Improve InceptionV4 network performance.

model_zoo/official/cv/faster_rcnn/scripts/run_distribute_train_ascend.sh

@@ -46,6 +46,7 @@ exit 1
 fi
 
 ulimit -u unlimited
+export HCCL_CONNECT_TIMEOUT=600
 export DEVICE_NUM=8
 export RANK_SIZE=8
 export RANK_TABLE_FILE=$PATH1

model_zoo/official/cv/inceptionv4/README.md

@@ -142,12 +142,12 @@ sh scripts/run_standalone_train_ascend.sh DEVICE_ID DATA_DIR
 Training result will be stored in the example path. Checkpoints will be stored at `ckpt_path` by default, and training log  will be redirected to `./log.txt` like followings.
 
 ```python
-epoch: 1 step: 1251, loss is 5.861846
-Epoch time: 701416.649, per step time: 560.685
-epoch: 2 step: 1251, loss is 4.295785
-Epoch time: 472524.154, per step time: 377.717
-epoch: 3 step: 1251, loss is 3.691987
-Epoch time: 472505.767, per step time: 377.702
+epoch: 1 step: 1251, loss is 5.4833196
+Epoch time: 520274.060, per step time: 415.887
+epoch: 2 step: 1251, loss is 4.093194
+Epoch time: 288520.628, per step time: 230.632
+epoch: 3 step: 1251, loss is 3.6242008
+Epoch time: 288507.506, per step time: 230.622
 ```
 
 ## [Eval process](#contents)
@@ -201,7 +201,7 @@ metric: {'Loss': 0.9849, 'Top1-Acc':0.7985, 'Top5-Acc':0.9460}
 | Outputs                    | probability                                                  |
 | Loss                       | 0.98486                                                      |
 | Accuracy (8p)              | ACC1[79.85%] ACC5[94.60%]                                    |
-| Total time (8p)            | 33h                                                          |
+| Total time (8p)            | 20h                                                          |
 | Params (M)                 | 153M                                                         |
 | Checkpoint for Fine tuning | 2135M                                                        |
 | Scripts                    | [inceptionv4 script](https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/inceptionv4) |
@@ -225,11 +225,11 @@ metric: {'Loss': 0.9849, 'Top1-Acc':0.7985, 'Top5-Acc':0.9460}
 
 | **Ascend** | train performance |
 | :--------: | :---------------: |
-|     1p     |     345 img/s     |
+|     1p     |     556 img/s     |
 
 | **Ascend** | train performance |
 | :--------: | :---------------: |
-|     8p     |     2708img/s     |
+|     8p     |     4430 img/s     |
 
 # [Description of Random Situation](#contents)
 

model_zoo/official/cv/inceptionv4/scripts/run_distribute_train_ascend.sh

@@ -14,6 +14,7 @@
 # limitations under the License.
 # ============================================================================
 
+export HCCL_CONNECT_TIMEOUT=600
 export RANK_TABLE_FILE=$1
 DATA_DIR=$2
 export RANK_SIZE=8

model_zoo/official/cv/inceptionv4/src/config.py

@@ -41,7 +41,4 @@ config_ascend = edict({
     'lr_max': 0.4,
     'warmup_epochs': 1,
     'start_epoch': 1,
-
-    'onnx_filename': 'inceptionv4.onnx',
-    'air_filename': 'inceptionv4.air'
 })

model_zoo/official/cv/inceptionv4/src/inceptionv4.py

@@ -15,40 +15,15 @@
 """InceptionV4"""
 import mindspore.nn as nn
 from mindspore.ops import operations as P
-from mindspore.common.initializer import Initializer
-
-
-
-class Avginitializer(Initializer):
-    """
-    Initialize the weight to 1/m*n,  (m, n) is the shape of kernel.
-    """
-
-
-    def _initialize(self, arr):
-        arr[:] = 0
-        for i in range(arr.shape[0]):
-            for j in range(arr.shape[2]):
-                for k in range(arr.shape[3]):
-                    arr[i][i][j][k] = 1/(arr.shape[2]*arr.shape[3])
-
 
 class Avgpool(nn.Cell):
-    """
-    Average pooling for temporal data.
-
-    Using a custom initializer to turn conv2d into avgpool2d. The weights won't be trained.
-
-    """
-    def __init__(self, channel, kernel_size, stride=1, pad_mode='same'):
+    """Avgpool"""
+    def __init__(self, kernel_size, stride=1, pad_mode='same'):
         super(Avgpool, self).__init__()
-        self.init = Avginitializer()
-        self.conv = nn.Conv2d(channel, channel, kernel_size,
-                              stride=stride, pad_mode=pad_mode, weight_init=self.init)
-        self.conv.set_train(False)
+        self.avg_pool = nn.AvgPool2d(kernel_size=kernel_size, stride=stride, pad_mode=pad_mode)
 
     def construct(self, x):
-        x = self.conv(x)
+        x = self.avg_pool(x)
         return x
 
 
@@ -141,7 +116,7 @@ class InceptionA(nn.Cell):
             Conv2d(96, 96, 3, stride=1, pad_mode='pad', padding=1, has_bias=False)])
 
         self.branch_3 = nn.SequentialCell([
-            Avgpool(384, kernel_size=3, stride=1, pad_mode='same'),
+            Avgpool(kernel_size=3, stride=1, pad_mode='same'),
             Conv2d(384, 96, 1, stride=1, padding=0, has_bias=False)])
 
         self.concat = P.Concat(1)
@@ -178,7 +153,7 @@ class InceptionB(nn.Cell):
             Conv2d(224, 256, (1, 7), pad_mode='same', stride=1, has_bias=False)
         ])
         self.branch_3 = nn.SequentialCell([
-            Avgpool(in_channels, kernel_size=3, stride=1, pad_mode='same'),
+            Avgpool(kernel_size=3, stride=1, pad_mode='same'),
             Conv2d(in_channels, 128, 1, stride=1, padding=0, has_bias=False)
         ])
         self.concat = P.Concat(1)
@@ -265,7 +240,7 @@ class InceptionC(nn.Cell):
             512, 256, (3, 1), pad_mode='same', stride=1, has_bias=False)
 
         self.branch_3 = nn.SequentialCell([
-            Avgpool(in_channels, kernel_size=3, stride=1, pad_mode='same'),
+            Avgpool(kernel_size=3, stride=1, pad_mode='same'),
             Conv2d(in_channels, 256, 1, stride=1, padding=0, has_bias=False)
         ])
         self.concat0 = P.Concat(1)

model_zoo/official/cv/unet/scripts/run_distribute_train.sh

@@ -26,6 +26,7 @@ then
     exit 1
 fi
 
+export HCCL_CONNECT_TIMEOUT=600
 export RANK_SIZE=8
 
 for((i=0;i<RANK_SIZE;i++))