+    env = app.builder.env
+    if not hasattr(env, '_viewcode_modules'):
+        return
+    highlighter = app.builder.highlighter  # type: ignore
+    urito = app.builder.get_relative_uri
+
+    modnames = set(env._viewcode_modules)  # type: ignore
+
+    for modname, entry in status_iterator(
+            sorted(env._viewcode_modules.items()),  # type: ignore
+            __('highlighting module code... '), "blue",
+            len(env._viewcode_modules),  # type: ignore
+            app.verbosity, lambda x: x[0]):
+        if not entry:
+            continue
+        code, tags, used, refname = entry
+        # construct a page name for the highlighted source
+        pagename = '_modules/' + modname.replace('.', '/')
+        # highlight the source using the builder's highlighter
+        if env.config.highlight_language in ('python3', 'default', 'none'):
+            lexer = env.config.highlight_language
+        else:
+            lexer = 'python'
+        highlighted = highlighter.highlight_block(code, lexer, linenos=False)
+        # split the code into lines
+        lines = highlighted.splitlines()
+        # split off wrap markup from the first line of the actual code
+        before, after = lines[0].split('<pre>')
+        lines[0:1] = [before + '<pre>', after]
+        # nothing to do for the last line; it always starts with </pre> anyway
+        # now that we have code lines (starting at index 1), insert anchors for
+        # the collected tags (HACK: this only works if the tag boundaries are
+        # properly nested!)
+        maxindex = len(lines) - 1
+        for name, docname in used.items():
+            type, start, end = tags[name]
+            backlink = urito(pagename, docname) + '#' + refname + '.' + name
+            lines[start] = (
+                '<div class="viewcode-block" id="%s"><a class="viewcode-back" '
+                'href="%s">%s</a>' % (name, backlink, _('[文档]')) +
+                lines[start])
+            lines[min(end, maxindex)] += '</div>'
+        # try to find parents (for submodules)
+        parents = []
+        parent = modname
+        while '.' in parent:
+            parent = parent.rsplit('.', 1)[0]
+            if parent in modnames:
+                parents.append({
+                    'link': urito(pagename, '_modules/' +
+                                  parent.replace('.', '/')),
+                    'title': parent})
+        parents.append({'link': urito(pagename, '_modules/index'),
+                        'title': _('Module code')})
+        parents.reverse()
+        # putting it all together
+        context = {
+            'parents': parents,
+            'title': modname,
+            'body': (_('<h1>Source code for %s</h1>') % modname +
+                     '\n'.join(lines)),
+        }
+        yield (pagename, context, 'page.html')
+
+    if not modnames:
+        return
+
+    html = ['\n']
+    # the stack logic is needed for using nested lists for submodules
+    stack = ['']
+    for modname in sorted(modnames):
+        if modname.startswith(stack[-1]):
+            stack.append(modname + '.')
+            html.append('<ul>')
+        else:
+            stack.pop()
+            while not modname.startswith(stack[-1]):
+                stack.pop()
+                html.append('</ul>')
+            stack.append(modname + '.')
+        html.append('<li><a href="%s">%s</a></li>\n' % (
+            urito('_modules/index', '_modules/' + modname.replace('.', '/')),
+            modname))
+    html.append('</ul>' * (len(stack) - 1))
+    context = {
+        'title': _('Overview: module code'),
+        'body': (_('<h1>All modules for which code is available</h1>') +
+                 ''.join(html)),
+    }
+
+    yield ('_modules/index', context, 'page.html')
+
+
+def migrate_viewcode_import(app: Sphinx, config: Config) -> None:
+    if config.viewcode_import is not None:
+        warnings.warn('viewcode_import was renamed to viewcode_follow_imported_members. '
+                      'Please update your configuration.',
+                      RemovedInSphinx30Warning, stacklevel=2)
+
+
+def setup(app: Sphinx) -> Dict[str, Any]:
+    app.add_config_value('viewcode_import', None, False)
+    app.add_config_value('viewcode_enable_epub', False, False)
+    app.add_config_value('viewcode_follow_imported_members', True, False)
+    app.connect('config-inited', migrate_viewcode_import)
+    app.connect('doctree-read', doctree_read)
+    app.connect('env-merge-info', env_merge_info)
+    app.connect('html-collect-pages', collect_pages)
+    app.connect('missing-reference', missing_reference)
+    # app.add_config_value('viewcode_include_modules', [], 'env')
+    # app.add_config_value('viewcode_exclude_modules', [], 'env')
+    app.add_event('viewcode-find-source')
+    app.add_event('viewcode-follow-imported')
+    return {
+        'version': sphinx.__display_version__,
+        'env_version': 1,
+        'parallel_read_safe': True
+    }
diff --git a/docs/xai/docs/requirements.txt b/docs/xai/docs/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..49a77fdec3a5c745edd40eaa223883c31500e975
--- /dev/null
+++ b/docs/xai/docs/requirements.txt
@@ -0,0 +1,10 @@
+sphinx >= 2.2.1, <= 2.4.4
+docutils == 0.16
+myst_parser == 0.14.0
+sphinx-markdown-tables
+sphinx_rtd_theme == 0.5.2
+numpy
+nbsphinx
+IPython
+ipykernel
+jieba
diff --git a/docs/xai/docs/source_en/conf.py b/docs/xai/docs/source_en/conf.py
new file mode 100644
index 0000000000000000000000000000000000000000..1603e8bf6b59f180623ecd95d7adf6c4d6c5d6ae
--- /dev/null
+++ b/docs/xai/docs/source_en/conf.py
@@ -0,0 +1,129 @@
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+import os
+import sys
+import IPython
+import re
+sys.path.append(os.path.abspath('../_ext'))
+from sphinx.ext import autodoc as sphinx_autodoc
+
+import mindspore_xai
+
+# -- Project information -----------------------------------------------------
+
+project = 'MindSpore'
+copyright = '2021, MindSpore'
+author = 'MindSpore'
+
+# The full version, including alpha/beta/rc tags
+release = 'master'
+
+
+# -- General configuration ---------------------------------------------------
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.doctest',
+    'sphinx.ext.intersphinx',
+    'sphinx.ext.todo',
+    'sphinx.ext.coverage',
+    'sphinx.ext.napoleon',
+    'sphinx.ext.viewcode',
+    'sphinx_markdown_tables',
+    'myst_parser',
+    'nbsphinx',
+    'sphinx.ext.mathjax',
+    'IPython.sphinxext.ipython_console_highlighting'
+]
+
+source_suffix = {
+    '.rst': 'restructuredtext',
+    '.md': 'markdown',
+}
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns = []
+
+suppress_warnings = [
+    'nbsphinx',
+]
+
+pygments_style = 'sphinx'
+
+autodoc_inherit_docstrings = False
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = 'sphinx_rtd_theme'
+
+# Example configuration for intersphinx: refer to the Python standard library.
+intersphinx_mapping = {
+    'python': ('https://docs.python.org/', '../../../../resource/python_objects.inv'),
+    'numpy': ('https://docs.scipy.org/doc/numpy/', '../../../../resource/numpy_objects.inv'),
+}
+
+# Modify default signatures for autodoc.
+autodoc_source_path = os.path.abspath(sphinx_autodoc.__file__)
+autodoc_source_re = re.compile(r'stringify_signature\(.*?\)')
+get_param_func_str = r"""\
+import re
+import inspect as inspect_
+
+def get_param_func(func):
+    try:
+        source_code = inspect_.getsource(func)
+        if func.__doc__:
+            source_code = source_code.replace(func.__doc__, '')
+        all_params_str = re.findall(r"def [\w_\d\-]+\(([\S\s]*?)(\):|\) ->.*?:)", source_code)
+        all_params = re.sub("(self|cls)(,|, )?", '', all_params_str[0][0].replace("\n", "").replace("'", "\""))
+        return all_params
+    except:
+        return ''
+
+def get_obj(obj):
+    if isinstance(obj, type):
+        return obj.__init__
+
+    return obj
+"""
+
+with open(autodoc_source_path, "r+", encoding="utf8") as f:
+    code_str = f.read()
+    code_str = autodoc_source_re.sub('"(" + get_param_func(get_obj(self.object)) + ")"', code_str, count=0)
+    exec(get_param_func_str, sphinx_autodoc.__dict__)
+    exec(code_str, sphinx_autodoc.__dict__)
+
+sys.path.append(os.path.abspath('../../../../resource/sphinx_ext'))
+import anchor_mod
+import nbsphinx_mod
+
+
+sys.path.append(os.path.abspath('../../../../resource/search'))
+import search_code
+
+sys.path.append(os.path.abspath('../../../../resource/custom_directives'))
+from custom_directives import IncludeCodeDirective
+
+def setup(app):
+    app.add_directive('includecode', IncludeCodeDirective)
diff --git a/docs/xai/docs/source_en/images/grad_cam_saliency.png b/docs/xai/docs/source_en/images/grad_cam_saliency.png
new file mode 100644
index 0000000000000000000000000000000000000000..62ef2e908621dbcebf805151a1f8e37bd894884b
Binary files /dev/null and b/docs/xai/docs/source_en/images/grad_cam_saliency.png differ
diff --git a/docs/xai/docs/source_en/images/lime_tabular.png b/docs/xai/docs/source_en/images/lime_tabular.png
new file mode 100644
index 0000000000000000000000000000000000000000..e0a05861c0f1c69a1d0f55e0b7b0aa7bc427201f
Binary files /dev/null and b/docs/xai/docs/source_en/images/lime_tabular.png differ
diff --git a/docs/xai/docs/source_en/images/rise_plus_saliency.png b/docs/xai/docs/source_en/images/rise_plus_saliency.png
new file mode 100644
index 0000000000000000000000000000000000000000..3c0f4c43395dd2baa10a1d1441a73acb7aa9d6a5
Binary files /dev/null and b/docs/xai/docs/source_en/images/rise_plus_saliency.png differ
diff --git a/docs/xai/docs/source_en/images/saliency_overlay.png b/docs/xai/docs/source_en/images/saliency_overlay.png
new file mode 100644
index 0000000000000000000000000000000000000000..19fa0ee5cce1bc041462747debf679b162a3f08a
Binary files /dev/null and b/docs/xai/docs/source_en/images/saliency_overlay.png differ
diff --git a/docs/xai/docs/source_en/images/shap_gradient.png b/docs/xai/docs/source_en/images/shap_gradient.png
new file mode 100644
index 0000000000000000000000000000000000000000..5ad88ddb6f32d1b50739892f8a2829431e8655e8
Binary files /dev/null and b/docs/xai/docs/source_en/images/shap_gradient.png differ
diff --git a/docs/xai/docs/source_en/images/shap_kernel.png b/docs/xai/docs/source_en/images/shap_kernel.png
new file mode 100644
index 0000000000000000000000000000000000000000..0cd7e1b5e1951dd6633b66afd2a5d7081fa83d90
Binary files /dev/null and b/docs/xai/docs/source_en/images/shap_kernel.png differ
diff --git a/docs/xai/docs/source_en/images/xai_en.png b/docs/xai/docs/source_en/images/xai_en.png
new file mode 100644
index 0000000000000000000000000000000000000000..a169b965dbbf4ea369279ef257c3e3ae77bd5bda
Binary files /dev/null and b/docs/xai/docs/source_en/images/xai_en.png differ
diff --git a/docs/xai/docs/source_en/index.rst b/docs/xai/docs/source_en/index.rst
new file mode 100644
index 0000000000000000000000000000000000000000..29e1d1a0d8796aeac66123119ff4052024c3c416
--- /dev/null
+++ b/docs/xai/docs/source_en/index.rst
@@ -0,0 +1,40 @@
+MindSpore XAI Documents
+===========================
+
+Currently, most deep learning models are black-box models with good performance but poor explainability. The MindSpore XAI - a MindSpore-based explainable AI toolbox - provides a variety of explanation and decision methods to help you better understand, trust, and improve models. It also evaluates the explanation methods from various dimensions, enabling you to compare and select methods best suited to your environment.
+
+.. raw:: html
+
+   <img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/docs/xai/docs/source_en/images/xai_en.png" width="700px" alt="" >
+
+.. toctree::
+   :glob:
+   :maxdepth: 1
+   :caption: Installation
+
+   installation
+
+.. toctree::
+   :glob:
+   :maxdepth: 1
+   :caption: Guide
+
+   using_cv_explainers
+   using_cv_benchmarks
+   using_tabular_explainers
+   using_tabsim
+   using_tbnet
+
+.. toctree::
+   :maxdepth: 1
+   :caption: API References
+
+   mindspore_xai.explainer
+   mindspore_xai.benchmark
+   mindspore_xai.tool
+
+.. toctree::
+   :maxdepth: 1
+   :caption: FAQ
+
+   troubleshoot
diff --git a/docs/xai/docs/source_en/installation.md b/docs/xai/docs/source_en/installation.md
new file mode 100644
index 0000000000000000000000000000000000000000..cbefaebe068ff298d7843cfbf4f40b865c3ee5b3
--- /dev/null
+++ b/docs/xai/docs/source_en/installation.md
@@ -0,0 +1,55 @@
+# MindSpore XAI Installation
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/installation.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## System Requirements
+
+- OS: EulerOS-aarch64, CentOS-aarch64, CentOS-x86, Ubuntu-aarch64 or Ubuntu-x86
+- Device: Ascend 910 or GPU CUDA 10.1, 11.1
+- Python 3.7.5 or 3.9.0
+- MindSpore 1.7 or above
+
+## Installing by pip
+
+Download the `.whl` package from [MindSpore XAI download page](https://www.mindspore.cn/versions/en) and install with `pip`.
