Update format (#160)

* Add ref

* Change `name` keyword to `typename`

* Remove `layer` key from model configuration

* Move model config generaiton to agent

README.md

@@ -19,11 +19,11 @@ from luchador.episode_runner import EpisodeRunner
 
 def main(env, agent, episodes, steps):
     # Create environment
-	Environment = get_env(env['name'])
-	env = Environment(**env['args'])
+    Environment = get_env(env['typename'])
+    env = Environment(**env['args'])
 
     # Create agent
-    Agent = get_agent(agent['name'])
+    Agent = get_agent(agent['typename'])
     agent = Agent(**agent['args'])
     agent.init(env)
 
@@ -109,7 +109,7 @@ $ python
 To use this new agent, we need to create configuration file. As this agent does not take any constructor argument, configuration file is as simple as follow.
 
 ```yaml
-name: MyRandomAgent
+typename: MyRandomAgent
 args: {}
 ```
 
@@ -163,23 +163,20 @@ Network architecture can be described using a set of layer configurations, and w
 model_type: Sequential
 layer_configs:
   - scope: layer1
-    layer:
-      name: Conv2D
-      args:
-        n_filters: 32
-        filter_width: 8
-        filter_height: 8
-        strides: 4
-        padding: valid
+    typename: Conv2D
+    args:
+      n_filters: 32
+      filter_width: 8
+      filter_height: 8
+      strides: 4
+      padding: valid
   - scope: layer2
-    layer:
-      name: ReLU
-      args: {}
+    typename: ReLU
+    args: {}
   - scope: layer3
-    layer:
-      name: Dense
-      args:
-        n_nodes: 3
+    typename: Dense
+    args:
+      n_nodes: 3
 ```
 
 You can feed this configuration to `luchador.nn.util.make_model` then the function will return the coresponding network architecture.
@@ -190,23 +187,20 @@ But having static parameters is sometimes inconvenient. For example, although th
 model_type: Sequential
 layer_configs:
   - scope: layer1
-    layer:
-      name: Conv2D
-      args:
-        n_filters: 32
-        filter_width: 8
-        filter_height: 8
-        strides: 4
-        padding: valid
+    typename: Conv2D
+    args:
+      n_filters: 32
+      filter_width: 8
+      filter_height: 8
+      strides: 4
+      padding: valid
   - scope: layer2
-    layer:
-      name: ReLU
-      args: {{}}
+    typename: ReLU
+    args: {{}}
   - scope: layer3
-    layer:
-      name: Dense
-      args:
-        n_nodes: {n_actions}
+    typename: Dense
+    args:
+      n_nodes: {n_actions}
 ```
 
 When you load this file with `luchador.nn.util.make_model('model.yml', n_actions=5)`, 5 is substituted at `{n_actions}`. Notice that `ReLU`'s `args` parameter became `{{}}` from `{}` so that it Python's `format` function will replace it to `{}`.

example/ALEEnvironment_test.yml

@@ -1,4 +1,4 @@
-name: ALEEnvironment
+typename: ALEEnvironment
 args:
   rom: space_invaders
 

example/ALEEnvironment_train.yml

@@ -1,4 +1,4 @@
-name: ALEEnvironment
+typename: ALEEnvironment
 args:
   rom: space_invaders
 

example/CartPole_agent.yml

@@ -1 +1 @@
-name: CartPoleAgent
+typename: CartPoleAgent

example/CartPole_env.yml

@@ -1,4 +1,4 @@
-name: CartPole
+typename: CartPole
 args:
   angle_limit: 12  # Degree
   distance_limit: 2.4  # meter

example/DQNAgent_test.yml

@@ -7,7 +7,7 @@ alias:
   save_prefix: &save_prefix DQN_integration_test
   initial_parameter: &initial_parameter example/space_invaders_vanilla_dqn_99000.h5
 
-name: DQNAgent
+typename: DQNAgent
 args:
   recorder_config:
     memory_size: 100
@@ -27,26 +27,27 @@ args:
       terminal:
         dtype: bool
 
+  model_config:
+    model_file: example/vanilla_dqn.yml
+    initial_parameter: *initial_parameter
+    input_channel: *stack
+    input_height: *height
+    input_width: *width
+
   q_network_config:
-    model_config:
-      name: example/vanilla_dqn.yml
-      initial_parameter: *initial_parameter
-      input_channel: *stack
-      input_height: *height
-      input_width: *width
     q_learning_config:
       discount_rate: 0.99
       # reward is clipped between the following min and max
       min_reward: -1.0
       max_reward: 1.0
     cost_config:
-      name: SSE2
+      typename: SSE2
       args:
         # error between predicted Q value and target Q value is clipped by the following min and max
         min_delta: -1.0
         max_delta: 1.0
     optimizer_config:
-      name: NeonRMSProp
+      typename: NeonRMSProp
       args:
         decay: 0.95
         epsilon: 0.000001

example/DQNAgent_train.yml

@@ -7,7 +7,7 @@ alias:
   save_prefix: &save_prefix DQN
   initial_parameter: &initial_parameter null
 
-name: DQNAgent
+typename: DQNAgent
 args:
   recorder_config:
     memory_size: 1000000
@@ -27,26 +27,27 @@ args:
       terminal:
         dtype: bool
 
