!1592 bug fix while evaluation

Merge pull request !1592 from SanjayChan/r0.3

mindspore/model_zoo/mobilenetv2/Readme.md

@@ -133,7 +133,7 @@ result: {'acc': 0.71976314102564111} ckpt=/path/to/checkpoint/mobilenet-200_625.
 
 #### Inference Performance
 
-| Parameters                 | GoogLeNet                     |                           |                      |
+| Parameters                 |                               |                           |                      |
 | -------------------------- | ----------------------------- | ------------------------- | -------------------- |
 | Model Version              | V1                            |                           |                      |
 | Resource                   | Huawei 910                    | NV SMX2 V100-32G          | Huawei 310           |

mindspore/model_zoo/mobilenetv2/eval.py

@@ -36,21 +36,23 @@ args_opt = parser.parse_args()
 
 if __name__ == '__main__':
     config_platform = None
+    net = None
     if args_opt.platform == "Ascend":
         config_platform = config_ascend
         device_id = int(os.getenv('DEVICE_ID'))
         context.set_context(mode=context.GRAPH_MODE, device_target="Ascend",
                             device_id=device_id, save_graphs=False)
+        net = mobilenet_v2(num_classes=config_platform.num_classes, platform="Ascend")
     elif args_opt.platform == "GPU":
         config_platform = config_gpu
         context.set_context(mode=context.GRAPH_MODE,
                             device_target="GPU", save_graphs=False)
+        net = mobilenet_v2(num_classes=config_platform.num_classes, platform="GPU")
     else:
         raise ValueError("Unsupport platform.")
 
     loss = nn.SoftmaxCrossEntropyWithLogits(
         is_grad=False, sparse=True, reduction='mean')
-    net = mobilenet_v2(num_classes=config_platform.num_classes)
 
     if args_opt.platform == "Ascend":
         net.to_float(mstype.float16)

mindspore/model_zoo/mobilenetv2/scripts/run_infer.sh

@@ -52,4 +52,4 @@ python ${BASEPATH}/../eval.py \
         --platform=$1 \
         --dataset_path=$2 \
         --checkpoint_path=$3 \
-        &> infer.log &  # dataset val folder path
+        &> ../infer.log &  # dataset val folder path

mindspore/model_zoo/mobilenetv2/scripts/run_train.sh

@@ -36,13 +36,13 @@ run_ascend()
     fi
     mkdir ../train
     cd ../train || exit
-    python ${BASEPATH}/../launch.py \
+    python ${BASEPATH}/../src/launch.py \
             --nproc_per_node=$2 \
             --visible_devices=$4 \
             --server_id=$3 \
-            --training_script=${BASEPATH}/train.py \
+            --training_script=${BASEPATH}/../train.py \
             --dataset_path=$5 \
-            --platform=$1 &> train.log &  # dataset train folder
+            --platform=$1 &> ../train.log &  # dataset train folder
 }
 
 run_gpu()
@@ -73,7 +73,7 @@ run_gpu()
     python ${BASEPATH}/../train.py \
         --dataset_path=$4 \
         --platform=$1 \
-        &> train.log &  # dataset train folder
+        &> ../train.log &  # dataset train folder
 }
 
 if [ $# -gt 5 ] || [ $# -lt 4 ]

mindspore/model_zoo/mobilenetv2/src/dataset.py

@@ -21,7 +21,6 @@ import mindspore.dataset.engine as de
 import mindspore.dataset.transforms.vision.c_transforms as C
 import mindspore.dataset.transforms.c_transforms as C2
 
-
 def create_dataset(dataset_path, do_train, config, platform, repeat_num=1, batch_size=32):
     """
     create a train or eval dataset
@@ -29,8 +28,8 @@ def create_dataset(dataset_path, do_train, config, platform, repeat_num=1, batch
     Args:
         dataset_path(string): the path of dataset.
         do_train(bool): whether dataset is used for train or eval.
-        repeat_num(int): the repeat times of dataset. Default: 1
-        batch_size(int): the batch size of dataset. Default: 32
+        repeat_num(int): the repeat times of dataset. Default: 1.
+        batch_size(int): the batch size of dataset. Default: 32.
 