+
+```bash
+pip install mindspore_xai-{version}-py3-none-any.whl
+```
+
+## Installing from Source Code
+
+1. Download source code from gitee.com:
+
+    ```bash
+    git clone https://gitee.com/mindspore/xai.git
+    ```
+
+2. Install the dependency python modules:
+
+    ```bash
+    cd xai
+    pip install -r requirements.txt
+    ```
+
+3. Install the XAI module from source code:
+
+    ```bash
+    python setup.py install
+    ```
+
+4. Optionally, you may build a `.whl` package for installation without step 3:
+
+    ```bash
+    bash package.sh
+    pip install output/mindspore_xai-{version}-py3-none-any.whl
+    ```
+
+## Installation Verification
+
+Upon successful installation, importing 'mindspore_xai' module in Python will cause no error:
+
+```python
+import mindspore_xai
+print(mindspore_xai.__version__)
+```
diff --git a/docs/xai/docs/source_en/mindspore_xai.benchmark.rst b/docs/xai/docs/source_en/mindspore_xai.benchmark.rst
new file mode 100644
index 0000000000000000000000000000000000000000..cdc230be7fb5a8ddb6cf52a735b9e247c0db0dc3
--- /dev/null
+++ b/docs/xai/docs/source_en/mindspore_xai.benchmark.rst
@@ -0,0 +1,5 @@
+mindspore_xai.benchmark
+=======================
+
+.. automodule:: mindspore_xai.benchmark
+    :members:
diff --git a/docs/xai/docs/source_en/mindspore_xai.explainer.rst b/docs/xai/docs/source_en/mindspore_xai.explainer.rst
new file mode 100644
index 0000000000000000000000000000000000000000..788e850f3ff56d21347009bc6a4c5f6c29f6f587
--- /dev/null
+++ b/docs/xai/docs/source_en/mindspore_xai.explainer.rst
@@ -0,0 +1,5 @@
+mindspore_xai.explainer
+=========================
+
+.. automodule:: mindspore_xai.explainer
+    :members:
diff --git a/docs/xai/docs/source_en/mindspore_xai.tool.rst b/docs/xai/docs/source_en/mindspore_xai.tool.rst
new file mode 100644
index 0000000000000000000000000000000000000000..98d0533c178fc70e3c50cc6cc25952ea7a352b2c
--- /dev/null
+++ b/docs/xai/docs/source_en/mindspore_xai.tool.rst
@@ -0,0 +1,5 @@
+mindspore_xai.tool
+=========================
+
+.. automodule:: mindspore_xai.tool.cv
+    :members:
diff --git a/docs/xai/docs/source_en/troubleshoot.md b/docs/xai/docs/source_en/troubleshoot.md
new file mode 100644
index 0000000000000000000000000000000000000000..44455a80f148e0ca31a55f41a8fb72b65c567248
--- /dev/null
+++ b/docs/xai/docs/source_en/troubleshoot.md
@@ -0,0 +1,36 @@
+# Troubleshooting
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/troubleshoot.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## Import Errors
+
+<font size=3>**Q: What can I do if libgomp `cannot allocate memory in static TLS block` error occurs when importing `mindspore_xai` or its subpackages?**</font>
+
+A: You have to do the following steps:
+
+Reinstall scikit-learn 1.0.2:
+
+```bash
+pip install --force-reinstall scikit-learn==1.0.2
+```
+
+List all site-packages directories:
+
+```bash
+python -m site
+```
+
+There is a list of directories in `sys.path` shown the previous, find the `scikit_learn.libs/` sub-directory in the directories of the list.
+Once you located the `scikit_learn.libs/` directory, say which is underneath `<SITE_PKGS>/`, list files inside it:
+
+```bash
+ls <SITE_PKGS>/scikit_learn.libs
+```
+
+There is a dynamical library libgomp inside with a filename like `libgomp-XXX.so.XXX`, append the absolute path to the environment variable `LD_PRELOAD`:
+
+```bash
+export LD_PRELOAD=$LD_PRELOAD:<SITE_PKGS>/scikit_learn.libs/libgomp-XXX.so.XXX
+```
+
+Run your MindSpore XAI scripts again.
diff --git a/docs/xai/docs/source_en/using_cv_benchmarks.md b/docs/xai/docs/source_en/using_cv_benchmarks.md
new file mode 100644
index 0000000000000000000000000000000000000000..06a75025d0133de79c02fa1e22c8665a292893fd
--- /dev/null
+++ b/docs/xai/docs/source_en/using_cv_benchmarks.md
@@ -0,0 +1,84 @@
+# Using CV Benchmarks
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/using_cv_benchmarks.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## What are CV Benchmarks
+
+Benchmarks are algorithms evaluating the goodness of saliency maps from explainers. MindSpore XAI currently provides 4 benchmarks for image classification scenario: `Robustness`, `Faithfulness`, `ClassSensitivity` and `Localization`.
+
+## Preparations
+
+The complete code of the tutorial below is [using_cv_benchmarks.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_cv_benchmarks.py).
+
+Please follow the [Downloading Data Package](https://www.mindspore.cn/xai/docs/en/r1.8/using_cv_explainers.html#downloading-data-package) instructions to download the necessary files for the tutorial.
+
+With the tutorial package, we have to get the sample image, trained classifier, explainer and optionally the saliency map ready:
+
+```python
+# have to change the current directory to xai/examples/ first
+import mindspore as ms
+from mindspore_xai.explainer import GradCAM
+
+from common.resnet import resnet50
+from common.dataset import load_image_tensor
+
+# only PYNATIVE_MODE is supported
+ms.set_context(mode=ms.PYNATIVE_MODE)
+
+# 20 classes
+num_classes = 20
+
+# load the trained classifier
+net = resnet50(num_classes)
+param_dict = ms.load_checkpoint("xai_examples_data/ckpt/resnet50.ckpt")
+ms.load_param_into_net(net, param_dict)
+
+# [1, 3, 224, 224] Tensor
+boat_image = load_image_tensor('xai_examples_data/test/boat.jpg')
+
+# explainer
+grad_cam = GradCAM(net, layer='layer4')
+
+# 3 is the class id of 'boat'
+saliency = grad_cam(boat_image, targets=3)
+```
+
+## Using Robustness
+
+`Robustness` is the simplest benchmark, it perturbs the inputs by adding random noise and outputs the maximum sensitivity as evaluation score from the perturbations:
+
+```python
+from mindspore.nn import Softmax
+from mindspore_xai.benchmark import Robustness
+
+# the classifier use Softmax as activation function
+robustness = Robustness(num_classes, activation_fn=Softmax())
+# the 'saliency' argument is optional
+score = robustness.evaluate(grad_cam, boat_image, targets=3, saliency=saliency)
+```
+
+The returned `score` is a 1D tensor with only one float value for an 1xCx224x224 image tensor.
+
+## Using Faithfulness and ClassSensitivity
+
+The ways of using `Faithfulness` and `ClassSensitivity` are very similar to `Robustness`. However, `ClassSensitivity` is class agnostic, `targets` can not be specified.
+
+## Using Localization
+
+If the object region or bounding box is provided, `Localization` can be used. It evaluates base on how many saliency pixels fall inside the object region:
+
+```python
+import numpy as np
+import mindspore as ms
+from mindspore_xai.benchmark import Localization
+
+# top-left:80,66 bottom-right:223,196 is the bounding box of a boat
+mask = np.zeros([1, 1, 224, 224])
+mask[:, :, 66:196, 80:223] = 1
+
+mask = ms.Tensor(mask, dtype=ms.float32)
+
+localization = Localization(num_classes)
+
+score = localization.evaluate(grad_cam, boat_image, targets=3, mask=mask)
+```
diff --git a/docs/xai/docs/source_en/using_cv_explainers.md b/docs/xai/docs/source_en/using_cv_explainers.md
new file mode 100644
index 0000000000000000000000000000000000000000..41673c3b90f188ddb0b396e46d93b3050aa4df51
--- /dev/null
+++ b/docs/xai/docs/source_en/using_cv_explainers.md
@@ -0,0 +1,246 @@
+# Using CV Explainers
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/using_cv_explainers.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## What are CV Explainers
+
+Explainers are algorithms explaining the decisions made by AI models. MindSpore XAI currently provides 7 explainers for image classification scenario. Saliency maps (or heatmaps) are the outputs, their brightness represents the importance of the corresponding regions on the original image.
+
+A saliency map overlay on top of the original image:
+
+![saliency_overlay](./images/saliency_overlay.png)
+
+There are 2 categories of explainers: gradient based and perturbation based. The gradient based explainers rely on the backpropagation method to compute the pixel importance while the perturbation based explainers exploit random perturbations on the original images.
+
+| Explainer          | Category        |
+|:------------------:|:---------------:|
+| Gradient           | gradient        |
+| GradCAM            | gradient        |
+| GuidedBackprop     | gradient        |
+| Deconvolution      | gradient        |
+| Occlusion          | perturbation    |
+| RISE               | perturbation    |
+| RISEPlus           | perturbation    |
+
+## Preparations
+
+### Downloading Data Package
+
+First of all, we have to download the data package and put it underneath the `xai/examples/` directory of a local XAI [source package](https://gitee.com/mindspore/xai):
+
+```bash
+wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/xai_examples_data.tar.gz
+tar -xf xai_examples_data.tar.gz
+
+git clone https://gitee.com/mindspore/xai
+mv xai_examples_data xai/examples/
+```
+
+`xai/examples/` files:
+
+```bash
+xai/examples/
+├── xai_examples_data/
+│    ├── ckpt/
+│    │    ├── resent50.ckpt
+│    ├── train/
+│    └── test/
+├── common/
+│    ├── dataset.py
+│    └── resnet.py
+├── using_cv_explainers.py
+├── using_rise_plus.py
+└── using_cv_benchmarks.py
+```
+
+- `xai_examples_data/`: The extracted data package.
+- `xai_examples_data/ckpt/resent50.ckpt`: ResNet50 checkpoint file.
+- `xai_examples_data/test`: Test dataset.
+- `xai_examples_data/train`: Training dataset.
+- `common/dataset.py`: Dataset loader.
+- `common/resnet.py`: ResNet model definitions.
+- `using_cv_explainers.py`: Example of using explainers.
+- `using_rise_plus.py`: Example of using RISEPlus explainer.
+- `using_cv_benchmarks.py`: Example of using benchmarks.
+
+### Preparing Python Environment
+
+The complete code of the tutorial below is [using_cv_explainers.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_cv_explainers.py).
+
+In order to explain an image classification predication, we have to have a trained CNN network (`nn.Cell`) and an image to be examined:
+
+```python
+# have to change the current directory to xai/examples/ first
+import mindspore as ms
+from common.resnet import resnet50
+from common.dataset import load_image_tensor
+
+# only PYNATIVE_MODE is supported
+ms.set_context(mode=ms.PYNATIVE_MODE)
+
+# 20 classes
+num_classes = 20
+
+# load the trained classifier
+net = resnet50(num_classes)
+param_dict = ms.load_checkpoint("xai_examples_data/ckpt/resnet50.ckpt")
+ms.load_param_into_net(net, param_dict)
+
+# [1, 3, 224, 224] Tensor
+boat_image = load_image_tensor("xai_examples_data/test/boat.jpg")
+```
+
+## Using GradCAM
+
+`GradCAM` is a typical and effective gradient based explainer:
+
+```python
+from PIL import Image
+import mindspore as ms
+from mindspore_xai.explainer import GradCAM
+from mindspore_xai.visual.cv import saliency_to_image
+
+# usually specify the last convolutional layer
+grad_cam = GradCAM(net, layer="layer4")
+
+# 3 is the class id of 'boat'
+saliency = grad_cam(boat_image, targets=3, show=False)
+
+# convert the saliency map to a PIL.Image.Image object
+boat_img = Image.open("xai_examples_data/test/boat.jpg")
+saliency_to_image(saliency, boat_img)
+```
+
+The returned `saliency` is a 1x1x224x224 tensor for an 1xCx224x224 image tensor, which stores all pixel importances (range:[0.0, 1.0]) to the classification decision of 'boat'. Users may specify any class to be explained.
+
+![grad_cam_saliency](./images/grad_cam_saliency.png)
+
+### Batch Explanation
+
+For gradient based explainers, batch explanation is usually more efficient. Other explainers may also batch the evaluations:
+
+```python
+from common.dataset import load_dataset
+
+test_ds = load_dataset('xai_examples_data/test').batch(4)
+
+for images, labels in test_ds:
+    saliencies = grad_cam(images, targets=ms.Tensor([3, 3, 3, 3], dtype=ms.int32))
+    # other custom operations ...
+```
+
+The returned `saliencies` is a 4x1x224x224 tensor for a 4xCx224x224 batched image tensor.
+
+### Using Other Explainers
+
+The ways of using other explainers are very similar to `GradCAM`, except `RISEPlus`.
+
+## Using RISEPlus
+
+The complete code of the tutorial below is [using_rise_plus.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_rise_plus.py).
+
+`RISEPlus` is based on `RISE` with an introduction of Out-of-Distribution(OoD) detector. It solves the degeneration problem of `RISE` on samples that the classifier had never seem the similar in training.
+
+First, we need to train an OoD detector(`OoDNet`) with the classifier training dataset:
+
+```python
+# have to change the current directory to xai/examples/ first
+import mindspore as ms
+from mindspore.nn import Softmax, SoftmaxCrossEntropyWithLogits
+from mindspore_xai.tool.cv import OoDNet
+from mindspore_xai.explainer import RISEPlus
+from common.dataset import load_dataset
+from common.resnet import resnet50
+
+# only PYNATIVE_MODE is supported
+ms.set_context(mode=ms.PYNATIVE_MODE)
+
+num_classes = 20
+
+# classifier training dataset
+train_ds = load_dataset('xai_examples_data/train').batch(4)
+
+# load the trained classifier
+net = resnet50(num_classes)
+param_dict = ms.load_checkpoint('xai_examples_data/ckpt/resnet50.ckpt')
+ms.load_param_into_net(net, param_dict)
+
+ood_net = OoDNet(underlying=net, num_classes=num_classes)
+
+# use SoftmaxCrossEntropyWithLogits as loss function if the activation function of
+# the classifier is Softmax, use BCEWithLogitsLoss if the activation function is Sigmoid
+ood_net.train(train_ds, loss_fn=SoftmaxCrossEntropyWithLogits())
+
+ms.save_checkpoint(ood_net, 'ood_net.ckpt')
+```
+
+The classifier for `OoDNet` must be a subclass of `nn.Cell`, in `__init__()` which must:
+
+- defines an `int` member attribute named `num_features` as the number of feature values to be returned by the feature layer.
