From ee5dd799c0a4deda8e91710cfa581ad7c056201d Mon Sep 17 00:00:00 2001
From: WMC <846995292@qq.com>
Date: Mon, 30 May 2022 20:30:42 +0800
Subject: [PATCH 1/5] update torch version

---
 .../cv/classification/SENet154/README.md      |  7 +++---
 .../classification/SENet154/test/run_2onnx.sh |  2 +-
 .../SENet154/test/train_eval_8p.sh            |  2 +-
 .../SENet154/test/train_full_1p.sh            |  2 +-
 .../SENet154/test/train_full_8p.sh            |  2 +-
 .../SENet154/test/train_performance_1p.sh     |  2 +-
 .../SENet154/test/train_performance_8p.sh     |  2 +-
 .../cv/classification/SENet154/train.py       | 22 +++++++++----------
 8 files changed, 20 insertions(+), 21 deletions(-)

diff --git a/PyTorch/contrib/cv/classification/SENet154/README.md b/PyTorch/contrib/cv/classification/SENet154/README.md
index 3a719047ec..9da2ec1fb0 100755
--- a/PyTorch/contrib/cv/classification/SENet154/README.md
+++ b/PyTorch/contrib/cv/classification/SENet154/README.md
@@ -42,8 +42,8 @@ Label smoothing is required for qualified model accuracy.
     4. `--batch-size N`：分配个每个设备的batch大小
 3. 开始训练
    ```
-   bash ./test/train_full_1p.sh  --data_path=数据集路径    # 精度训练
-   bash ./test/train_performance_1p.sh  --data_path=数据集路径  # 性能训练
+   bash ./test/train_full_8p.sh  --data_path=数据集路径    # 精度训练
+   bash ./test/train_performance_8p.sh  --data_path=数据集路径  # 性能训练
    ```
 
 ### 训练结果
@@ -67,5 +67,4 @@ $E$为当前一轮的Epoch序号，从0开始
 ### NPU 8p
 |Epochs|Learning rate                         |Optimization type|FPS     |Acc@1 |Acc@5 |
 |:----:|:------------------------------------:|:---------------:|:------:|:----:|:----:|
-|120   |$0.6\times 0.1^{\lfloor E/30 \rfloor}$|O2               |1524.537|78.599|93.849|
-|120   |$0.3\times (1+\cos{\frac{E\pi}{120}})$|O2               |1522.120|80.048|94.799|
\ No newline at end of file
+|120   |$0.3\times (1+\cos{\frac{E\pi}{120}})$|O2               |1022.920|80.564|94.428|
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh b/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
index 949af6d740..38cec26b8f 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
@@ -1,3 +1,3 @@
 source test/env_npu.sh
 
-python3.7.5 pth2onnx.py "$@" checkpoint-120.pth.rar
\ No newline at end of file
+python3.7 pth2onnx.py "$@" checkpoint-120.pth.rar
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
index 18aa841d70..5e47f1196c 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
@@ -75,7 +75,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 train.py \
+taskset -c $pid_start-$pid_end python3.7 train.py \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
index e6e42752b8..d68cf96ec4 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
@@ -78,7 +78,7 @@ if [ x"${etp_flag}" != x"true" ];then
     source ${test_path_dir}/env_npu.sh
 fi
 kernel_num=$(($(nproc) / 8))
-python3.7.5 -u train.py "$@" \
+python3.7 -u train.py "$@" \
     --num-workers $kernel_num \
     --data $data_path \
     --device npu:$ASCEND_DEVICE_ID \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
index de36d5a226..eedaa7aeb7 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
@@ -74,7 +74,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 -u train.py "$@" \
+taskset -c $pid_start-$pid_end python3.7 -u train.py "$@" \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
index 27f4881c05..18c1844a5a 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
@@ -78,7 +78,7 @@ if [ x"${etp_flag}" != x"true" ];then
     source ${test_path_dir}/env_npu.sh
 fi
 kernel_num=$(($(nproc) / 8))
-python3.7.5 -u train.py "$@" \
+python3.7 -u train.py "$@" \
     --num-workers $kernel_num \
     --data $data_path \
     --device npu:$ASCEND_DEVICE_ID \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
index cc9e364d11..d857dd3d1b 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
@@ -74,7 +74,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 -u train.py "$@" \
+taskset -c $pid_start-$pid_end python3.7 -u train.py "$@" \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/train.py b/PyTorch/contrib/cv/classification/SENet154/train.py
index 02f9927ea6..2e5e5e2f7a 100755
--- a/PyTorch/contrib/cv/classification/SENet154/train.py
+++ b/PyTorch/contrib/cv/classification/SENet154/train.py
@@ -25,7 +25,8 @@ import torch.distributed as dist
 import torch.nn as nn 
 import torch.nn.parallel as par
 import torch.optim as optim
-
+if torch.__version__ >= '1.8.1':
+    import torch_npu
 sys.path.append('.')
 import checkpoint
 import data
@@ -80,11 +81,11 @@ def main_worker(args):
             pretrained_dict.pop('module.last_linear.weight')
             pretrained_dict.pop('module.last_linear.bias')
 