     Returns:
         dataset
@@ -44,7 +43,12 @@ def create_dataset(dataset_path, do_train, config, platform, repeat_num=1, batch
             ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True,
                                          num_shards=rank_size, shard_id=rank_id)
     elif platform == "GPU":
-        ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)
+        if do_train:
+            from mindspore.communication.management import get_rank, get_group_size
+            ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True,
+                                         num_shards=get_group_size(), shard_id=get_rank())
+        else:
+            ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)
     else:
         raise ValueError("Unsupport platform.")
 

mindspore/model_zoo/mobilenetv2/train.py

@@ -32,7 +32,7 @@ from mindspore.train.model import Model, ParallelMode
 from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, Callback
 from mindspore.train.loss_scale_manager import FixedLossScaleManager
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_group_size
 import mindspore.dataset.engine as de
 from src.dataset import create_dataset
 from src.lr_generator import get_lr
@@ -146,7 +146,7 @@ class Monitor(Callback):
         self.losses.append(step_loss)
         cur_step_in_epoch = (cb_params.cur_step_num - 1) % cb_params.batch_num
 
-        print("epoch: [{:3d}/{:3d}], step:[{:5d}/{:5d}], loss:[{:5.3f}/{:5.3f}], time:[{:5.3f}], lr:[{:5.3f}]".format(
+        print("epoch: [{:3d}/{:3d}], step:[{:5d}/{:5d}], loss:[{:5.3f}/{:5.3f}], time:[{:5.3f}], lr:[{:5.5f}]".format(
             cb_params.cur_epoch_num -
             1, cb_params.epoch_num, cur_step_in_epoch, cb_params.batch_num, step_loss,
             np.mean(self.losses), step_mseconds, self.lr_init[cb_params.cur_step_num - 1]))
@@ -157,6 +157,11 @@ if __name__ == '__main__':
         # train on gpu
         print("train args: ", args_opt, "\ncfg: ", config_gpu)
 
+        init('nccl')
+        context.set_auto_parallel_context(parallel_mode="data_parallel",
+                                          mirror_mean=True,
+                                          device_num=get_group_size())
+
         # define net
         net = mobilenet_v2(num_classes=config_gpu.num_classes, platform="GPU")
         # define loss
@@ -216,14 +221,14 @@ if __name__ == '__main__':
             init()
 
         epoch_size = config_ascend.epoch_size
-        net = mobilenet_v2(num_classes=config_ascend.num_classes)
+        net = mobilenet_v2(num_classes=config_ascend.num_classes, platform="Ascend")
         net.to_float(mstype.float16)
         for _, cell in net.cells_and_names():
             if isinstance(cell, nn.Dense):
                 cell.to_float(mstype.float32)
         if config_ascend.label_smooth > 0:
             loss = CrossEntropyWithLabelSmooth(
-                smooth_factor=config_ascend.label_smooth, num_classes=config.num_classes)
+                smooth_factor=config_ascend.label_smooth, num_classes=config_ascend.num_classes)
         else:
             loss = SoftmaxCrossEntropyWithLogits(
                 is_grad=False, sparse=True, reduction='mean')

mindspore/model_zoo/mobilenetv3/Readme.md

@@ -133,7 +133,7 @@ result: {'acc': 0.71976314102564111} ckpt=/path/to/checkpoint/mobilenet-200_625.
 
 #### Inference Performance
 
-| Parameters                 | GoogLeNet                     |                           |                      |
+| Parameters                 |                               |                           |                      |
 | -------------------------- | ----------------------------- | ------------------------- | -------------------- |
 | Model Version              | V1                            |                           |                      |
 | Resource                   | Huawei 910                    | NV SMX2 V100-32G          | Huawei 310           |

mindspore/model_zoo/mobilenetv3/eval.py

@@ -24,7 +24,8 @@ from mindspore.train.serialization import load_checkpoint, load_param_into_net
 from mindspore.common import dtype as mstype
 from src.dataset import create_dataset
 from src.config import config_ascend, config_gpu
-from src.mobilenetV2 import mobilenet_v2
+from src.mobilenetV3 import mobilenet_v3_large
+
 
 parser = argparse.ArgumentParser(description='Image classification')
 parser.add_argument('--checkpoint_path', type=str, default=None, help='Checkpoint file path')
@@ -49,7 +50,7 @@ if __name__ == '__main__':
 
     loss = nn.SoftmaxCrossEntropyWithLogits(
         is_grad=False, sparse=True, reduction='mean')
-    net = mobilenet_v2(num_classes=config_platform.num_classes)
+    net = mobilenet_v3_large(num_classes=config_platform.num_classes)
 
     if args_opt.platform == "Ascend":
         net.to_float(mstype.float16)

mindspore/model_zoo/mobilenetv3/scripts/run_infer.sh

@@ -42,14 +42,14 @@ export RANK_ID=0
 export RANK_SIZE=1
 if [ -d "eval" ];
 then
-    rm -rf ./eval
+    rm -rf ../eval
 fi
-mkdir ./eval
-cd ./eval || exit
+mkdir ../eval
+cd ../eval || exit
 