+
+- defines a `bool` member attribute named `output_features` with `False` as initial value, OoDNet tells the classifier to return the feature tensor in `construct()` by setting `output_features` to `True`.
+
+A LeNet5 example of underlying classifier:
+
+```python
+from mindspore import nn
+from mindspore.common.initializer import Normal
+
+class MyLeNet5(nn.Cell):
+
+    def __init__(self, num_class, num_channel):
+        super(MyLeNet5, self).__init__()
+
+        # must add the following 2 attributes to your model
+        self.num_features = 84 # no. of features, int
+        self.output_features = False # output features flag, bool
+
+        self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid')
+        self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid')
+        self.relu = nn.ReLU()
+        self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.flatten = nn.Flatten()
+        self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02))
+        self.fc2 = nn.Dense(120, self.num_features, weight_init=Normal(0.02))
+        self.fc3 = nn.Dense(self.num_features, num_class, weight_init=Normal(0.02))
+
+    def construct(self, x):
+        x = self.conv1(x)
+        x = self.relu(x)
+        x = self.max_pool2d(x)
+        x = self.conv2(x)
+        x = self.relu(x)
+        x = self.max_pool2d(x)
+        x = self.flatten(x)
+        x = self.relu(self.fc1(x))
+        x = self.relu(self.fc2(x))
+
+        # return the features tensor if output_features is True
+        if self.output_features:
+            return x
+
+        x = self.fc3(x)
+        return x
+```
+
+Now we can use `RISEPlus` with the trained `OoDNet`:
+
+```python
+from PIL import Image
+from mindspore_xai.visual.cv import saliency_to_image
+
+# create a new classifier as the underlying when loading OoDNet from a checkpoint
+ood_net = OoDNet(underlying=resnet50(num_classes), num_classes=num_classes)
+param_dict = ms.load_checkpoint('ood_net.ckpt')
+ms.load_param_into_net(ood_net, param_dict)
+
+rise_plus = RISEPlus(ood_net=ood_net, network=net, activation_fn=Softmax())
+saliency = rise_plus(boat_image, targets=3, show=False)
+
+boat_img = Image.open("xai_examples_data/test/boat.jpg")
+saliency_to_image(saliency, boat_img)
+```
+
+The returned `saliency` is an 1x1x224x224 tensor for an 1xCx224x224 image tensor.
+
+![rise_plus_saliency](./images/rise_plus_saliency.png)
diff --git a/docs/xai/docs/source_en/using_tabsim.md b/docs/xai/docs/source_en/using_tabsim.md
new file mode 100644
index 0000000000000000000000000000000000000000..b41b3bb8d59a16cbfecacaac50d137fbd9a98d6b
--- /dev/null
+++ b/docs/xai/docs/source_en/using_tabsim.md
@@ -0,0 +1,176 @@
+# Using TabSim Data Simulator
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/using_tabsim.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## Introduction
+
+Sometimes, it is impossible to get sufficient amount of data for modeling or algorithm development, TabSim can be used for capturing the distribution of the tabular data and generating simulated data afterward.
+
+The complete code of the tutorial below is [using_tabsim.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_tabsim.py).
+
+## Installation
+
+TabSim is part of the XAI package, no extra installation is required besides [MindSpore](https://mindspore.cn/install/en) and [XAI](https://www.mindspore.cn/xai/docs/en/r1.8/installation.html).
+
+## User Flow
+
+There are 2 phases in the TabSim user flow:
+
+1. Digestion: Analyzing the real tabular data, capturing the statistic characteristics and output the digest file. Accomplished by the commandline tool `mindspore_xai tabdig`.
+2. Simulation: Generating simulated data according to the statistics stores in the digested file. Accomplished by the commandline tool `mindspore_xai tabsim`.
+
+## Digestion Phase
+
+```bash
+mindspore_xai tabdig <real datafile> <digest file> [--bins <bins>] [--clip-sd <clip sd>]
+```
+
+`<real datafile>`: Path of the real CSV table to be simulated.
+
+`<digest file>`: Path of the digest file to be saved.
+
+`<bins>`: [optional] Number of bins [2 - 32] for discretizing numeric columns, default: 10
+
+`<clip sd>`: [optional] Number of standard deviations away from the mean that defines the outliers, outlier values
+will be clipped. default: 3, setting to 0 or less will disable the value clipping.
+
+### File Format of Real Data
+
+The real data must be a CSV file with header that contains the names and type of all columns.
+
+Header pattern: `<col name>|<col type>,<col name>|<col type>,<col name>|<col type>,...`
+
+`<col name>`: Column name, allowed pattern: `[0-9a-zA-Z_\-]+`
+
+`<col type>`: Column type, options: 'int', 'float', 'str', 'cat'
+
+- 'int': Integers
+- 'float': Float numbers
+- 'str': Strings, allowed pattern: `[0-9a-zA-Z_\-\+\.]*`
+- 'cat': Catergorical values, the underlying data type is integer without order
+
+'int' and 'float' are numeric, while 'str' and 'cat' are discrete columns. There are at most 256 distinct values allowed
+in each discrete column.
+
+Optionally, users may specify at most one label column by adding a '*' before a discrete column (numeric columns are not
+allowed).
+
+Header example: `col_A|int,col_B|float,col_C|str,*col_D|cat`
+
+It is recommended users randomly pick around 1 million records from the real database to form the real data file for the
+statistic accuracy and the memory constraints.
+
+### File Format of Digest Files
+
+Digest files are clear text json, they stores the name, type and value distributions of columns without any actual record.
+Users should never modify the digest file manually; otherwise that may corrupt it.
+
+### Digestion Example
+
+We used the [Iris](https://scikit-learn.org/stable/auto_examples/datasets/plot_iris_dataset.html) dataset for the
+demonstration. This dataset consists of 3 different types of irises’ petal and sepal lengths. The Python code below
+writes tabular data to `real_table.csv`.
+
+```python
+import sklearn.datasets
+
+iris = sklearn.datasets.load_iris()
+features = iris.data
+labels = iris.target
+# save the tabular data to file
+header = 'sepal_length|float,sepal_width|float,petal_length|float,petal_width|float,*class|cat'
+with open('real_table.csv', 'w') as f:
+    f.write(header + '\n')
+    for i in range(len(labels)):
+        for feat in features[i]:
+            f.write("{},".format(feat))
+        f.write("{}\n".format(labels[i]))
+```
+
+Content of `real_table.csv`:
+
+```text
+sepal_length|float,sepal_width|float,petal_length|float,petal_width|float,*class|cat
+5.1,3.5,1.4,0.2,0
+4.9,3.0,1.4,0.2,0
+4.7,3.2,1.3,0.2,0
+4.6,3.1,1.5,0.2,0
+5.0,3.6,1.4,0.2,0
+5.4,3.9,1.7,0.4,0
+4.6,3.4,1.4,0.3,0
+5.0,3.4,1.5,0.2,0
+4.4,2.9,1.4,0.2,0
+...
+```
+
+Then, we analyze the real tabular data, capturing the statistic characteristics and output it to `digest.json`.
+
+```bash
+mindspore_xai tabdig real_table.csv digest.json
+```
+
+Content of `digest.json`:
+
+```json
+{
+    "label_col_idx": 4,
+    "columns": [
+        {
+            "name": "sepal_length",
+            "idx": 0,
+            "ctype": "float",
+            "dtype": "float",
+            "is_numeric": true,
+            "is_label": false,
+            ...
+```
+
+## Simulation Phase
+
+```bash
+mindspore_xai tabsim <digest file> <sim datafile> <rows> [--batch-size <batch size>] [--noise <noise>]
+```
+
+`<digest file>`: Path of the digest file of the real data.
+
+`<sim datafile>`: Path of the simulated CSV table.
+
+`<rows>`: Number of rows to be generated to `<sim datafile>`.
+
+`<batch size>`: [optional] Number of rows in each batch, default: 10000
+
+`<noise>`: [optional] Noise level (0.0-1.0) of value picking probabilities, 0.0 means 100% follows the digested joint
+distributions, higher the noise level more even the probabilities. default: 0
+
+### File Format of Simulated Data
+
+The simulated CSV file has a similar format to the real data CSV file, but the header is different:
+
+`<col name>,<col name>,<col name>,...`
+
+It contains no `<col type>` and no `*` while the column order remains the same.
+
+### Simulation Example
+
+Here we generate 200000 rows of simulated data according to the statistics stored in the digested file, the simulated
+data is output to `sim_table.csv`.
+
+```bash
+mindspore_xai tabsim digest.json sim_table.csv 200000
+```
+
+Content of `sim_table.csv`:
+
+```text
+sepal_length,sepal_width,petal_length,petal_width,class
+5.577184113278916,2.600922272560204,4.432243573999988,1.3937476921377445,1
+6.723739024436704,2.7995789972671985,4.093195099230183,1.377081159510022,1
+4.787110003892638,2.8994714750972608,1.221068662892122,0.18023497892950327,0
+5.47601589088659,2.683719381022501,4.429520567795243,1.44376166769605,1
+5.713634033969561,2.238437659593092,4.468051986603512,1.5218876291352155,1
+6.014412107785783,2.921972441210267,4.066770696930024,0.9183809029577147,1
+6.188386742135447,2.92122446931648,5.288862927543273,1.4537708701756062,2
+7.394485586937094,2.867479423550221,5.730391070749579,1.998759192383688,2
+5.468839597899383,2.8957462954323083,4.4090170094158525,1.502682955942951,1
+...
+```
diff --git a/docs/xai/docs/source_en/using_tabular_explainers.md b/docs/xai/docs/source_en/using_tabular_explainers.md
new file mode 100644
index 0000000000000000000000000000000000000000..af60703c60d6f313be509e3f9a99fc2d77d948d6
--- /dev/null
+++ b/docs/xai/docs/source_en/using_tabular_explainers.md
@@ -0,0 +1,183 @@
+# Using Tabular Explainers
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/using_tabular_explainers.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## Introduction
+
+In this tutorial we explain the tabular data classification using 3 different explainers, including `LIMETabular`,
+`SHAPKernel`, and `SHAPGradient`.
+
+All explainers support `PYNATIVE_MODE`. All explainers except `SHAPGradient` support `GRAPH_MODE`.
+
+|  Explainer   |    PYNATIVE_MODE    |     GRAPH_MODE     |
+|:------------:|:-------------------:|:------------------:|
+| LIMETabular  |      Supported      |     Supported      |
+|  SHAPKernel  |      Supported      |     Supported      |
+| SHAPGradient |      Supported      |      <blank>       |
+
+The complete code of the tutorial below is [using_tabular_explainers.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_tabular_explainers.py).
+
+## Import Dataset
+
+We use the [Iris](https://scikit-learn.org/stable/auto_examples/datasets/plot_iris_dataset.html) dataset for the demonstration.
+These data sets consist of 3 different types of irises’ petal and sepal lengths.
+
+```python
+import sklearn.datasets
+import mindspore as ms
+
+iris = sklearn.datasets.load_iris()
+
+# feature_names: ['sepal length (cm)', 'sepal width (cm)', 'petal length (cm)', 'petal width (cm)']
+feature_names = iris.feature_names
+# class_names: ['setosa', 'versicolor', 'virginica']
+class_names = list(iris.target_names)
+
+# convert data and labels from numpy array to mindspore tensor
+# use the first 100 samples
+data = ms.Tensor(iris.data, ms.float32)[:100]
+labels = ms.Tensor(iris.target, ms.int32)[:100]
+
+# explain the first sample
+inputs = data[:1]
+# explain the label 'setosa'(class index 0)
+targets = 0
+```
+
+## Import Model
+
+Here we define a simple linear classifier.
+
+```python
+import numpy as np
+import mindspore.nn as nn
+
+
+class LinearNet(nn.Cell):
+    def __init__(self):
+        super(LinearNet, self).__init__()
+        # input features: 4
+        # output classes: 3
+        self.linear = nn.Dense(4, 3, activation=nn.Softmax())
+
+    def construct(self, x):
+        x = self.linear(x)
+        return x
+
+
+net = LinearNet()
+
+# load pre-trained parameters
+weight = np.array([[0.648, 1.440, -2.05, -0.977], [0.507, -0.276, -0.028, -0.626], [-1.125, -1.183, 2.099, 1.605]])
+bias = np.array([0.308, 0.343, -0.652])
+net.linear.weight.set_data(ms.Tensor(weight, ms.float32))
+net.linear.bias.set_data(ms.Tensor(bias, ms.float32))
+```
+
+## Using LIMETabular
+
+`LIMETabular` approximates the machine learning model with a local, interpretable model to explain each individual
+prediction.
+
+```python
+from mindspore_xai.explainer import LIMETabular
+
+# convert features to feature stats
+feature_stats = LIMETabular.to_feat_stats(data, feature_names=feature_names)
+# initialize the explainer
+lime = LIMETabular(net, feature_stats, feature_names=feature_names, class_names=class_names)
+# explain
+lime_outputs = lime(inputs, targets, show=True)
+print("LIMETabular:")
+for i, exps in enumerate(lime_outputs):
+    for exp in exps:
+        print("Explanation for sample {} class {}:".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+output:
+> LIMETabular:
+>
+> Explanation for sample 0 class setosa:
+>
+> [('petal length (cm) <= 1.60', 0.8182714590301656),
+> ('sepal width (cm) > 3.30', 0.0816516722404966), ('petal width (cm) <= 0.30', 0.03557190104069489),
+> ('sepal length (cm) <= 5.10', -0.021441399016492325)]
+
+![lime_tabular](./images/lime_tabular.png)
+
+`LIMETabular` also supports a callable function, for example:
+
+```python
+def predict_fn(x):
+    return net(x)
+
+
+# initialize the explainer
+lime = LIMETabular(predict_fn, feature_stats, feature_names=feature_names, class_names=class_names)
+```
+
+## Using SHAPKernel
+
+`SHAPKernel` is a method that uses a special weighted linear regression to compute the importance of each feature.
