1
0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 11:10:20 +01:00
Files
coach/rl_coach/architectures/tensorflow_components/heads/q_head.py
2019-07-14 18:43:48 +03:00

70 lines
3.0 KiB
Python

#
# Copyright (c) 2017 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
from rl_coach.base_parameters import AgentParameters
from rl_coach.core_types import QActionStateValue
from rl_coach.spaces import SpacesDefinition, BoxActionSpace, DiscreteActionSpace
class QHead(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, output_bias_initializer=None):
super().__init__(agent_parameters, spaces, network_name, head_idx, loss_weight, is_local, activation_function,
dense_layer=dense_layer)
self.name = 'q_values_head'
if isinstance(self.spaces.action, BoxActionSpace):
self.num_actions = 1
elif isinstance(self.spaces.action, DiscreteActionSpace):
self.num_actions = len(self.spaces.action.actions)
else:
raise ValueError(
'QHead does not support action spaces of type: {class_name}'.format(
class_name=self.spaces.action.__class__.__name__,
)
)
self.return_type = QActionStateValue
if agent_parameters.network_wrappers[self.network_name].replace_mse_with_huber_loss:
self.loss_type = tf.losses.huber_loss
else:
self.loss_type = tf.losses.mean_squared_error
self.output_bias_initializer = output_bias_initializer
def _build_module(self, input_layer):
# Standard Q Network
self.q_values = self.output = self.dense_layer(self.num_actions)\
(input_layer, name='output', bias_initializer=self.output_bias_initializer)
# used in batch-rl to estimate a probablity distribution over actions
self.softmax = self.add_softmax_with_temperature()
def __str__(self):
result = [
"Dense (num outputs = {})".format(self.num_actions)
]
return '\n'.join(result)
def add_softmax_with_temperature(self):
temperature = self.ap.network_wrappers[self.network_name].softmax_temperature
temperature_scaled_outputs = self.q_values / temperature
return tf.nn.softmax(temperature_scaled_outputs, name="softmax")