-        for param in net.parameters():
+        for param in model.parameters():
             param.requires_gard = False
 
-        net.last_linear = nn.Linear(2048, 1000)
-        net.load_state_dict(pretrained_dict, strict=False)
+        model.last_linear = nn.Linear(2048, 1000)
+        model.load_state_dict(pretrained_dict, strict=False)
 
     if args.label_smoothing_epsilon > 0.0:
         logging.info('Using label smoothing with epsilon = {:.3f}'.format(args.label_smoothing_epsilon))
@@ -126,11 +127,11 @@ def main_worker(args):
     _, val_loader = data.create_val_loader(
         model, val_dir, args, scale, distributed=args.distributed
     )
-
+    
     model, optimizer = apex.amp.initialize(
         model, optimizer, 
         opt_level=args.opt_level, 
-        loss_scale=args.loss_scale
+        loss_scale=args.loss_scale if args.loss_scale != -1 else 'dynamic'
     )
     if args.distributed:
         model = par.DistributedDataParallel(model, device_ids=[args.local_rank])
@@ -196,8 +197,8 @@ def train(train_loader, model, criterion, optimizer, epoch_id, args):
         input = input.to(args.device, non_blocking=True)
         target = target.to(args.device, non_blocking=True)
         # compute output
-        output = model(input)
-        loss = criterion(output, target)
+        output = model(input)         
+        loss = criterion(output, target)      
         # measure accuracy and record loss 
         prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
         losses.update(loss.data.item(), input.size(0))
@@ -419,8 +420,8 @@ def parse_args():
     parser.add_argument('--opt-level', default='O2', type=str, choices=['O1', 'O2'],
                         help='optimization level of amp.initialize (default: \'O2\')')
 
-    parser.add_argument('--loss-scale', default=128, type=int,
-                        help='loss scale of amp.initialize (default: 128)')
+    parser.add_argument('--loss-scale', default=-1, type=int,
+                        help='loss scale of amp.initialize (default:128), -1 means dynamic')
 
     parser.add_argument('--epochs', default=100, type=int, metavar='N',
                         help='number of total epochs to run (default: 100)')
@@ -498,7 +499,6 @@ def parse_args():
 if __name__ == '__main__':
     args = parse_args()
     if 'npu' in args.device:
-        import torch.npu 
         if args.distributed:
             os.environ['MASTER_ADDR'] = '127.0.0.1'
             os.environ['MASTER_PORT'] = '23333'
-- 
Gitee


From 8b5653bde116a1f9eb6b785c601f153889f6a8e1 Mon Sep 17 00:00:00 2001
From: WMC <846995292@qq.com>
Date: Wed, 1 Jun 2022 13:49:03 +0800
Subject: [PATCH 2/5] update

---
 PyTorch/contrib/cv/classification/SENet154/train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/PyTorch/contrib/cv/classification/SENet154/train.py b/PyTorch/contrib/cv/classification/SENet154/train.py
index 2e5e5e2f7a..00fe6d17e5 100755
--- a/PyTorch/contrib/cv/classification/SENet154/train.py
+++ b/PyTorch/contrib/cv/classification/SENet154/train.py
@@ -131,7 +131,7 @@ def main_worker(args):
     model, optimizer = apex.amp.initialize(
         model, optimizer, 
         opt_level=args.opt_level, 
-        loss_scale=args.loss_scale if args.loss_scale != -1 else 'dynamic'
+        loss_scale='dynamic'
     )
     if args.distributed:
         model = par.DistributedDataParallel(model, device_ids=[args.local_rank])
-- 
Gitee


From 0836051cbdbcc955bf7eeb34e93e177fbec696e8 Mon Sep 17 00:00:00 2001
From: WMC <846995292@qq.com>
Date: Mon, 30 May 2022 20:30:42 +0800
Subject: [PATCH 3/5] update torch version