+
+```python
+from mindspore_xai.explainer import SHAPKernel
+
+# initialize the explainer
+shap_kernel = SHAPKernel(net, data, feature_names=feature_names, class_names=class_names)
+# explain
+shap_kernel_outputs = shap_kernel(inputs, targets, show=True)
+print("SHAPKernel:")
+for i, exps in enumerate(shap_kernel_outputs):
+    for exp in exps:
+        print("Explanation for sample {} class {}:".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+output:
+> SHAPKernel:
+>
+> Explanation for sample 0 class setosa:
+>
+> [-0.00403276  0.03651359  0.59952676  0.01399141]
+
+![shap_kernel](./images/shap_kernel.png)
+
+`SHAPKernel` also supports a callable function, for example:
+
+```python
+# initialize the explainer
+shap_kernel = SHAPKernel(predict_fn, data, feature_names=feature_names, class_names=class_names)
+```
+
+## Using SHAPGradient
+
+`SHAPGradient` explains a model using expected gradients (an extension of integrated gradients).
+
+```python
+from mindspore_xai.explainer import SHAPGradient
+import mindspore as ms
+
+# Gradient only works under PYNATIVE_MODE.
+ms.set_context(mode=ms.PYNATIVE_MODE)
+# initialize the explainer
+shap_gradient = SHAPGradient(net, data, feature_names=feature_names, class_names=class_names)
+# explain
+shap_gradient_outputs = shap_gradient(inputs, targets, show=True)
+print("SHAPGradient:")
+for i, exps in enumerate(shap_gradient_outputs):
+    for exp in exps:
+        print("Explanation for sample {} class {}:".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+output:
+
+> SHAPGradient:
+>
+> Explanation for sample 0 class setosa:
+>
+> [-0.0112452   0.08389313  0.47006473  0.0373782 ]
+
+![shap_gradient](./images/shap_gradient.png)
diff --git a/docs/xai/docs/source_en/using_tbnet.md b/docs/xai/docs/source_en/using_tbnet.md
new file mode 100644
index 0000000000000000000000000000000000000000..abf6abaec9614decac48f04594bc9ab73cec39b6
--- /dev/null
+++ b/docs/xai/docs/source_en/using_tbnet.md
@@ -0,0 +1,287 @@
+# Using TB-Net Whitebox Recommendation Model
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_en/using_tbnet.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source_en.png"></a>
+
+## What is TB-Net
+
+TB-Net is a white box recommendation model, which constructs subgraphs in knowledge graphs based on the interaction between users and items as well as the features of items, and then calculates paths in the graphs using a bidirectional conduction algorithm. Finally, we can obtain explainable recommendation results.
+
+Paper: Shendi Wang, Haoyang Li, Xiao-Hui Li, Caleb Chen Cao, Lei Chen. Tower Bridge Net (TB-Net): Bidirectional Knowledge Graph Aware Embedding Propagation for Explainable Recommender Systems
+
+## Preparations
+
+### Downloading Data Package
+
+First of all, we have to download the data package and put it underneath the `models/whitebox/tbnet` directory of a local XAI [source package](https://gitee.com/mindspore/xai):
+
+```bash
+wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/tbnet_data.tar.gz
+tar -xf tbnet_data.tar.gz
+
+git clone https://gitee.com/mindspore/xai.git
+mv data xai/models/whitebox/tbnet
+```
+
+`xai/models/whitebox/tbnet/` files:
+
+```bash
+.
+└─tbnet
+  ├─README.md
+  ├─README_CN.md
+  ├─data
+  │ └─steam                         # Steam user purchase history dataset
+  │   ├─LICENSE
+  │   ├─config.json                 # hyper-parameters and training configuration
+  │   ├─src_infer.csv               # source datafile for inference
+  │   ├─src_test.csv                # source datafile for evaluation
+  │   └─src_train.csv               # source datafile for training
+  ├─src
+  │ ├─dataset.py                    # dataset loader
+  │ ├─embedding.py                  # embedding module
+  │ ├─metrics.py                    # model metrics
+  │ ├─path_gen.py                   # data preprocessor
+  │ ├─recommend.py                  # result aggregator
+  │ └─tbnet.py                      # TB-Net architecture
+  ├─export.py                       # export MINDIR/AIR script
+  ├─preprocess.py                   # data pre-processing script
+  ├─eval.py                         # evaluation script
+  ├─infer.py                        # inference and explaining script
+  ├─train.py                        # training script
+  └─tbnet_config.py                 # configuration reader
+```
+
+### Preparing Python Environment
+
+TB-Net is part of the XAI package, no extra installation is required besides [MindSpore](https://mindspore.cn/install/en) and [XAI](https://www.mindspore.cn/xai/docs/en/r1.8/installation.html). GPUs are supported.
+
+## Data Pre-processing
+
+The complete example code of this step is [preprocess.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/preprocess.py).
+
+Before training the TB-Net, we have to convert the source datafile to relation path data.
+
+### Source Datafile Format
+
+The source datafiles of the steam dataset all share the exact same CSV format with headers:
+
+`user,item,rating,developer,genre,category,release_year`
+
+The first 3 columns must be present with specific order and meaning:
+
+- `user`: String, user ID, records of the same user must be grouped in consecutive rows in a single file. Splitting the records across different files will give misleading results.
+- `item`: String, item ID.
+- `rating`: Character, either `c`(user had interactions (e.g. clicked) with the item but not purchased), `p`(user purchased the item) or `x`(other items).
+
+(Remark: There is no `c` rating item in the steam dataset.)
+
+Since the order and meaning of these columns are fixed, the names do not matter, users may choose other names like `uid,iid,act`, etc.
+
+The later columns `developer,genre,category,release_year` are for the item's string attribute IDs. Users should decide the column names (i.e. relation names) and keep them consistent in all source datafiles. There must be at least one attribute column with no maximum limit. In some cases, there are more than one values in each attribute, they should be separated by `;`. Leaving the attribute blank means the item has no such attribute.
+
+The content of source datafiles for different purposes are slightly different:
+
+- `src_train.csv`: For training, the numbers of rows of `p` rating and `c` + `x` rating items should be roughly the same by re-sampling, there is no need to list all items in every user.
+- `src_test.csv`: For evaluation, very similar to `src_train.csv` but with less amount of data.
+- `src_infer.csv`: For inference, must contain data of ONLY ONE user. ALL `c`, `p` and `x` rating items should be listed. In [preprocess.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/preprocess.py), only the `c` and `x` items are put as recommendation candidates in path data.
+
+### Converting to Relation Path Data
+
+```python
+import io
+import json
+from src.path_gen import PathGen
+
+path_gen = PathGen(per_item_paths=39)
+
+path_gen.generate("./data/steam/src_train.csv", "./data/steam/train.csv")
+
+# save id maps for the later use by Recommender for inference
+with io.open("./data/steam/id_maps.json", mode="w", encoding="utf-8") as f:
+    json.dump(path_gen.id_maps(), f, indent=4)
+
+# treat newly met items and references in src_test.csv and src_infer.csv as unseen entities
+# dummy internal id 0 will be assigned to them
+path_gen.grow_id_maps = False
+
+path_gen.generate("./data/steam/src_test.csv", "./data/steam/test.csv")
+
+# for inference, only take interacted('c') and other('x') items as candidate items,
+# the purchased('p') items won't be recommended.
+# assume there is only one user in src_infer.csv
+path_gen.subject_ratings = "cx"
+
+path_gen.generate("./data/steam/src_infer.csv", "./data/steam/infer.csv")
+```
+
+`PathGen` is responsible for converting source datafile into relation path data.
+
+### Relation Path Data Format
+
+Relation path data are header-less CSV (all integer values), with columns:
+
+`<subject item>,<label>,<relation1>,<reference>,<relation2>,<historical item>,...`
+
+- `subject item`: Internal id of the sample item for training or the recommendation candidate item for inference.`0` represents unseen items.
+- `label`: Ground truth label, `0` - not purchased, `1` - purchased. Ignored when inferencing.
+- `relation1`: Internal id of the relation 1 (column name of the subject item's attribute).
+- `reference`: Internal id of the reference entity (the common attribute ID).`0` represents unseen references.
+- `relation2`: Internal id of the relation 2 (column name of the historical item's attribute).
+- `historical item`: Internal id of the historical (`c` or `p` rating) item.`0` represents unseen items.
+
+The column sequence `<relation1>,<reference>,<relation2>,<historical item>` is going to repeat `per_item_paths` times in each row.
+
+### Running preprocess.py
+
+Have to change the current directory to `xai/models/whitebox/tbnet` first.
+
+```bash
+python preprocess.py
+```
+
+Dataset `./data/stream` will be processed with path data `train.csv`, `test.csv` and `infer.csv` and the source-internal ID maps `id_maps.json` generated in the dataset directory.
+
+## Training and Evaluation
+
+The complete example code of this step is [train.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/train.py).
+
+```python
+from src.tbnet import TBNet, NetWithLossCell, TrainStepWrapCell, EvalNet
+from src.dataset import create_dataset
+from src.metrics import AUC, ACC
+
+...
+
+train_ds = create_dataset(train_csv_path, cfg.per_item_paths).batch(cfg.batch_size)
+test_ds = create_dataset(test_csv_path, cfg.per_item_paths).batch(cfg.batch_size)
+
+print("creating TBNet for training...")
+network = TBNet(cfg.num_items, cfg.num_references, cfg.num_relations, cfg.embedding_dim)
+loss_net = NetWithLossCell(network, cfg.kge_weight, cfg.node_weight, cfg.l2_weight)
+train_net = TrainStepWrapCell(loss_net, cfg.lr)
+train_net.set_train()
+eval_net = EvalNet(network)
+time_callback = TimeMonitor(data_size=train_ds.get_dataset_size())
+loss_callback = MyLossMonitor()
+model = Model(network=train_net, eval_network=eval_net, metrics={'auc': AUC(), 'acc': ACC()})
+print("training...")
+for i in range(args.epochs):
+    print(f'===================== Epoch {i} =====================')
+    model.train(epoch=1, train_dataset=train_ds, callbacks=[time_callback, loss_callback], dataset_sink_mode=False)
+    train_out = model.eval(train_ds, dataset_sink_mode=False)
+    test_out = model.eval(test_ds, dataset_sink_mode=False)
+    print(f'Train AUC:{train_out["auc"]} ACC:{train_out["acc"]}  Test AUC:{test_out["auc"]} ACC:{test_out["acc"]}')
+
+    ckpt_path = os.path.join(ckpt_dir_path, f'tbnet_epoch{i}.ckpt')
+    save_checkpoint(network, ckpt_path)
+    print(f'checkpoint saved: {ckpt_path}')
+```
+
+We can see from the above code snap of [train.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/train.py) that `tbnet.py`, `metrics.py` and `dataset.py` provides all the classes for training TB-Net like other MindSpore models. It trains with dataset created from `./data/steam/train.csv` and `./data/steam/test.csv` (for evaluation).
+
+### Running train.py
+
+Have to change the current directory to `xai/models/whitebox/tbnet` first.
+
+```bash
+python train.py
+```
+
+20 epochs will be trained with the steam dataset and checkpoints will be saved to `./checkpoints/steam` for each epoch.
+
+## Inference with Explanations
+
+The complete example code of this step is [infer.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/infer.py).
+
+```python
+from src.tbnet import TBNet
+from src.recommend import Recommender
+from src.dataset import create_dataset
+
+...
+
+print(f"creating dataset from {data_path}...")
+infer_ds = create_dataset(data_path, cfg.per_item_paths)
+infer_ds = infer_ds.batch(cfg.batch_size)
+
+print("inferring...")
+# infer and aggregate results
+
+with io.open(id_maps_path, mode="r", encoding="utf-8") as f:
+    id_maps = json.load(f)
+recommender = Recommender(network, id_maps, args.items)
+
+for item, rl1, ref, rl2, hist_item, _ in infer_ds:
+    # assume there is data of only one user in infer_ds and all items inside are candidates.
+    recommender(item, rl1, ref, rl2, hist_item)
+
+# show recommendations with explanations
+suggestions = recommender.suggest()
+for suggest in suggestions:
+    print("")
+    print(f'Recommends item:"{suggest.item}" (score:{suggest.score}) because:')
+    # show explanations
+    explanation = 0
+    for path in suggest.paths:
+        if path.relation1 == path.relation2:
+            print(f'- it shares the same {path.relation1}:"{path.reference}" with user\'s '
+                  f'historical item:"{path.hist_item}".\n  (importance:{path.importance})')
+        else:
+            print(f'- it has {path.relation1}:"{path.reference}" while which is {path.relation2} '
+                  f'of user\'s historical item:"{path.hist_item}".\n  (importance:{path.importance})')
+        explanation += 1
+        if explanation >= args.explanations:
+            break
+```
+
+We can see from the above code snap of [infer.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/infer.py) that class `Recommender` aggregates all inference results from TB-Net and gives top-k item recommendations. Each recommended item comes with a sorted list of relation paths beginning from the most important one. All IDs and relation names returned by `Recommender` are from `./data/steam/id_maps.json` that originated from the source datafile `./data/steam/src_train.csv`.
+
+### Running infer.py
+
+Have to change the current directory to `xai/models/whitebox/tbnet` first.
+
+```bash
+python infer.py --checkpoint_id 19
+```
+
+The last checkpoint `./checkpoints/steam/tbnet_epoch19.ckpt` from [Running train.py](#running-train-py) will be used to infer on `./data/steam/infer.csv`. 3 recommended items with 3 relation path explanations each will be given.
+
+## Exporting Trained Model
+
+The complete example code of this step is [export.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/export.py).
+
+```python
+import mindspore as ms
+from src.tbnet import TBNet
+
+...
