gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 03:30:19 +01:00
Code Issues Projects Releases Wiki Activity

network_imporvements branch merge

Browse Source
This commit is contained in:
Shadi Endrawis
2018-10-02 13:41:46 +03:00
parent 72ea933384
commit 51726a5b80
110 changed files with 1639 additions and 1161 deletions
Download Patch File Download Diff File Expand all files Collapse all files

202
rl_coach/base_parameters.py
View File

@@ -22,7 +22,8 @@ from collections import OrderedDict
from enum import Enum
from typing import Dict, List, Union
from rl_coach.core_types import TrainingSteps, EnvironmentSteps, GradientClippingMethod
from rl_coach.core_types import TrainingSteps, EnvironmentSteps, GradientClippingMethod, RunPhase
# from rl_coach.environments.environment import SelectedPhaseOnlyDumpMethod, MaxDumpMethod
from rl_coach.filters.filter import NoInputFilter
@@ -173,7 +174,71 @@ class PresetValidationParameters(Parameters):
class NetworkParameters(Parameters):
def __init__(self):
def __init__(self,
force_cpu = False,
async_training = False,
shared_optimizer = True,
scale_down_gradients_by_number_of_workers_for_sync_training = True,
clip_gradients = None,
gradients_clipping_method = GradientClippingMethod.ClipByGlobalNorm,
l2_regularization = 0,
learning_rate = 0.00025,
learning_rate_decay_rate = 0,
learning_rate_decay_steps = 0,
input_embedders_parameters = {},
embedding_merger_type = EmbeddingMergerType.Concat,
middleware_parameters = None,
heads_parameters = [],
use_separate_networks_per_head = False,
optimizer_type = 'Adam',
optimizer_epsilon = 0.0001,
adam_optimizer_beta1 = 0.9,
adam_optimizer_beta2 = 0.99,
rms_prop_optimizer_decay = 0.9,
batch_size = 32,
replace_mse_with_huber_loss = False,
create_target_network = False,
tensorflow_support = True):
"""
:param force_cpu:
Force the neural networks to run on the CPU even if a GPU is available
:param async_training:
If set to True, asynchronous training will be used, meaning that each workers will progress in its own
speed, while not waiting for the rest of the workers to calculate their gradients.
:param shared_optimizer:
If set to True, a central optimizer which will be shared with all the workers will be used for applying
gradients to the network. Otherwise, each worker will have its own optimizer with its own internal
parameters that will only be affected by the gradients calculated by that worker
:param scale_down_gradients_by_number_of_workers_for_sync_training:
If set to True, in synchronous training, the gradients of each worker will be scaled down by the
number of workers. This essentially means that the gradients applied to the network are the average
of the gradients over all the workers.
:param clip_gradients:
A value that will be used for clipping the gradients of the network. If set to None, no gradient clipping
will be applied. Otherwise, the gradients will be clipped according to the gradients_clipping_method.
:param gradients_clipping_method:
A gradient clipping method, defined by a GradientClippingMethod enum, and that will be used to clip the
gradients of the network. This will only be used if the clip_gradients value is defined as a value other
than None.
:param l2_regularization:
:param learning_rate:
:param learning_rate_decay_rate:
:param learning_rate_decay_steps:
:param input_embedders_parameters:
:param embedding_merger_type:
:param middleware_parameters:
:param heads_parameters:
:param use_separate_networks_per_head:
:param optimizer_type:
:param optimizer_epsilon:
:param adam_optimizer_beta1:
:param adam_optimizer_beta2:
:param rms_prop_optimizer_decay:
:param batch_size:
:param replace_mse_with_huber_loss:
:param create_target_network:
:param tensorflow_support:
"""
super().__init__()
self.framework = Frameworks.tensorflow
self.sess = None
@@ -182,9 +247,6 @@ class NetworkParameters(Parameters):
self.force_cpu = False
# distributed training options
self.num_threads = 1
self.synchronize_over_num_threads = 1
self.distributed = False
self.async_training = False
self.shared_optimizer = True
self.scale_down_gradients_by_number_of_workers_for_sync_training = True
@@ -192,7 +254,6 @@ class NetworkParameters(Parameters):
# regularization
self.clip_gradients = None
self.gradients_clipping_method = GradientClippingMethod.ClipByGlobalNorm
self.kl_divergence_constraint = None
self.l2_regularization = 0
# learning rate
@@ -205,9 +266,6 @@ class NetworkParameters(Parameters):
self.embedding_merger_type = EmbeddingMergerType.Concat
self.middleware_parameters = None
self.heads_parameters = []
self.num_output_head_copies = 1
self.loss_weights = []
self.rescale_gradient_from_head_by_factor = [1]
self.use_separate_networks_per_head = False
self.optimizer_type = 'Adam'
self.optimizer_epsilon = 0.0001
@@ -227,35 +285,113 @@ class NetworkComponentParameters(Parameters):
self.dense_layer = dense_layer
class VisualizationParameters(Parameters):
def __init__(self):
def __init__(self,
print_networks_summary=False,
dump_csv=True,
dump_signals_to_csv_every_x_episodes=5,
dump_gifs=False,
dump_mp4=False,
video_dump_methods=[],
dump_in_episode_signals=False,
dump_parameters_documentation=True,
render=False,
native_rendering=False,
max_fps_for_human_control=10,
tensorboard=False,
add_rendered_image_to_env_response=False):
"""
:param print_networks_summary:
If set to True, a summary of all the networks structure will be printed at the beginning of the experiment
:param dump_csv:
If set to True, the logger will dump logs to a csv file once in every dump_signals_to_csv_every_x_episodes
episodes. The logs can be later used to visualize the training process using Coach Dashboard.
:param dump_signals_to_csv_every_x_episodes:
Defines the number of episodes between writing new data to the csv log files. Lower values can affect
performance, as writing to disk may take time, and it is done synchronously.