---
 .../cv/classification/SENet154/README.md      |  7 +++---
 .../classification/SENet154/test/run_2onnx.sh |  2 +-
 .../SENet154/test/train_eval_8p.sh            |  2 +-
 .../SENet154/test/train_full_1p.sh            |  2 +-
 .../SENet154/test/train_full_8p.sh            |  2 +-
 .../SENet154/test/train_performance_1p.sh     |  2 +-
 .../SENet154/test/train_performance_8p.sh     |  2 +-
 .../cv/classification/SENet154/train.py       | 22 +++++++++----------
 8 files changed, 20 insertions(+), 21 deletions(-)

diff --git a/PyTorch/contrib/cv/classification/SENet154/README.md b/PyTorch/contrib/cv/classification/SENet154/README.md
index 3a719047ec..9da2ec1fb0 100755
--- a/PyTorch/contrib/cv/classification/SENet154/README.md
+++ b/PyTorch/contrib/cv/classification/SENet154/README.md
@@ -42,8 +42,8 @@ Label smoothing is required for qualified model accuracy.
     4. `--batch-size N`：分配个每个设备的batch大小
 3. 开始训练
    ```
-   bash ./test/train_full_1p.sh  --data_path=数据集路径    # 精度训练
-   bash ./test/train_performance_1p.sh  --data_path=数据集路径  # 性能训练
+   bash ./test/train_full_8p.sh  --data_path=数据集路径    # 精度训练
+   bash ./test/train_performance_8p.sh  --data_path=数据集路径  # 性能训练
    ```
 
 ### 训练结果
@@ -67,5 +67,4 @@ $E$为当前一轮的Epoch序号，从0开始
 ### NPU 8p
 |Epochs|Learning rate                         |Optimization type|FPS     |Acc@1 |Acc@5 |
 |:----:|:------------------------------------:|:---------------:|:------:|:----:|:----:|
-|120   |$0.6\times 0.1^{\lfloor E/30 \rfloor}$|O2               |1524.537|78.599|93.849|
-|120   |$0.3\times (1+\cos{\frac{E\pi}{120}})$|O2               |1522.120|80.048|94.799|
\ No newline at end of file
+|120   |$0.3\times (1+\cos{\frac{E\pi}{120}})$|O2               |1022.920|80.564|94.428|
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh b/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
index 949af6d740..38cec26b8f 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/run_2onnx.sh
@@ -1,3 +1,3 @@
 source test/env_npu.sh
 
-python3.7.5 pth2onnx.py "$@" checkpoint-120.pth.rar
\ No newline at end of file
+python3.7 pth2onnx.py "$@" checkpoint-120.pth.rar
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
index 18aa841d70..5e47f1196c 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_eval_8p.sh
@@ -75,7 +75,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 train.py \
+taskset -c $pid_start-$pid_end python3.7 train.py \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
index e6e42752b8..d68cf96ec4 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
@@ -78,7 +78,7 @@ if [ x"${etp_flag}" != x"true" ];then
     source ${test_path_dir}/env_npu.sh
 fi
 kernel_num=$(($(nproc) / 8))
-python3.7.5 -u train.py "$@" \
+python3.7 -u train.py "$@" \
     --num-workers $kernel_num \
     --data $data_path \
     --device npu:$ASCEND_DEVICE_ID \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
index de36d5a226..eedaa7aeb7 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_full_8p.sh
@@ -74,7 +74,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 -u train.py "$@" \
+taskset -c $pid_start-$pid_end python3.7 -u train.py "$@" \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
index 27f4881c05..18c1844a5a 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_1p.sh
@@ -78,7 +78,7 @@ if [ x"${etp_flag}" != x"true" ];then
     source ${test_path_dir}/env_npu.sh
 fi
 kernel_num=$(($(nproc) / 8))
-python3.7.5 -u train.py "$@" \
+python3.7 -u train.py "$@" \
     --num-workers $kernel_num \
     --data $data_path \
     --device npu:$ASCEND_DEVICE_ID \
diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
index cc9e364d11..d857dd3d1b 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_performance_8p.sh
@@ -74,7 +74,7 @@ kernel_num=$(($(nproc) / 8))
 pid_start=$((kernel_num * rank))
 pid_end=$((pid_start + kernel_num - 1))
 
-taskset -c $pid_start-$pid_end python3.7.5 -u train.py "$@" \
+taskset -c $pid_start-$pid_end python3.7 -u train.py "$@" \
     --distributed \
     --num-devices 8 \
     --local-rank $rank \
diff --git a/PyTorch/contrib/cv/classification/SENet154/train.py b/PyTorch/contrib/cv/classification/SENet154/train.py
index 02f9927ea6..2e5e5e2f7a 100755
--- a/PyTorch/contrib/cv/classification/SENet154/train.py
+++ b/PyTorch/contrib/cv/classification/SENet154/train.py
@@ -25,7 +25,8 @@ import torch.distributed as dist
 import torch.nn as nn 
 import torch.nn.parallel as par
 import torch.optim as optim
-
+if torch.__version__ >= '1.8.1':
+    import torch_npu
 sys.path.append('.')
 import checkpoint
 import data
@@ -80,11 +81,11 @@ def main_worker(args):
             pretrained_dict.pop('module.last_linear.weight')
             pretrained_dict.pop('module.last_linear.bias')
 