+
+network = TBNet(cfg.num_items, cfg.num_references, cfg.num_relations, cfg.embedding_dim)
+param_dict = ms.load_checkpoint(ckpt_path)
+ms.load_param_into_net(network, param_dict)
+
+item = ms.Tensor(np.ones((1,)).astype(np.int))
+rl1 = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+ref = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+rl2 = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+his = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+inputs = [item, rl1, ref, rl2, his]
+file_name = os.path.realpath(args.file_name)
+ms.export(network, *inputs, file_name=file_name, file_format=args.file_format)
+```
+
+The above code snap of [export.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/export.py) shows it is straightforward with [mindspore.export](https://www.mindspore.cn/docs/zh-CN/r1.8/api_python/mindspore/mindspore.export.html#mindspore.export).
+
+### Running export.py
+
+Have to change the current directory to `xai/models/whitebox/tbnet` first.
+
+```bash
+python export.py --config_path ./data/steam/config.json --checkpoint_path ./checkpoints/steam/tbnet_epoch19.ckpt
+```
+
+The trained model will be exported as `./tbnet.mindir`.
+
+## Example Script Arguments and Model Performance
+
+For the detail descriptions of the example script arguments and model performance, please refer to [README.md](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/README.md#).
diff --git a/docs/xai/docs/source_zh_cn/conf.py b/docs/xai/docs/source_zh_cn/conf.py
new file mode 100644
index 0000000000000000000000000000000000000000..ccccc6358e2b375d4ebbbb8915e1fc5dc339a434
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/conf.py
@@ -0,0 +1,154 @@
+# Configuration file for the Sphinx documentation builder.
+#
+# This file only contains a selection of the most common options. For a full
+# list see the documentation:
+# https://www.sphinx-doc.org/en/master/usage/configuration.html
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+import os
+import sys
+import IPython
+import re
+sys.path.append(os.path.abspath('../_ext'))
+from sphinx.ext import autodoc as sphinx_autodoc
+
+# -- Project information -----------------------------------------------------
+
+project = 'MindSpore'
+copyright = '2021, MindSpore'
+author = 'MindSpore'
+
+# The full version, including alpha/beta/rc tags
+release = 'master'
+
+
+# -- General configuration ---------------------------------------------------
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.doctest',
+    'sphinx.ext.intersphinx',
+    'sphinx.ext.todo',
+    'sphinx.ext.coverage',
+    'sphinx.ext.napoleon',
+    'sphinx.ext.viewcode',
+    'sphinx_markdown_tables',
+    'myst_parser',
+    'nbsphinx',
+    'sphinx.ext.mathjax',
+    'IPython.sphinxext.ipython_console_highlighting'
+]
+
+source_suffix = {
+    '.rst': 'restructuredtext',
+    '.md': 'markdown',
+}
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns = []
+
+pygments_style = 'sphinx'
+
+autodoc_inherit_docstrings = False
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = 'sphinx_rtd_theme'
+
+html_search_language = 'zh'
+
+html_search_options = {'dict': '../../../resource/jieba.txt'}
+
+# Example configuration for intersphinx: refer to the Python standard library.
+intersphinx_mapping = {
+    'python': ('https://docs.python.org/', '../../../../resource/python_objects.inv'),
+    'numpy': ('https://docs.scipy.org/doc/numpy/', '../../../../resource/numpy_objects.inv'),
+}
+
+from sphinx import directives
+with open('../_ext/overwriteobjectiondirective.txt', 'r', encoding="utf8") as f:
+    exec(f.read(), directives.__dict__)
+
+from sphinx.ext import viewcode
+with open('../_ext/overwriteviewcode.txt', 'r', encoding="utf8") as f:
+    exec(f.read(), viewcode.__dict__)
+
+# Modify default signatures for autodoc.
+autodoc_source_path = os.path.abspath(sphinx_autodoc.__file__)
+autodoc_source_re = re.compile(r'stringify_signature\(.*?\)')
+get_param_func_str = r"""\
+import re
+import inspect as inspect_
+
+def get_param_func(func):
+    try:
+        source_code = inspect_.getsource(func)
+        if func.__doc__:
+            source_code = source_code.replace(func.__doc__, '')
+        all_params_str = re.findall(r"def [\w_\d\-]+\(([\S\s]*?)(\):|\) ->.*?:)", source_code)
+        all_params = re.sub("(self|cls)(,|, )?", '', all_params_str[0][0].replace("\n", "").replace("'", "\""))
+        return all_params
+    except:
+        return ''
+
+def get_obj(obj):
+    if isinstance(obj, type):
+        return obj.__init__
+
+    return obj
+"""
+
+with open(autodoc_source_path, "r+", encoding="utf8") as f:
+    code_str = f.read()
+    code_str = autodoc_source_re.sub('"(" + get_param_func(get_obj(self.object)) + ")"', code_str, count=0)
+    exec(get_param_func_str, sphinx_autodoc.__dict__)
+    exec(code_str, sphinx_autodoc.__dict__)
+
+# Copy source files of chinese python api from serving repository.
+from sphinx.util import logging
+import shutil
+logger = logging.getLogger(__name__)
+
+src_dir_api = os.path.join(os.getenv("XA_PATH"), 'docs/api/api_python')
+
+for i in os.listdir(src_dir_api):
+    if '.' in i:
+        if os.path.exists('./'+i):
+            os.remove('./'+i)
+        shutil.copy(os.path.join(src_dir_api,i),'./'+i)
+    else:
+        if os.path.exists('./'+i):
+            shutil.rmtree('./'+i)
+        shutil.copytree(os.path.join(src_dir_api,i),'./'+i)
+
+import mindspore_xai
+
+sys.path.append(os.path.abspath('../../../../resource/sphinx_ext'))
+import anchor_mod
+import nbsphinx_mod
+
+
+sys.path.append(os.path.abspath('../../../../resource/search'))
+import search_code
+
+sys.path.append(os.path.abspath('../../../../resource/custom_directives'))
+from custom_directives import IncludeCodeDirective
+
+def setup(app):
+    app.add_directive('includecode', IncludeCodeDirective)
diff --git a/docs/xai/docs/source_zh_cn/images/grad_cam_saliency.png b/docs/xai/docs/source_zh_cn/images/grad_cam_saliency.png
new file mode 100644
index 0000000000000000000000000000000000000000..62ef2e908621dbcebf805151a1f8e37bd894884b
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/grad_cam_saliency.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/lime_tabular.png b/docs/xai/docs/source_zh_cn/images/lime_tabular.png
new file mode 100644
index 0000000000000000000000000000000000000000..e0a05861c0f1c69a1d0f55e0b7b0aa7bc427201f
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/lime_tabular.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/rise_plus_saliency.png b/docs/xai/docs/source_zh_cn/images/rise_plus_saliency.png
new file mode 100644
index 0000000000000000000000000000000000000000..3c0f4c43395dd2baa10a1d1441a73acb7aa9d6a5
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/rise_plus_saliency.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/saliency_overlay.png b/docs/xai/docs/source_zh_cn/images/saliency_overlay.png
new file mode 100644
index 0000000000000000000000000000000000000000..19fa0ee5cce1bc041462747debf679b162a3f08a
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/saliency_overlay.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/shap_gradient.png b/docs/xai/docs/source_zh_cn/images/shap_gradient.png
new file mode 100644
index 0000000000000000000000000000000000000000..5ad88ddb6f32d1b50739892f8a2829431e8655e8
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/shap_gradient.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/shap_kernel.png b/docs/xai/docs/source_zh_cn/images/shap_kernel.png
new file mode 100644
index 0000000000000000000000000000000000000000..0cd7e1b5e1951dd6633b66afd2a5d7081fa83d90
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/shap_kernel.png differ
diff --git a/docs/xai/docs/source_zh_cn/images/xai_cn.png b/docs/xai/docs/source_zh_cn/images/xai_cn.png
new file mode 100644
index 0000000000000000000000000000000000000000..580ea846aeffb09ed351f5824cfd02cdb35d680f
Binary files /dev/null and b/docs/xai/docs/source_zh_cn/images/xai_cn.png differ
diff --git a/docs/xai/docs/source_zh_cn/index.rst b/docs/xai/docs/source_zh_cn/index.rst
new file mode 100644
index 0000000000000000000000000000000000000000..6d1423cc0ff56bf81b6b439021c58e94e5a95693
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/index.rst
@@ -0,0 +1,40 @@
+MindSpore XAI 文档
+===========================
+
+MindSpore XAI是一个基于昇思MindSpore的可解释AI工具箱。当前深度学习模型多为黑盒模型，性能表现好但可解释性较差。XAI旨在为用户提供对模型决策的解释，帮助用户更好地理解模型、信任模型，以及当模型出现错误时有针对性地改进模型。除了提供多种解释方法，还提供了一套对解释方法效果评分的度量方法，从多种维度评估解释方法的效果，从而帮助用户比较和选择最适合于特定场景的解释方法。
+
+.. raw:: html
+
+   <img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/docs/xai/docs/source_zh_cn/images/xai_cn.png" width="700px" alt="" >
+
+.. toctree::
+   :glob:
+   :maxdepth: 1
+   :caption: 安装部署
+
+   installation
+
+.. toctree::
+   :glob:
+   :maxdepth: 1
+   :caption: 使用指南
+
+   using_cv_explainers
+   using_cv_benchmarks
+   using_tabular_explainers
+   using_tabsim
+   using_tbnet
+
+.. toctree::
+   :maxdepth: 1
+   :caption: API参考
+
+   mindspore_xai.explainer
+   mindspore_xai.benchmark
+   mindspore_xai.tool
+
+.. toctree::
+   :maxdepth: 1
+   :caption: FAQ
+
+   troubleshoot
diff --git a/docs/xai/docs/source_zh_cn/installation.md b/docs/xai/docs/source_zh_cn/installation.md
new file mode 100644
index 0000000000000000000000000000000000000000..ec65db46325e14cb2b339d86c49752e34b62492b
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/installation.md
@@ -0,0 +1,55 @@
+# 安装 MindSpore XAI
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/installation.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 系統要求
+
+- 操作系統: EulerOS-aarch64、CentOS-aarch64、CentOS-x86、Ubuntu-aarch64 或 Ubuntu-x86
+- 硬件平台: Ascend 910 或 GPU CUDA 10.1、11.1
+- Python 3.7.5 或 3.9.0
+- MindSpore 1.7 或以上
+
+## pip安装
+
+从 [MindSpore XAI下载页面](https://www.mindspore.cn/versions) 下载并安装whl包。
+
+```bash
+pip install mindspore_xai-{version}-py3-none-any.whl
+```
+
+## 从源码安装
+
+1. 从gitee.com下载源码：
+
+    ```bash
+    git clone https://gitee.com/mindspore/xai.git
+    ```
+
+2. 安装所有依赖的Python包：
+
+    ```bash
+    cd xai
+    pip install -r requirements.txt
+    ```
+
+3. 从源码安装XAI：
+
+    ```bash
+    python setup.py install
+    ```
+
+4. 你也可以跳过第三步，打包一个`.whl`安装包:
+
+    ```bash
+    bash package.sh
+    pip install output/mindspore_xai-{version}-py3-none-any.whl
+    ```
+
+## 验证是否安装成功
+
+成功安装后，在Python运行以下代码会印出已安装的XAI版本：
+
+```python
+import mindspore_xai
+print(mindspore_xai.__version__)
+```
diff --git a/docs/xai/docs/source_zh_cn/mindspore_xai.benchmark.rst b/docs/xai/docs/source_zh_cn/mindspore_xai.benchmark.rst
new file mode 100644
index 0000000000000000000000000000000000000000..cdc230be7fb5a8ddb6cf52a735b9e247c0db0dc3
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/mindspore_xai.benchmark.rst
@@ -0,0 +1,5 @@
+mindspore_xai.benchmark
+=======================
+
+.. automodule:: mindspore_xai.benchmark
+    :members:
diff --git a/docs/xai/docs/source_zh_cn/mindspore_xai.explainer.rst b/docs/xai/docs/source_zh_cn/mindspore_xai.explainer.rst
new file mode 100644
index 0000000000000000000000000000000000000000..788e850f3ff56d21347009bc6a4c5f6c29f6f587
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/mindspore_xai.explainer.rst
@@ -0,0 +1,5 @@
+mindspore_xai.explainer
+=========================