 # luanch
-python ${BASEPATH}/eval.py \
+python ${BASEPATH}/../eval.py \
         --platform=$1 \
         --dataset_path=$2 \
         --checkpoint_path=$3 \
-        &> infer.log &  # dataset val folder path
+        &> ../infer.log &  # dataset val folder path

mindspore/model_zoo/mobilenetv3/scripts/run_train.sh

@@ -31,17 +31,17 @@ run_ascend()
     export PYTHONPATH=${BASEPATH}:$PYTHONPATH
     if [ -d "train" ];
     then
-        rm -rf ./train
+        rm -rf ../train
     fi
-    mkdir ./train
-    cd ./train || exit
-    python ${BASEPATH}/launch.py \
+    mkdir ../train
+    cd ../train || exit
+    python ${BASEPATH}/../src/launch.py \
             --nproc_per_node=$2 \
             --visible_devices=$4 \
             --server_id=$3 \
-            --training_script=${BASEPATH}/train.py \
+            --training_script=${BASEPATH}/../train.py \
             --dataset_path=$5 \
-            --platform=$1 &> train.log &  # dataset train folder
+            --platform=$1 &> ../train.log &  # dataset train folder
 }
 
 run_gpu()
@@ -62,17 +62,17 @@ run_gpu()
     export PYTHONPATH=${BASEPATH}:$PYTHONPATH
     if [ -d "train" ];
     then
-        rm -rf ./train
+        rm -rf ../train
     fi
-    mkdir ./train
-    cd ./train || exit
+    mkdir ../train
+    cd ../train || exit
 
     export CUDA_VISIBLE_DEVICES="$3"
     mpirun -n $2 --allow-run-as-root \
-    python ${BASEPATH}/train.py \
+    python ${BASEPATH}/../train.py \
         --dataset_path=$4 \
         --platform=$1 \
-        &> train.log &  # dataset train folder
+        &> ../train.log &  # dataset train folder
 }
 
 if [ $# -gt 5 ] || [ $# -lt 4 ]

mindspore/model_zoo/mobilenetv3/src/dataset.py

@@ -44,7 +44,12 @@ def create_dataset(dataset_path, do_train, config, platform, repeat_num=1, batch
             ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True,
                                          num_shards=rank_size, shard_id=rank_id)
     elif platform == "GPU":
-        ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)
+        if do_train:
+            from mindspore.communication.management import get_rank, get_group_size
+            ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True,
+                                         num_shards=get_group_size(), shard_id=get_rank())
+        else:
+            ds = de.ImageFolderDatasetV2(dataset_path, num_parallel_workers=8, shuffle=True)
     else:
         raise ValueError("Unsupport platform.")
 

mindspore/model_zoo/mobilenetv3/train.py

@@ -33,7 +33,7 @@ from mindspore.train.callback import ModelCheckpoint, CheckpointConfig, Callback
 from mindspore.train.loss_scale_manager import FixedLossScaleManager
 from mindspore.train.serialization import load_checkpoint, load_param_into_net
 import mindspore.dataset.engine as de
-from mindspore.communication.management import init
+from mindspore.communication.management import init, get_group_size
 from src.dataset import create_dataset
 from src.lr_generator import get_lr
 from src.config import config_gpu, config_ascend
@@ -157,6 +157,11 @@ if __name__ == '__main__':
         # train on gpu
         print("train args: ", args_opt, "\ncfg: ", config_gpu)
 
+        init('nccl')
+        context.set_auto_parallel_context(parallel_mode="data_parallel",
+                                          mirror_mean=True,
+                                          device_num=get_group_size())
+
         # define net
         net = mobilenet_v3_large(num_classes=config_gpu.num_classes)
         # define loss