TD3 (#338)

rl_coach/architectures/tensorflow_components/architecture.py

@@ -203,7 +203,6 @@ class TensorFlowArchitecture(Architecture):
             self._create_gradient_accumulators()
 
         # gradients of the outputs w.r.t. the inputs
-        # at the moment, this is only used by ddpg
         self.gradients_wrt_inputs = [{name: tf.gradients(output, input_ph) for name, input_ph in
                                       self.inputs.items()} for output in self.outputs]
         self.gradients_weights_ph = [tf.placeholder('float32', self.outputs[i].shape, 'output_gradient_weights')

rl_coach/architectures/tensorflow_components/heads/__init__.py

@@ -16,6 +16,7 @@ from .sac_head import SACPolicyHead
 from .sac_q_head import SACQHead
 from .classification_head import ClassificationHead
 from .cil_head import RegressionHead
+from .td3_v_head import TD3VHead
 from .ddpg_v_head import DDPGVHead
 
 __all__ = [
@@ -37,5 +38,6 @@ __all__ = [
     'SACQHead',
     'ClassificationHead',
     'RegressionHead',
+    'TD3VHead'
     'DDPGVHead'
 ]

rl_coach/architectures/tensorflow_components/heads/head.py

@@ -13,7 +13,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-from typing import Type
 
 import numpy as np
 import tensorflow as tf
@@ -22,7 +21,7 @@ from rl_coach.architectures.tensorflow_components.layers import Dense, convert_l
 from rl_coach.base_parameters import AgentParameters
 from rl_coach.spaces import SpacesDefinition
 from rl_coach.utils import force_list
-
+from rl_coach.architectures.tensorflow_components.utils import squeeze_tensor
 
 # Used to initialize weights for policy and value output layers
 def normalized_columns_initializer(std=1.0):
@@ -72,8 +71,9 @@ class Head(object):
         :param input_layer: the input to the graph
         :return: the output of the last layer and the target placeholder
         """
+
         with tf.variable_scope(self.get_name(), initializer=tf.contrib.layers.xavier_initializer()):
-            self._build_module(input_layer)
+            self._build_module(squeeze_tensor(input_layer))
 
             self.output = force_list(self.output)
             self.target = force_list(self.target)

rl_coach/architectures/tensorflow_components/heads/td3_v_head.py

@@ -0,0 +1,67 @@
+#
+# Copyright (c) 2019 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import tensorflow as tf
+
+from rl_coach.architectures.tensorflow_components.layers import Dense
+from rl_coach.architectures.tensorflow_components.heads.head import Head, normalized_columns_initializer
+from rl_coach.base_parameters import AgentParameters
+from rl_coach.core_types import VStateValue
+from rl_coach.spaces import SpacesDefinition
+
+
+class TD3VHead(Head):
+    def __init__(self, agent_parameters: AgentParameters, spaces: SpacesDefinition, network_name: str,
+                 head_idx: int = 0, loss_weight: float = 1., is_local: bool = True, activation_function: str='relu',
+                 dense_layer=Dense, initializer='xavier'):
+        super().__init__(agent_parameters, spaces, network_name, head_idx, loss_weight, is_local, activation_function,
+                         dense_layer=dense_layer)
+        self.name = 'td3_v_values_head'
+        self.return_type = VStateValue
+        self.loss_type = []
+        self.initializer = initializer
+        self.loss = []
+        self.output = []
+
+    def _build_module(self, input_layer):
+        # Standard V Network
+        q_outputs = []
+        self.target = tf.placeholder(tf.float32, shape=(None, 1), name="q_networks_min_placeholder")
+
+        for i in range(input_layer.shape[0]): # assuming that the actual size is 2, as there are two critic networks
+            if self.initializer == 'normalized_columns':
+                q_outputs.append(self.dense_layer(1)(input_layer[i], name='q_output_{}'.format(i + 1),
+                                                     kernel_initializer=normalized_columns_initializer(1.0)))
+            elif self.initializer == 'xavier' or self.initializer is None:
+                q_outputs.append(self.dense_layer(1)(input_layer[i], name='q_output_{}'.format(i + 1)))
+
+            self.output.append(q_outputs[i])
+            self.loss.append(tf.reduce_mean((self.target-q_outputs[i])**2))
+
+        self.output.append(tf.reduce_min(q_outputs, axis=0))
+        self.output.append(tf.reduce_mean(self.output[0]))
+        self.loss = sum(self.loss)
+        tf.losses.add_loss(self.loss)
+
+    def __str__(self):
+        result = [
+            "Q1 Action-Value Stream",
+            "\tDense (num outputs = 1)",
+            "Q2 Action-Value Stream",
+            "\tDense (num outputs = 1)",
+            "Min (Q1, Q2)"
+        ]
+        return '\n'.join(result)

rl_coach/architectures/tensorflow_components/middlewares/fc_middleware.py

@@ -28,23 +28,28 @@ class FCMiddleware(Middleware):
     def __init__(self, activation_function=tf.nn.relu,
                  scheme: MiddlewareScheme = MiddlewareScheme.Medium,
                  batchnorm: bool = False, dropout_rate: float = 0.0,
-                 name="middleware_fc_embedder", dense_layer=Dense, is_training=False):
+                 name="middleware_fc_embedder", dense_layer=Dense, is_training=False, num_streams: int = 1):
         super().__init__(activation_function=activation_function, batchnorm=batchnorm,
                          dropout_rate=dropout_rate, scheme=scheme, name=name, dense_layer=dense_layer,
                          is_training=is_training)
         self.return_type = Middleware_FC_Embedding
-        self.layers = []
+
+        assert(isinstance(num_streams, int) and num_streams >= 1)
+        self.num_streams = num_streams
 
     def _build_module(self):
-        self.layers.append(self.input)
+        self.output = []
 
-        for idx, layer_params in enumerate(self.layers_params):
-            self.layers.extend(force_list(
-                layer_params(self.layers[-1], name='{}_{}'.format(layer_params.__class__.__name__, idx),
-                             is_training=self.is_training)
-            ))
+        for stream_idx in range(self.num_streams):
+            layers = [self.input]
 
-        self.output = self.layers[-1]
+            for idx, layer_params in enumerate(self.layers_params):
+                layers.extend(force_list(
+                    layer_params(layers[-1], name='{}_{}'.format(layer_params.__class__.__name__,
+                                                                 idx + stream_idx * len(self.layers_params)),
+                                 is_training=self.is_training)
+                ))
+            self.output.append((layers[-1]))
 
     @property
     def schemes(self):
@@ -72,3 +77,15 @@ class FCMiddleware(Middleware):
                 ]
         }
 
+    def __str__(self):
+        stream = [str(l) for l in self.layers_params]
+        if self.layers_params:
+            if self.num_streams > 1:
+                stream = [''] + ['\t' + l for l in stream]
+                result = stream * self.num_streams
+                result[0::len(stream)] = ['Stream {}'.format(i) for i in range(self.num_streams)]
+            else:
+                result = stream
+            return '\n'.join(result)
+        else:
+            return 'No layers'

rl_coach/architectures/tensorflow_components/utils.py

@@ -38,3 +38,10 @@ def get_activation_function(activation_function_string: str):
         "Activation function must be one of the following {}. instead it was: {}" \
             .format(activation_functions.keys(), activation_function_string)
     return activation_functions[activation_function_string]
+
+
+def squeeze_tensor(tensor):
+    if tensor.shape[0] == 1:
+        return tensor[0]
+    else:
+        return tensor