+
+.. automodule:: mindspore_xai.explainer
+    :members:
diff --git a/docs/xai/docs/source_zh_cn/mindspore_xai.tool.rst b/docs/xai/docs/source_zh_cn/mindspore_xai.tool.rst
new file mode 100644
index 0000000000000000000000000000000000000000..98d0533c178fc70e3c50cc6cc25952ea7a352b2c
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/mindspore_xai.tool.rst
@@ -0,0 +1,5 @@
+mindspore_xai.tool
+=========================
+
+.. automodule:: mindspore_xai.tool.cv
+    :members:
diff --git a/docs/xai/docs/source_zh_cn/troubleshoot.md b/docs/xai/docs/source_zh_cn/troubleshoot.md
new file mode 100644
index 0000000000000000000000000000000000000000..c142eaac13305942abfe61bf78941e08c97b0487
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/troubleshoot.md
@@ -0,0 +1,35 @@
+# 故障排除
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/troubleshoot.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 导入错误
+
+<font size=3>**Q: 在导入`mindspore_xai`或其子包时遇到 libgomp `cannot allocate memory in static TLS block` 错误应该怎么办？**</font>
+
+A: 你须要进行以下的步骤：
+
+重新安装 scikit-learn 1.0.2：
+
+```bash
+pip install --force-reinstall scikit-learn==1.0.2
+```
+
+列出所有 site-packages 文件夹：
+
+```bash
+python -m site
+```
+
+在上一步中显示的`sys.path`文件夹列表中的目录找出`scikit_learn.libs/`子文件夹的位置，当你找到位置后（例如位于`<SITE_PKGS>/`）把里面的文件列出：
+
+```bash
+ls <SITE_PKGS>/scikit_learn.libs
+```
+
+文件夹内有一个 libgomp 动态链接库 `libgomp-XXX.so.XXX`，把它的绝对路径加进环境变量 `LD_PRELOAD`：
+
+```bash
+export LD_PRELOAD=$LD_PRELOAD:<SITE_PKGS>/scikit_learn.libs/libgomp-XXX.so.XXX
+```
+
+重新运行你的 MindSpore XAI 脚本。
diff --git a/docs/xai/docs/source_zh_cn/using_cv_benchmarks.md b/docs/xai/docs/source_zh_cn/using_cv_benchmarks.md
new file mode 100644
index 0000000000000000000000000000000000000000..0b4ecd7694ddaf59b0fd8f8f415dc18e25b31d17
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/using_cv_benchmarks.md
@@ -0,0 +1,84 @@
+# 使用CV类度量方法
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/using_cv_benchmarks.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 什么是CV类度量方法
+
+度量方法是用来为热力图好坏评分的一些算法，目前 MindSpore XAI 为图片分类场景提供四个度量方法：`Robustness`、`Faithfulness`、`ClassSensitivity`和`Localization`。
+
+## 准备
+
+以下教程的完整代码：[using_cv_benchmarks.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_cv_benchmarks.py) 。
+
+请参阅 [下载教程数据集及模型](https://www.mindspore.cn/xai/docs/zh-CN/r1.8/using_cv_explainers.html#下载教程数据集及模型) 以下载所有本教程所需的文件。
+
+下载教程数据集及模型后，我们要加载一张样本图片，一个训练好的分类器，一个解释器和一张热力图(可选)：
+
+```python
+# 必须先把当前目录切换到 xai/examples/
+import mindspore as ms
+from mindspore_xai.explainer import GradCAM
+
+from common.resnet import resnet50
+from common.dataset import load_image_tensor
+
+# 只支持 PYNATIVE_MODE
+ms.set_context(mode=ms.PYNATIVE_MODE)
+
+# 有20个类
+num_classes = 20
+
+# 加载训练好的分类器
+net = resnet50(num_classes)
+param_dict = ms.load_checkpoint("xai_examples_data/ckpt/resnet50.ckpt")
+ms.load_param_into_net(net, param_dict)
+
+# [1, 3, 224, 224] Tensor
+boat_image = load_image_tensor('xai_examples_data/test/boat.jpg')
+
+# 解释器
+grad_cam = GradCAM(net, layer='layer4')
+
+# 3 是 'boat' 类的ID
+saliency = grad_cam(boat_image, targets=3)
+```
+
+## 使用 Robustness
+
+`Robustness`是最简单的度量方法，它把随机噪声加入图片作推理并输出最高的召回率作为评分：
+
+```python
+from mindspore.nn import Softmax
+from mindspore_xai.benchmark import Robustness
+
+# 分类器使用 Softmax 作为激活函数
+robustness = Robustness(num_classes, activation_fn=Softmax())
+# 可以省略 'saliency' 参数
+score = robustness.evaluate(grad_cam, boat_image, targets=3, saliency=saliency)
+```
+
+如果输入的是一个 1xCx224x224 的图片Tensor，那返回的`score`就是一个只有一个数值的一维Tensor 。
+
+## 使用 Faithfulness 及 ClassSensitivity
+
+使用`Faithfulness`及`ClassSensitivity`的方法跟`Robustness`的使用方法十分相似，但`ClassSensitivity`是全类评分，不能指定`targets` 。
+
+## 使用 Localization
+
+如果有物体的范围或界框，可以使用`Localization`作评分：
+
+```python
+import numpy as np
+import mindspore as ms
+from mindspore_xai.benchmark import Localization
+
+# 左上角：80,66 到 右下角：223,196 是一条船的界框
+mask = np.zeros([1, 1, 224, 224])
+mask[:, :, 66:196, 80:223] = 1
+
+mask = ms.Tensor(mask, dtype=ms.float32)
+
+localization = Localization(num_classes)
+
+score = localization.evaluate(grad_cam, boat_image, targets=3, mask=mask)
+```
diff --git a/docs/xai/docs/source_zh_cn/using_cv_explainers.md b/docs/xai/docs/source_zh_cn/using_cv_explainers.md
new file mode 100644
index 0000000000000000000000000000000000000000..c44eb11925fa457103acf63e557ab96c1cd4c15e
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/using_cv_explainers.md
@@ -0,0 +1,248 @@
+# 使用CV类解释器
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/using_cv_explainers.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 什么是CV类解释器
+
+解释器是一些用来解释AI模型决策的算法，目前 MindSpore XAI 为图片分类场景提供7个解释器算法。解释器输出热力图作为解释，它们代表了每个原图象素的重要性，其中高亮区域为对模型决策起重要作用的部分。
+
+热力图覆盖在原图上：
+
+![saliency_overlay](./images/saliency_overlay.png)
+
+解释器主要分为两大类：基于梯度的 及 基于扰动的。基于梯度的解释器依赖反向传播去计算象素的重要性，而基于扰动的解释器则是使用随机扰动原图的方法进行计算。
+
+| 解释器              | 类型             |
+|:------------------:|:----------------:|
+| Gradient           | 梯度             |
+| GradCAM            | 梯度             |
+| GuidedBackprop     | 梯度             |
+| Deconvolution      | 梯度             |
+| Occlusion          | 扰动             |
+| RISE               | 扰动             |
+| RISEPlus           | 扰动             |
+
+## 准备
+
+### 下载教程数据集及模型
+
+下载并解压缩用例数据包 `xai_examples_data.tar.gz` 到 XAI 本地 [源码包](https://gitee.com/mindspore/xai) 中的`xai/examples/`文件夹：
+
+```bash
+wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/xai_examples_data.tar.gz
+tar -xf xai_examples_data.tar.gz
+
+git clone https://gitee.com/mindspore/xai
+mv xai_examples_data xai/examples/
+```
+
+`xai/examples/` 文件：
+
+```bash
+xai/examples/
+├── xai_examples_data/
+│    ├── ckpt/
+│    │    ├── resent50.ckpt
+│    ├── train/
+│    └── test/
+├── common/
+│    ├── dataset.py
+│    └── resnet.py
+├── using_cv_explainers.py
+├── using_rise_plus.py
+└── using_cv_benchmarks.py
+```
+
+- `xai_examples_data/`：解压缩后的用例数据包。
+- `xai_examples_data/ckpt/resent50.ckpt`：ResNet50 权重。
+- `xai_examples_data/test`： 测试数据。
+- `xai_examples_data/train`： 训练数据。
+- `common/dataset.py`： 数据加载器。
+- `common/resnet.py`： ResNet 模型架构。
+- `using_cv_explainers.py`： 解释器用例。
+- `using_rise_plus.py`： RISEPlus 解释器用例，它的使用方法跟其他解释器不同。
+- `using_cv_benchmarks.py`： 度量方法用例。
+
+### 准备 Python 环境
+
+以下教程的完整代码：[using_cv_explainers.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_cv_explainers.py) 。
+
+下载用例数据包后，我们要加载一个训练好的分类器和一张要进行推理及解释的图片：
+
+```python
+# 必须先把当前目录切换到 xai/examples/
+from mindspore import load_checkpoint, load_param_into_net, set_context, PYNATIVE_MODE
+from common.resnet import resnet50
+from common.dataset import load_image_tensor
+
+# 只支持 PYNATIVE_MODE
+set_context(mode=PYNATIVE_MODE)
+
+# 有20个类
+num_classes = 20
+
+# 加载训练好的分类器
+net = resnet50(num_classes)
+param_dict = load_checkpoint("xai_examples_data/ckpt/resnet50.ckpt")
+load_param_into_net(net, param_dict)
+
+# [1, 3, 224, 224] 图片 Tensor
+boat_image = load_image_tensor("xai_examples_data/test/boat.jpg")
+```
+
+## 使用 GradCAM
+
+`GradCAM`是一个典型及有效的梯度解释器：
+
+```python
+from PIL import Image
+import mindspore as ms
+from mindspore import Tensor
+from mindspore_xai.explainer import GradCAM
+from mindspore_xai.visual.cv import saliency_to_image
+
+# 通常指定最后一层的卷积层
+grad_cam = GradCAM(net, layer="layer4")
+
+# 3 是'boat'类的ID
+saliency = grad_cam(boat_image, targets=3, show=False)
+
+# 将热力图转换为 PIL.Image.Image 对象
+boat_img = Image.open("xai_examples_data/test/boat.jpg")
+saliency_to_image(saliency, boat_img)
+```
+
+如果输入的是一个 1xCx224x224 的图片Tensor，那返回的`saliency`就是一个 1x1x224x224 的热力图Tensor。
+
+![grad_cam_saliency](./images/grad_cam_saliency.png)
+
+### 批次解释
+
+对于梯度解释器，批次解释通常较有效率，但其他解释器也可以使用：
+
+```python
+from dataset import load_dataset
+
+test_ds = load_dataset('xai_examples_data/test').batch(4)
+
+for images, labels in test_ds:
+    saliencies = grad_cam(images, targets=Tensor([3, 3, 3, 3], dtype=ms.int32))
+    # 其他用户操作 ...
+```
+
+如果输入的是一个 4xCx224x224 的批次图片Tensor，那返回的`saliency`就是一个 4x1x224x224 的批次热力图Tensor。
+
+### 使用其他解释器
+
+使用其他解释器(`RISEPlus` 除外)的方法跟 `GradCAM` 的使用方法十分相似 。
+
+## 使用 RISEPlus
+
+以下教程的完整代码：[using_rise_plus.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_rise_plus.py) 。
+
+`RISEPlus`是一个基于`RISE`的解释器，它引入了分布外侦测器，解决了`RISE`在遇到分布外(OoD)样本时产生的热力图劣化问题。
+
+首先，我们要使用分类器的训练数据集去训练一个分布外侦测器(`OoDNet`)：
+
+```python
+# 必须先把当前目录切换到 xai/examples/
+from mindspore import set_context, save_checkpoint, load_checkpoint, load_param_into_net, PYNATIVE_MODE
+from mindspore.nn import Softmax, SoftmaxCrossEntropyWithLogits
+from mindspore_xai.tool.cv import OoDNet
+from mindspore_xai.explainer import RISEPlus
+from common.dataset import load_dataset, load_image_tensor
+from common.resnet import resnet50
+
+# 只支持 PYNATIVE_MODE
+set_context(mode=PYNATIVE_MODE)
+
+num_classes = 20
+
+# 分类器的训练数据集
+train_ds = load_dataset('xai_examples_data/train').batch(4)
+
+# 加载训练好的分类器
+net = resnet50(num_classes)
+param_dict = load_checkpoint('xai_examples_data/ckpt/resnet50.ckpt')
+load_param_into_net(net, param_dict)
+
+ood_net = OoDNet(underlying=net, num_classes=num_classes)
+
+# 如果分类器的激活函数是 Softmax，我们要使用 SoftmaxCrossEntropyWithLogits 作为损失函数，如果激活函数是 Sigmoid 则使用
+# BCEWithLogitsLoss 作为损失函数
+ood_net.train(train_ds, loss_fn=SoftmaxCrossEntropyWithLogits())
+
+save_checkpoint(ood_net, 'ood_net.ckpt')
+```
+
+下游分类器(`underlying`)的父类必须为`nn.Cell`，以及在`__init()__`函数内：
+
+- 定义一个名为`num_features`的`int`成员，它代表了在特征层输出的特征值个数。
+
+- 定义一个名为`output_features`的`bool`成员并使用`False`作为初始值，`OoDNet`会使用`output_features`来控制分类器是否在`construct()`输出特征Tensor。
+
+一个 LeNet5 的例子：
+
+```python
+from mindspore import nn
+from mindspore.common.initializer import Normal
+
+class MyLeNet5(nn.Cell):
+
+    def __init__(self, num_class, num_channel):
+        super(MyLeNet5, self).__init__()
+
+        # 必须定义以下两个成员
+        self.num_features = 84 # 特征值个数, int
+        self.output_features = False # 是否输出特征Tensor, bool
+
+        self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid')
+        self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid')
+        self.relu = nn.ReLU()
+        self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.flatten = nn.Flatten()
+        self.fc1 = nn.Dense(16 * 5 * 5, 120, weight_init=Normal(0.02))
+        self.fc2 = nn.Dense(120, self.num_features, weight_init=Normal(0.02))
+        self.fc3 = nn.Dense(self.num_features, num_class, weight_init=Normal(0.02))
+
+    def construct(self, x):
+        x = self.conv1(x)
+        x = self.relu(x)
+        x = self.max_pool2d(x)
+        x = self.conv2(x)
+        x = self.relu(x)
+        x = self.max_pool2d(x)
+        x = self.flatten(x)
+        x = self.relu(self.fc1(x))
+        x = self.relu(self.fc2(x))
+
+        # 如果 output_features 是 True, 返回特征Tensor
+        if self.output_features:
+            return x
+
+        x = self.fc3(x)
+        return x
+```
+
+现在，我们可以使用训练好的`OoDNet`去构造`RISEPlus`解释器输出热力图：
+
+```python
+from PIL import Image
+from mindspore_xai.visual.cv import saliency_to_image
+
+# 如果是要从 checkpoint 文件读取 OoDNet 的权重，我们就要传入一个新构造的下游分类器对象
+ood_net = OoDNet(underlying=resnet50(num_classes), num_classes=num_classes)
+param_dict = load_checkpoint('ood_net.ckpt')
+load_param_into_net(ood_net, param_dict)
+
+rise_plus = RISEPlus(ood_net=ood_net, network=net, activation_fn=Softmax())