+  model_config:
+    model_file: vanilla_dqn
+    initial_parameter: *initial_parameter
+    input_channel: *stack
+    input_height: *height
+    input_width: *width
+
   q_network_config:
-    model_config:
-      name: vanilla_dqn
-      initial_parameter: *initial_parameter
-      input_channel: *stack
-      input_height: *height
-      input_width: *width
     q_learning_config:
       discount_rate: 0.99
       # reward is clipped between the following min and max
       min_reward: -1.0
       max_reward: 1.0
     cost_config:
-      name: SSE2
+      typename: SSE2
       args:
         # error between predicted Q value and target Q value is clipped by the following min and max
         min_delta: -1.0
         max_delta: 1.0
     optimizer_config:
-      name: NeonRMSProp
+      typename: NeonRMSProp
       args:
         decay: 0.95
         epsilon: 0.000001

example/FlappyBirdEnv.yml

@@ -1,4 +1,4 @@
-name: FlappyBird
+typename: FlappyBird
 args:
   random_seed: null
 

example/MyRandomAgent.yml

@@ -1 +1 @@
-name: MyRandomAgent
+typename: MyRandomAgent

example/RemoteEnv.yml

@@ -1,4 +1,4 @@
-name: RemoteEnv
+typename: RemoteEnv
 args:
   host: 0.0.0.0
   port: 12345

example/vanilla_dqn.yml

@@ -5,100 +5,89 @@ input:
   name: state
 layer_configs:
   - scope: layer0/preprocessing
-    layer:
-      name: TrueDiv
-      args:
-        denom: 255
+    typename: TrueDiv
+    args:
+      denom: 255
   - scope: layer1/conv2D
-    layer:
-      name: Conv2D
-      args:
-        n_filters: 32
-        filter_width: 8
-        filter_height: 8
-        strides: 4
-        padding: valid
-        initializers:
-          bias: &initializer1
-            name: Uniform
-            args:
-              # 1 / sqrt(8 * 8 * 32) = 0.022097
-              maxval: 0.022
-              minval: -0.022
-          weight: *initializer1
+    typename: Conv2D
+    args:
+      n_filters: 32
+      filter_width: 8
+      filter_height: 8
+      strides: 4
+      padding: valid
+      initializers:
+        bias: &initializer1
+          typename: Uniform
+          args:
+            # 1 / sqrt(8 * 8 * 32) = 0.022097
+            maxval: 0.022
+            minval: -0.022
+        weight: *initializer1
   - scope: layer1/ReLU
-    layer:
-      name: ReLU
+    typename: ReLU
   - scope: layer2/conv2D
-    layer:
-      name: Conv2D
-      args:
-        n_filters: 64
-        filter_width: 4
-        filter_height: 4
-        strides: 2
-        padding: valid
-        initializers:
-          bias: &initializer2
-            name: Uniform
-            args:
-              # 1 / sqrt(4 * 4 * 64) = 0.03125
-              maxval: 0.031
-              minval: -0.031
-          weight: *initializer2
+    typename: Conv2D
+    args:
+      n_filters: 64
+      filter_width: 4
+      filter_height: 4
+      strides: 2
+      padding: valid
+      initializers:
+        bias: &initializer2
+          typename: Uniform
+          args:
+            # 1 / sqrt(4 * 4 * 64) = 0.03125
+            maxval: 0.031
+            minval: -0.031
+        weight: *initializer2
   - scope: layer2/ReLU
-    layer:
-      name: ReLU
+    typename: ReLU
   - scope: layer3/conv2D
-    layer:
-      name: Conv2D
-      args:
-        filter_width: 3
-        filter_height: 3
-        n_filters: 64
-        strides: 1
-        padding: valid
-        initializers:
-          bias: &initializer3
-            name: Uniform
-            args:
-              # 1 / sqrt(3 * 3 * 64) = 0.04166
-              maxval: 0.042
-              minval: -0.042
-          weight: *initializer3
+    typename: Conv2D
+    args:
+      filter_width: 3
+      filter_height: 3
+      n_filters: 64
+      strides: 1
+      padding: valid
+      initializers:
+        bias: &initializer3
+          typename: Uniform
+          args:
+            # 1 / sqrt(3 * 3 * 64) = 0.04166
+            maxval: 0.042
+            minval: -0.042
+        weight: *initializer3
   - scope: layer3/ReLU
-    layer:
-      name: ReLU
+    typename: ReLU
   - scope: layer4/flatten
-    layer:
-      name: Flatten
+    typename: Flatten
   - scope: layer5/dense
-    layer:
-      name: Dense
-      args:
-        n_nodes: 512
-        initializers:
-          bias: &initializer5
-            name: Uniform
-            args:
-              # 1 / sqrt(3136) = 0.01785
-              # 3136 is expected #inputs to this layer when the input size to layer0 is 84 * 84 * 4
-              maxval: 0.018
-              minval: -0.018
-          weight: *initializer5
+    typename: Dense
+    args:
+      n_nodes: 512
+      initializers:
+        bias: &initializer5
+          typename: Uniform
+          args:
+            # 1 / sqrt(3136) = 0.01785
+            # 3136 is expected #inputs to this layer when the input size to layer0 is 84 * 84 * 4
+            maxval: 0.018
+            minval: -0.018
+        weight: *initializer5
   - scope: layer5/ReLU
-    layer:
-      name: ReLU
+    typename: ReLU
   - scope: layer6/dense
-    layer:
-      name: Dense
-      args:
-        n_nodes: {n_actions}
-        initializers:
-          bias: &initializer6
-            name: Uniform
-            args:
-              # 1 / sqrt(512) = 0.04419
-              maxval: 0.044
-              minval: -0.044
-          weight: *initializer6
+    typename: Dense
+    args:
+      n_nodes: {n_actions}
+      initializers:
+        bias: &initializer6
+          typename: Uniform
+          args:
+            # 1 / sqrt(512) = 0.04419
+            maxval: 0.044
+            minval: -0.044
+        weight: *initializer6