-        for param in net.parameters():
+        for param in model.parameters():
             param.requires_gard = False
 
-        net.last_linear = nn.Linear(2048, 1000)
-        net.load_state_dict(pretrained_dict, strict=False)
+        model.last_linear = nn.Linear(2048, 1000)
+        model.load_state_dict(pretrained_dict, strict=False)
 
     if args.label_smoothing_epsilon > 0.0:
         logging.info('Using label smoothing with epsilon = {:.3f}'.format(args.label_smoothing_epsilon))
@@ -126,11 +127,11 @@ def main_worker(args):
     _, val_loader = data.create_val_loader(
         model, val_dir, args, scale, distributed=args.distributed
     )
-
+    
     model, optimizer = apex.amp.initialize(
         model, optimizer, 
         opt_level=args.opt_level, 
-        loss_scale=args.loss_scale
+        loss_scale=args.loss_scale if args.loss_scale != -1 else 'dynamic'
     )
     if args.distributed:
         model = par.DistributedDataParallel(model, device_ids=[args.local_rank])
@@ -196,8 +197,8 @@ def train(train_loader, model, criterion, optimizer, epoch_id, args):
         input = input.to(args.device, non_blocking=True)
         target = target.to(args.device, non_blocking=True)
         # compute output
-        output = model(input)
-        loss = criterion(output, target)
+        output = model(input)         
+        loss = criterion(output, target)      
         # measure accuracy and record loss 
         prec1, prec5 = accuracy(output.data, target, topk=(1, 5))
         losses.update(loss.data.item(), input.size(0))
@@ -419,8 +420,8 @@ def parse_args():
     parser.add_argument('--opt-level', default='O2', type=str, choices=['O1', 'O2'],
                         help='optimization level of amp.initialize (default: \'O2\')')
 
-    parser.add_argument('--loss-scale', default=128, type=int,
-                        help='loss scale of amp.initialize (default: 128)')
+    parser.add_argument('--loss-scale', default=-1, type=int,
+                        help='loss scale of amp.initialize (default:128), -1 means dynamic')
 
     parser.add_argument('--epochs', default=100, type=int, metavar='N',
                         help='number of total epochs to run (default: 100)')
@@ -498,7 +499,6 @@ def parse_args():
 if __name__ == '__main__':
     args = parse_args()
     if 'npu' in args.device:
-        import torch.npu 
         if args.distributed:
             os.environ['MASTER_ADDR'] = '127.0.0.1'
             os.environ['MASTER_PORT'] = '23333'
-- 
Gitee


From c66f24fc1a403550521d2b4e6418d92a027d36ce Mon Sep 17 00:00:00 2001
From: WMC <846995292@qq.com>
Date: Wed, 1 Jun 2022 13:49:03 +0800
Subject: [PATCH 4/5] update

---
 PyTorch/contrib/cv/classification/SENet154/train.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/PyTorch/contrib/cv/classification/SENet154/train.py b/PyTorch/contrib/cv/classification/SENet154/train.py
index 2e5e5e2f7a..00fe6d17e5 100755
--- a/PyTorch/contrib/cv/classification/SENet154/train.py
+++ b/PyTorch/contrib/cv/classification/SENet154/train.py
@@ -131,7 +131,7 @@ def main_worker(args):
     model, optimizer = apex.amp.initialize(
         model, optimizer, 
         opt_level=args.opt_level, 
-        loss_scale=args.loss_scale if args.loss_scale != -1 else 'dynamic'
+        loss_scale='dynamic'
     )
     if args.distributed:
         model = par.DistributedDataParallel(model, device_ids=[args.local_rank])
-- 
Gitee


From ceff51072d904fbb28ee851fda02c0edb5644cc1 Mon Sep 17 00:00:00 2001
From: WMC <846995292@qq.com>
Date: Wed, 1 Jun 2022 14:07:32 +0800
Subject: [PATCH 5/5] init

---
 .../contrib/cv/classification/SENet154/test/train_full_1p.sh    | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
index d68cf96ec4..1e695a4b06 100755
--- a/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
+++ b/PyTorch/contrib/cv/classification/SENet154/test/train_full_1p.sh
@@ -87,7 +87,7 @@ python3.7 -u train.py "$@" \
     --scheduler cosine \
     --label-smoothing-epsilon 0.1 \
     --batch-size $batch_size \
-    --print-freq 10  ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
+    --print-freq 10 > ${test_path_dir}/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log 2>&1 &
 
 wait
 
-- 
Gitee