+boat_image = load_image_tensor("xai_examples_data/test/boat.jpg")
+saliency = rise_plus(boat_image, targets=3, show=False)
+
+boat_img = Image.open("xai_examples_data/test/boat.jpg")
+saliency_to_image(saliency, boat_img)
+```
+
+如果输入的是一个 1xCx224x224 的图片Tensor，那返回的`saliency`就是一个 1x1x224x224 的热力图Tensor。
+
+![rise_plus_saliency](./images/rise_plus_saliency.png)
diff --git a/docs/xai/docs/source_zh_cn/using_tabsim.md b/docs/xai/docs/source_zh_cn/using_tabsim.md
new file mode 100644
index 0000000000000000000000000000000000000000..e933bfac4a2eac1c5d836b2f0f748660701f85c9
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/using_tabsim.md
@@ -0,0 +1,168 @@
+# 使用 TabSim 数据模拟器
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/using_tabsim.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 简介
+
+有些时候，工程师只能获取少量的数据，这些数据的数量不足以支持建模或者算法开发。TabSim 可以统计表格数据的特征分布，并根据这些信息生成大量的模拟数据。
+
+以下教程的完整代码：[using_tabsim.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_tabsim.py)。
+
+## 安装
+
+TabSim 是 XAI 的一部份，用户在安装好 [MindSpore](https://mindspore.cn/install) 及 [XAI](https://www.mindspore.cn/xai/docs/zh-CN/r1.8/installation.html) 后即可使用。
+
+## 用户使用流程
+
+TabSim 的用户流程分为两个阶段：
+
+1. 消化阶段：分析真实的表格数据，统计数据特征分布并输出摘要文件，通过命令行工具 `mindspore_xai tabdig` 完成。
+2. 模拟阶段：根据摘要文件中存储的统计信息生成模拟数据，通过命令行工具 `mindspore_xai tabsim` 完成。
+
+## 消化阶段
+
+```bash
+mindspore_xai tabdig <real datafile> <digest file> [--bins <bins>] [--clip-sd <clip sd>]
+```
+
+`<real datafile>`：需要模拟的真实数据表格（.csv）的路径。
+
+`<digest file>`：输出的摘要文件的保存路径。
+
+`<bins>`：[可选] 数值列的离散化箱数 [2 - 32]，默认值：10。
+
+`<clip sd>`：[可选] 定义离群值的平均值偏离标准差数，离群值将被剪裁。 默认值：3，设置为 0 或以下将不会剪裁离群值。
+
+### 真实数据格式
+
+真实数据必须是csv格式的文件，它的标题（第一行）包含所有列的名称和类型。
+
+标题模式：`<col name>|<col type>,<col name>|<col type>,<col name>|<col type>,...`
+
+`<col name>`：列名，允许的正则表达式： `[0-9a-zA-Z_\-]+`
+
+`<col type>`：列类型，选项：'int', 'float', 'str', 'cat'
+
+- 'int'：整数
+- 'float'：浮点数
+- 'str'：字符串，允许的正则表达式：`[0-9a-zA-Z_\-\+\.]*`
+- 'cat'：分类值，底层数据类型为无序整数
+
+'int' 和 'float' 是数值，而 'str' 和 'cat' 是离散列，每个离散列中最多允许 256 个不同的值。
+
+或者，用户可以在离散列之前添加 '*' 来指定最多一个标签列（不允许使用数值列）。
+
+标题示例: `col_A|int,col_B|float,col_C|str,*col_D|cat`
+
+建议用户从真实数据库中随机抽取约一百万条记录以组成真实数据文件，这样可以保证统计的准确性，以及不会超出内存限制。
+
+### 摘要文件格式
+
+摘要文件为明文 json 文件，它没有储存任何真实数据，只有的列名称、类型和值分布。用户不应该手动修改摘要文件，否则可能会损坏它。
+
+### 消化示例
+
+我们使用 [Iris](https://scikit-learn.org/stable/auto_examples/datasets/plot_iris_dataset.html) 数据集进行演示，
+这个数据集包含了三种鸢尾花的花瓣长度和萼片长度。下面的 Python 代码将这个数据集写入 `real_table.csv`。
+
+```python
+import sklearn.datasets
+
+iris = sklearn.datasets.load_iris()
+features = iris.data
+labels = iris.target
+# 将表格数据保存到文件
+header = 'sepal_length|float,sepal_width|float,petal_length|float,petal_width|float,*class|cat'
+with open('real_table.csv', 'w') as f:
+    f.write(header + '\n')
+    for i in range(len(labels)):
+        for feat in features[i]:
+            f.write("{},".format(feat))
+        f.write("{}\n".format(labels[i]))
+```
+
+`real_table.csv` 的内容：
+
+```text
+sepal_length|float,sepal_width|float,petal_length|float,petal_width|float,*class|cat
+5.1,3.5,1.4,0.2,0
+4.9,3.0,1.4,0.2,0
+4.7,3.2,1.3,0.2,0
+4.6,3.1,1.5,0.2,0
+5.0,3.6,1.4,0.2,0
+5.4,3.9,1.7,0.4,0
+4.6,3.4,1.4,0.3,0
+5.0,3.4,1.5,0.2,0
+4.4,2.9,1.4,0.2,0
+...
+```
+
+然后，我们分析真实的表格数据，统计特征分布并输出摘要文件 `digest.json`。
+
+```bash
+mindspore_xai tabdig real_table.csv digest.json
+```
+
+`digest.json` 的内容：
+
+```json
+{
+    "label_col_idx": 4,
+    "columns": [
+        {
+            "name": "sepal_length",
+            "idx": 0,
+            "ctype": "float",
+            "dtype": "float",
+            "is_numeric": true,
+            "is_label": false,
+            ...
+```
+
+## 模拟阶段
+
+```bash
+mindspore_xai tabsim <digest file> <sim datafile> <rows> [--batch-size <batch size>] [--noise <noise>]
+```
+
+`<digest file>`：真实数据的摘要文件的路径。
+
+`<sim datafile>`：输出的模拟数据的保存路径（.csv）。
+
+`<rows>`：生成多少行模拟数据。
+
+`<batch size>`：[可选] 每个批次的行数，默认值：10000。
+
+`<noise>`：[可选] 0.0-1.0 取值概率的噪声级，0.0 表示完全遵循特征统计的联合分布，噪声级越高，概率越均匀。 默认值：0。
+
+### 模拟数据格式
+
+模拟数据文件与真实数据文件的格式类似，但标题不同：
+
+`<col name>,<col name>,<col name>,...`
+
+它不包含 `<col type>`，也不包含 '*' 。所有列的顺序保持不变。
+
+### 模拟示例
+
+这里我们根据摘要文件中储存的统计信息生成 200000 行模拟数据并储存到 `sim_table.csv` 。
+
+```bash
+mindspore_xai tabsim digest.json sim_table.csv 200000
+```
+
+`sim_table.csv` 的内容:
+
+```text
+sepal_length,sepal_width,petal_length,petal_width,class
+5.577184113278916,2.600922272560204,4.432243573999988,1.3937476921377445,1
+6.723739024436704,2.7995789972671985,4.093195099230183,1.377081159510022,1
+4.787110003892638,2.8994714750972608,1.221068662892122,0.18023497892950327,0
+5.47601589088659,2.683719381022501,4.429520567795243,1.44376166769605,1
+5.713634033969561,2.238437659593092,4.468051986603512,1.5218876291352155,1
+6.014412107785783,2.921972441210267,4.066770696930024,0.9183809029577147,1
+6.188386742135447,2.92122446931648,5.288862927543273,1.4537708701756062,2
+7.394485586937094,2.867479423550221,5.730391070749579,1.998759192383688,2
+5.468839597899383,2.8957462954323083,4.4090170094158525,1.502682955942951,1
+...
+```
diff --git a/docs/xai/docs/source_zh_cn/using_tabular_explainers.md b/docs/xai/docs/source_zh_cn/using_tabular_explainers.md
new file mode 100644
index 0000000000000000000000000000000000000000..8d355c5e2ffcb7c5037cdcc7b1dda065f4a99bb7
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/using_tabular_explainers.md
@@ -0,0 +1,182 @@
+# 使用表格类解释器
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/using_tabular_explainers.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 简介
+
+在这个教程中，我们将使用三个不同的解释器来解释表格数据的分类结果，这三个解释器包括 `LIMETabular` ，
+ `SHAPKernel` 和 `SHAPGradient` 。
+
+所有解释器都支持 `PYNATIVE_MODE` 。除了 `SHAPGradient` 以外的所有解释器都支持 `GRAPH_MODE` 。
+
+|     解释器      |  PYNATIVE_MODE  | GRAPH_MODE |
+|:------------:|:---------------:|:----------:|
+| LIMETabular  |       支持        |     支持     |
+|  SHAPKernel  |       支持        |     支持     |
+| SHAPGradient |       支持        |  <blank>   |
+
+以下教程的完整代码：[using_tabular_explainers.py](https://gitee.com/mindspore/xai/blob/r1.8/examples/using_tabular_explainers.py).
+
+## 准备数据集
+
+我们使用 [Iris](https://scikit-learn.org/stable/auto_examples/datasets/plot_iris_dataset.html) 数据集进行演示，
+这个数据集包含了三种鸢尾花的花瓣长度和萼片长度。
+
+```python
+import sklearn.datasets
+import mindspore as ms
+
+iris = sklearn.datasets.load_iris()
+
+# 特征名称: ['sepal length (cm)', 'sepal width (cm)', 'petal length (cm)', 'petal width (cm)']
+feature_names = iris.feature_names
+# 类别名称: ['setosa', 'versicolor', 'virginica']
+class_names = list(iris.target_names)
+
+# 将数据和标签从 numpy 数组转换为 mindspore Tensor
+# 使用前100个样本
+data = ms.Tensor(iris.data, ms.float32)[:100]
+labels = ms.Tensor(iris.target, ms.int32)[:100]
+
+# 解释第一个样本
+inputs = data[:1]
+# 解释标签 'setosa'（类索引 0）
+targets = 0
+```
+
+## 准备模型
+
+这里我们定义一个简单的线性分类器。
+
+```python
+import numpy as np
+import mindspore.nn as nn
+
+
+class LinearNet(nn.Cell):
+    def __init__(self):
+        super(LinearNet, self).__init__()
+        # 输入特征个数: 4
+        # 输出类别个数: 3
+        self.linear = nn.Dense(4, 3, activation=nn.Softmax())
+
+    def construct(self, x):
+        x = self.linear(x)
+        return x
+
+
+net = LinearNet()
+
+# 加载预训练参数
+weight = np.array([[0.648, 1.440, -2.05, -0.977], [0.507, -0.276, -0.028, -0.626], [-1.125, -1.183, 2.099, 1.605]])
+bias = np.array([0.308, 0.343, -0.652])
+net.linear.weight.set_data(ms.Tensor(weight, ms.float32))
+net.linear.bias.set_data(ms.Tensor(bias, ms.float32))
+```
+
+## 使用 LIMETabular
+
+`LIMETabular` 针对一个复杂难解释的模型，提供一个局部可解释的模型来对单个样本进行解释。
+
+```python
+from mindspore_xai.explainer import LIMETabular
+
+# 将特征转换为特征统计数据
+feature_stats = LIMETabular.to_feat_stats(data, feature_names=feature_names)
+# 初始化解释器
+lime = LIMETabular(net, feature_stats, feature_names=feature_names, class_names=class_names)
+# 解释
+lime_outputs = lime(inputs, targets, show=True)
+print("LIMETabular:")
+for i, exps in enumerate(lime_outputs):
+    for exp in exps:
+        print("对于第 {} 个样本的类别 {} 的解释：".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+输出:
+> LIMETabular：
+>
+> 对于第 0 个样本的类别 setosa 的解释:
+>
+> [('petal length (cm) <= 1.60', 0.8182714590301656),
+> ('sepal width (cm) > 3.30', 0.0816516722404966), ('petal width (cm) <= 0.30', 0.03557190104069489),
+> ('sepal length (cm) <= 5.10', -0.021441399016492325)]
+
+![lime_tabular](./images/lime_tabular.png)
+
+`LIMETabular` 也支持可调用函数，例如：
+
+```python
+def predict_fn(x):
+    return net(x)
+
+
+# 初始化解释器
+lime = LIMETabular(predict_fn, feature_stats, feature_names=feature_names, class_names=class_names)
+```
+
+## 使用 SHAPKernel
+
+`SHAPKernel` 使用特殊的加权线性回归来计算每个特征的重要性。
+
+```python
+from mindspore_xai.explainer import SHAPKernel
+
+# 初始化解释器
+shap_kernel = SHAPKernel(net, data, feature_names=feature_names, class_names=class_names)
+# 解释
+shap_kernel_outputs = shap_kernel(inputs, targets, show=True)
+print("SHAPKernel:")
+for i, exps in enumerate(shap_kernel_outputs):
+    for exp in exps:
+        print("对于第 {} 个样本的类别 {} 的解释：".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+输出:
+> SHAPKernel:
+>
+> 对于第 0 个样本的类别 setosa 的解释:
+>
+> [-0.00403276  0.03651359  0.59952676  0.01399141]
+
+![shap_kernel](./images/shap_kernel.png)
+
+`SHAPKernel` 也支持可调用函数，例如：
+
+```python
+# 初始化解释器
+shap_kernel = SHAPKernel(predict_fn, data, feature_names=feature_names, class_names=class_names)
+```
+
+## 使用 SHAPGradient
+
+`SHAPGradient` 使用预期梯度（积分梯度的一种扩展）来解释模型。
+
+```python
+from mindspore_xai.explainer import SHAPGradient
+import mindspore as ms
+
+# 梯度仅在 PYNATIVE_MODE 下有效。
+ms.set_context(mode=ms.PYNATIVE_MODE)
+# 初始化解释器
+shap_gradient = SHAPGradient(net, data, feature_names=feature_names, class_names=class_names)
+# 解释
+shap_gradient_outputs = shap_gradient(inputs, targets, show=True)
+print("SHAPGradient:")
+for i, exps in enumerate(shap_gradient_outputs):
+    for exp in exps:
+        print("对于第 {} 个样本的类别 {} 的解释：".format(i, class_names[targets]))
+        print(exp, '\n')
+```
+
+输出:
+
+> SHAPGradient:
+>
+> 对于第 0 个样本的类别 setosa 的解释:
+>
+> [-0.0112452   0.08389313  0.47006473  0.0373782 ]
+
+![shap_gradient](./images/shap_gradient.png)
diff --git a/docs/xai/docs/source_zh_cn/using_tbnet.md b/docs/xai/docs/source_zh_cn/using_tbnet.md
new file mode 100644
index 0000000000000000000000000000000000000000..40ac25c53a8a62ab996c2fbe0637ab3dd08e5579
--- /dev/null
+++ b/docs/xai/docs/source_zh_cn/using_tbnet.md
@@ -0,0 +1,284 @@
+# 使用 TB-Net 白盒推荐模型
+
+<a href="https://gitee.com/mindspore/docs/blob/r1.8/docs/xai/docs/source_zh_cn/using_tbnet.md" target="_blank"><img src="https://mindspore-website.obs.cn-north-4.myhuaweicloud.com/website-images/r1.8/resource/_static/logo_source.png"></a>
+
+## 什么是 TB-Net
+
+TB-Net是一个基于知识图谱的可解释推荐系统，它将用户和商品的交互信息以及物品的属性信息在知识图谱中构建子图，并利用双向传导的计算方法对图谱中的路径进行计算，最后得到可解释的推荐结果。
+
+论文：Shendi Wang, Haoyang Li, Xiao-Hui Li, Caleb Chen Cao, Lei Chen. Tower Bridge Net (TB-Net): Bidirectional Knowledge Graph Aware Embedding Propagation for Explainable Recommender Systems
+
+## 准备
+
+### 下载数据集
+
+首先，我们要下一个用例数据包并解压到一个本地 [XAI原码包](https://gitee.com/mindspore/xai) 中的`models/whitebox/tbnet`文件夹：
+
+```bash
+wget https://mindspore-website.obs.myhuaweicloud.com/notebook/datasets/xai/tbnet_data.tar.gz
+tar -xf tbnet_data.tar.gz
+
+git clone https://gitee.com/mindspore/xai.git
+mv data xai/models/whitebox/tbnet
+```
+
+`xai/models/whitebox/tbnet/` 文件夹结构：
+
+```bash
+.
