polish config

.github/workflows/platform_test.yml

@@ -21,7 +21,7 @@ jobs:
         with:
           python-version: ${{ (matrix.os == 'macos-latest' && matrix.python-version == '3.7') && '3.7.16' || matrix.python-version }}
       - name: do_platform_test
-        timeout-minutes: 120
+        timeout-minutes: 30
         run: |
           python -m pip install .
           python -m pip install ".[test,k8s]"

.github/workflows/unit_test.yml

@@ -20,7 +20,7 @@ jobs:
         with:
           python-version: ${{ matrix.python-version }}
       - name: do_unittest
-        timeout-minutes: 120
+        timeout-minutes: 40
         run: |
           python -m pip install box2d-py
           python -m pip install .

ding/model/common/encoder.py

@@ -168,7 +168,9 @@ def __init__(
             else:
                 layers = []
                 for i in range(len(hidden_size_list)):
-                    layers.append(ResFCBlock(hidden_size_list[0], activation=self.act, norm_type=norm_type))
+                    layers.append(
+                        ResFCBlock(hidden_size_list[0], activation=self.act, norm_type=norm_type, dropout=dropout)
+                    )
                 self.main = nn.Sequential(*layers)
         else:
             layers = []

ding/model/template/q_learning.py

@@ -54,6 +54,7 @@ def __init__(
             )
         # Conv Encoder
         elif len(obs_shape) == 3:
+            assert dropout is None, "dropout is not supported in ConvEncoder"
             self.encoder = ConvEncoder(obs_shape, encoder_hidden_size_list, activation=activation, norm_type=norm_type)
         else:
             raise RuntimeError(

ding/policy/dqn.py

@@ -100,8 +100,6 @@ class DQNPolicy(Policy):
         model=dict(
             #(list(int)) Sequence of ``hidden_size`` of subsequent conv layers and the final dense layer.
             encoder_hidden_size_list=[128, 128, 64],
-            # (float) Dropout rate for dropout layers. When set to ``None`` means no dropout layers.
-            dropout=None,
         ),
         learn=dict(
             # (int) How many updates(iterations) to train after collector's one collection.

ding/torch_utils/network/res_block.py

@@ -127,6 +127,8 @@ def __init__(
         self.act = activation
         if dropout is not None:
             self.dropout = nn.Dropout(dropout)
+        else:
+            self.dropout = None
         self.fc1 = fc_block(in_channels, in_channels, activation=self.act, norm_type=norm_type)
         self.fc2 = fc_block(in_channels, in_channels, activation=None, norm_type=norm_type)
 
@@ -143,6 +145,6 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
         x = self.fc1(x)
         x = self.fc2(x)
         x = self.act(x + identity)
-        if hasattr(self, 'dropout'):
+        if self.dropout is not None:
             x = self.dropout(x)
         return x