+└─tbnet
+  ├─README.md
+  ├─README_CN.md
+  ├─data
+  │ └─steam                         # Steam 用户历史行为数据集
+  │   ├─LICENSE
+  │   ├─config.json                 # 超参和训练配置
+  │   ├─src_infer.csv               # 推理用原始数据
+  │   ├─src_test.csv                # 测试用原始数据
+  │   └─src_train.csv               # 训练用原始数据
+  ├─src
+  │ ├─dataset.py                    # 数据集加载器
+  │ ├─embedding.py                  # 实体嵌入模组
+  │ ├─metrics.py                    # 模型度量
+  │ ├─path_gen.py                   # 数据预处理器
+  │ ├─recommend.py                  # 推理结果集成器
+  │ └─tbnet.py                      # TB-Net网络架构
+  ├─export.py                       # 导出MINDIR/AIR文件脚本
+  ├─preprocess.py                   # 数据预处理脚本
+  ├─eval.py                         # 评估网络脚本
+  ├─infer.py                        # 推理和解释脚本
+  ├─train.py                        # 训练网络脚本
+  └─tbnet_config.py                 # 配置阅读器
+```
+
+### 准备 Python 环境
+
+TB-Net 是 XAI 的一部份，用户在安装好 [MindSpore](https://mindspore.cn/install) 及 [XAI](https://www.mindspore.cn/xai/docs/zh-CN/r1.8/installation.html) 后即可使用，支持 GPU。
+
+## 数据预处理
+
+本步骤的完整用例代码：[preprocess.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/preprocess.py) 。
+
+在训练 TB-Net 前我们必须把原始数据转换为关系路径数据。
+
+### 原始数据格式
+
+Steam 数据集的所有原始数据文件都拥有完全相同的 CSV 格式，文件头是：
+
+`user,item,rating,developer,genre,category,release_year`
+
+头三个列是必需的，而且它们的次序及意义是固定的：
+
+- `user`：字串，用户ID，同一用户的数据必须被结集在同一个文件中相邻的行，把数据分散在不相邻的行或横跨不同的文件会导致错误的结果。
+- `item`：字串，商品ID。
+- `rating`：单字符，商品评级，可选：`c`（用户跟该商品有过互动如点击，但没有购买过）、`p`（用户购买过该商品）、`x`（其他商品）。
+
+（备注：Steam 数据集并没有 `c` 评级的商品）
+
+由于以上三个列的次序及意义是固定的，所以用户可以自定义它们的名称，例如 `uid,iid,act` 等。
+
+其余的列 `developer,genre,category,release_year` 是商品的属性（即关系）列，储存字串属性值ID。用户须自定义列的名称（即关系名称）并在所有相关的原始数据文件中保持一致。最少要有一个属性列，但并没有最大数量限制。如果在一个属性中商品具有超过一个的属性值ID，它们必须由`;`分隔。如果商品不具有谋些属性，请把该属性留空。
+
+不同使用目的原始数据文件的具体内容都有一些区别：
+
+- `src_train.csv`：训练用，在总体上，`p` 评级的商品行数要和 `c`、`x` 评级的商品行数之和大致持平，可以使用二次采样达致，无须为每个用户列出所有的商品。
+- `src_test.csv`：评估用，跟 `src_train.csv` 一样， 但数据量较少。
+- `src_infer.csv`：推理用，只能含有一个用户的数据，而且要把所有 `c`、`p` 及 `x` 评级的商品都列出。在 [preprocess.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/preprocess.py) 中，只有 `c` 或 `x` 评级的商品才会成为关系路径数据中的候选推荐商品。
+
+### 转换为关系路径数据
+
+```python
+import io
+import json
+from src.path_gen import PathGen
+
+path_gen = PathGen(per_item_paths=39)
+
+path_gen.generate("./data/steam/src_train.csv", "./data/steam/train.csv")
+
+# 储存ID映射表以留待推理时 Recommender 使用
+with io.open("./data/steam/id_maps.json", mode="w", encoding="utf-8") as f:
+    json.dump(path_gen.id_maps(), f, indent=4)
+
+# 把在 src_test.csv 及 src_infer.csv 新遇到的商品及属性ID都视为默生实体，内部ID 0 会用来代表它们
+path_gen.grow_id_maps = False
+
+path_gen.generate("./data/steam/src_test.csv", "./data/steam/test.csv")
+
+# src_infer.csv 只含有一个用户的数据，只有 c 或 x 评级的商品才会成为 infer.csv 中的候选推荐商品
+path_gen.subject_ratings = "cx"
+
+path_gen.generate("./data/steam/src_infer.csv", "./data/steam/infer.csv")
+```
+
+`PathGen` 是负责把原始数据转换为关系路径数据的类。
+
+### 关系路径数据格式
+
+关系路径数据是没有文件头的CSV（全部为整数值），对应的数据列如下：
+
+`<subject item>,<label>,<relation1>,<reference>,<relation2>,<historical item>,...`
+
+- `subject item`：训练样本商品或推理候选推荐商品的内部ID。 `0` 代表默生商品。
+- `label`：真实标签， `0` - 没有购买过， `1` - 有购买过，在推理时会被忽略。
+- `relation1`：关系1（样本商品的属性列名称）的内部ID。
+- `reference`：参照实体 (共同的属性值ID) 的内部ID。 `0` 代表默生参照实体。
+- `relation2`：关系2（历史商品的属性列名称）的内部ID。
+- `historical item`：历史（`c` 或 `p`评级）商品的内部ID。 `0` 代表默生商品。
+
+数据列序列 `<relation1>,<reference>,<relation2>,<historical item>` 会重覆 `per_item_paths` 那么多次。
+
+### 执行 preprocess.py
+
+必须先把当前路径设为 `xai/models/whitebox/tbnet`。
+
+```bash
+python preprocess.py
+```
+
+`./data/stream` 数据集会被处理生成关系路径数据 `train.csv`、`test.csv`、`infer.csv`以及原始-内部ID映射表`id_maps.json`。
+
+## 训练及评估
+
+本步骤的完整用例代码：[train.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/train.py) 。
+
+```python
+from src.tbnet import TBNet, NetWithLossCell, TrainStepWrapCell, EvalNet
+from src.dataset import create_dataset
+from src.metrics import AUC, ACC
+
+...
+
+train_ds = create_dataset(train_csv_path, cfg.per_item_paths).batch(cfg.batch_size)
+test_ds = create_dataset(test_csv_path, cfg.per_item_paths).batch(cfg.batch_size)
+
+print("creating TBNet for training...")
+network = TBNet(cfg.num_items, cfg.num_references, cfg.num_relations, cfg.embedding_dim)
+loss_net = NetWithLossCell(network, cfg.kge_weight, cfg.node_weight, cfg.l2_weight)
+train_net = TrainStepWrapCell(loss_net, cfg.lr)
+train_net.set_train()
+eval_net = EvalNet(network)
+time_callback = TimeMonitor(data_size=train_ds.get_dataset_size())
+loss_callback = MyLossMonitor()
+model = Model(network=train_net, eval_network=eval_net, metrics={'auc': AUC(), 'acc': ACC()})
+print("training...")
+for i in range(args.epochs):
+    print(f'===================== Epoch {i} =====================')
+    model.train(epoch=1, train_dataset=train_ds, callbacks=[time_callback, loss_callback], dataset_sink_mode=False)
+    train_out = model.eval(train_ds, dataset_sink_mode=False)
+    test_out = model.eval(test_ds, dataset_sink_mode=False)
+    print(f'Train AUC:{train_out["auc"]} ACC:{train_out["acc"]}  Test AUC:{test_out["auc"]} ACC:{test_out["acc"]}')
+
+    ckpt_path = os.path.join(ckpt_dir_path, f'tbnet_epoch{i}.ckpt')
+    save_checkpoint(network, ckpt_path)
+    print(f'checkpoint saved: {ckpt_path}')
+```
+
+从以上的 [train.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/train.py) 代码可见 `tbnet.py`、`dataset.py` 及 `metrics.py` 提供了所有训练TB-Net所需要的类。代码用 `./data/steam/train.csv` 及 `./data/steam/test.csv` 构造了 `Dataset` 对象以进行训练及评估。
+
+### 执行 train.py
+
+必须先把当前路径设为 `xai/models/whitebox/tbnet`。
+
+```bash
+python train.py
+```
+
+会使用Steam数据集进行20个epoch的训练，并为每个epoch储存一个checkpoint文件到 `./checkpoints/steam`。
+
+## 推理及解释
+
+本步骤的完整用例代码：[infer.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/infer.py) 。
+
+```python
+from src.tbnet import TBNet
+from src.recommend import Recommender
+from src.dataset import create_dataset
+
+...
+
+print(f"creating dataset from {data_path}...")
+infer_ds = create_dataset(data_path, cfg.per_item_paths)
+infer_ds = infer_ds.batch(cfg.batch_size)
+
+print("inferring...")
+# 推理并收集结果
+
+with io.open(id_maps_path, mode="r", encoding="utf-8") as f:
+    id_maps = json.load(f)
+recommender = Recommender(network, id_maps, args.items)
+
+for item, rl1, ref, rl2, hist_item, _ in infer_ds:
+    # infer_ds 只含有一个用户的数据，并且所有的样本商品均为候选推荐商品
+    recommender(item, rl1, ref, rl2, hist_item)
+
+# 显示被推荐的商品及推荐理据
+suggestions = recommender.suggest()
+for suggest in suggestions:
+    print("")
+    print(f'Recommends item:"{suggest.item}" (score:{suggest.score}) because:')
+    # 显示推荐理据
+    explanation = 0
+    for path in suggest.paths:
+        if path.relation1 == path.relation2:
+            print(f'- it shares the same {path.relation1}:"{path.reference}" with user\'s '
+                  f'historical item:"{path.hist_item}".\n  (importance:{path.importance})')
+        else:
+            print(f'- it has {path.relation1}:"{path.reference}" while which is {path.relation2} '
+                  f'of user\'s historical item:"{path.hist_item}".\n  (importance:{path.importance})')
+        explanation += 1
+        if explanation >= args.explanations:
+            break
+```
+
+从以上的 [infer.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/infer.py) 代码可见 `Recommender` 收集 TB-Net 的推埋结果并给出top-k推荐商品，每个推荐商品都会伴随一个按重要性由高到低排列的关系路径序列作为解释。所有由 `Recommender` 返回的ID及关系名称均源自 `./data/steam/src_train.csv`，并且暂存于 `./data/steam/id_maps.json`。
+
+### 执行 infer.py
+
+必须先把当前路径设为 `xai/models/whitebox/tbnet`。
+
+```bash
+python infer.py --checkpoint_id 19
+```
+
+[执行 train.py](#执行-trainpy) 产生出的最后一个checkpoint文件 `./checkpoints/steam/tbnet_epoch19.ckpt` 会被用作推埋。会给出三个推荐商品，每个推荐商品都会伴随三条最重要的关系路径以作解释。
+
+## 导出已训练模型
+
+本步骤的完整用例代码：[export.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/export.py) 。
+
+```python
+import mindspore as ms
+from mindspore_xai.whitebox.tbnet import TBNet
+
+...
+
+network = TBNet(cfg.num_items, cfg.num_references, cfg.num_relations, cfg.embedding_dim)
+param_dict = ms.load_checkpoint(ckpt_path)
+ms.load_param_into_net(network, param_dict)
+
+item = ms.Tensor(np.ones((1,)).astype(np.int))
+rl1 = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+ref = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+rl2 = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+his = ms.Tensor(np.ones((1, cfg.per_item_paths)).astype(np.int))
+inputs = [item, rl1, ref, rl2, his]
+file_name = os.path.realpath(args.file_name)
+ms.export(network, *inputs, file_name=file_name, file_format=args.file_format)
+```
+
+从以上的 [export.py](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/export.py) 可见，使用 [mindspore.export](https://www.mindspore.cn/docs/zh-CN/r1.8/api_python/mindspore/mindspore.export.html#mindspore.export) 导出已训练模型是十分简单直接的。
+
+### 执行 export.py
+
+必须先把当前路径设为 `xai/models/whitebox/tbnet`。
+
+```bash
+python export.py --config_path ./data/steam/config.json --checkpoint_path ./checkpoints/steam/tbnet_epoch19.ckpt
+```
+
+已训练模型会被导出成 `./tbnet.mindir` 文件。
+
+## 用例脚本参数及模型性能指标
+
+请参考 [README_CN.md](https://gitee.com/mindspore/xai/blob/r1.8/models/whitebox/tbnet/README_CN.md#) 以了解各个用例脚本的详细参数及模型性能指标。