:param dump_gifs:
If set to True, GIF videos of the environment will be stored into the experiment directory according to
the filters defined in video_dump_methods.
:param dump_mp4:
If set to True, MP4 videos of the environment will be stored into the experiment directory according to
the filters defined in video_dump_methods.
:param dump_in_episode_signals:
If set to True, csv files will be dumped for each episode for inspecting different metrics within the
episode. This means that for each step in each episode, different metrics such as the reward, the
future return, etc. will be saved. Setting this to True may affect performance severely, and therefore
this should be used only for debugging purposes.
:param dump_parameters_documentation:
If set to True, a json file containing all the agent parameters will be saved in the experiment directory.
This may be very useful for inspecting the values defined for each parameters and making sure that all
the parameters are defined as expected.
:param render:
If set to True, the environment render function will be called for each step, rendering the image of the
environment. This may affect the performance of training, and is highly dependent on the environment.
By default, Coach uses PyGame to render the environment image instead of the environment specific rendered.
To change this, use the native_rendering flag.
:param native_rendering:
If set to True, the environment native renderer will be used for rendering the environment image.
In some cases this can be slower than rendering using PyGame through Coach, but in other cases the
environment opens its native renderer by default, so rendering with PyGame is an unnecessary overhead.
:param max_fps_for_human_control:
The maximum number of frames per second used while playing the environment as a human. This only has
effect while using the --play flag for Coach.
:param tensorboard:
If set to True, TensorBoard summaries will be stored in the experiment directory. This can later be
loaded in TensorBoard in order to visualize the training process.
:param video_dump_methods:
A list of dump methods that will be used as filters for deciding when to save videos.
The filters in the list will be checked one after the other until the first dump method that returns
false for should_dump() in the environment class. This list will only be used if dump_mp4 or dump_gif are
set to True.
:param add_rendered_image_to_env_response:
Some environments have a different observation compared to the one displayed while rendering.
For some cases it can be useful to pass the rendered image to the agent for visualization purposes.
If this flag is set to True, the rendered image will be added to the environment EnvResponse object,
which will be passed to the agent and allow using those images.
"""
super().__init__()
# Visualization parameters
self.print_summary = True
self.dump_csv = True
self.dump_gifs = False
self.dump_mp4 = False
self.dump_signals_to_csv_every_x_episodes = 5
self.dump_in_episode_signals = False
self.dump_parameters_documentation = True
self.render = False
self.native_rendering = False
self.max_fps_for_human_control = 10
self.tensorboard = False
self.video_dump_methods = [] # a list of dump methods which will be checked one after the other until the first
# dump method that returns false for should_dump()
self.add_rendered_image_to_env_response = False
self.print_networks_summary = print_networks_summary
self.dump_csv = dump_csv
self.dump_gifs = dump_gifs
self.dump_mp4 = dump_mp4
self.dump_signals_to_csv_every_x_episodes = dump_signals_to_csv_every_x_episodes
self.dump_in_episode_signals = dump_in_episode_signals
self.dump_parameters_documentation = dump_parameters_documentation
self.render = render
self.native_rendering = native_rendering
self.max_fps_for_human_control = max_fps_for_human_control
self.tensorboard = tensorboard
self.video_dump_methods = video_dump_methods
self.add_rendered_image_to_env_response = add_rendered_image_to_env_response
class AgentParameters(Parameters):
def __init__(self, algorithm: AlgorithmParameters, exploration: 'ExplorationParameters', memory: 'MemoryParameters',
networks: Dict[str, NetworkParameters], visualization: VisualizationParameters=VisualizationParameters()):
"""
:param algorithm: the algorithmic parameters
:param exploration: the exploration policy parameters
:param memory: the memory module parameters
:param networks: the parameters for the networks of the agent
:param visualization: the visualization parameters
:param algorithm:
A class inheriting AlgorithmParameters.
The parameters used for the specific algorithm used by the agent.
These parameters can be later referenced in the agent implementation through self.ap.algorithm.
:param exploration:
Either a class inheriting ExplorationParameters or a dictionary mapping between action
space types and their corresponding ExplorationParameters. If a dictionary was used,
when the agent will be instantiated, the correct exploration policy parameters will be used
according to the real type of the environment action space.
These parameters will be used to instantiate the exporation policy.
:param memory:
A class inheriting MemoryParameters. It defines all the parameters used by the memory module.
:param networks:
A dictionary mapping between network names and their corresponding network parmeters, defined
as a class inheriting NetworkParameters. Each element will be used in order to instantiate
a NetworkWrapper class, and all the network wrappers will be stored in the agent under
self.network_wrappers. self.network_wrappers is a dict mapping between the network name that
was given in the networks dict, and the instantiated network wrapper.
:param visualization:
A class inheriting VisualizationParameters and defining various parameters that can be
used for visualization purposes, such as printing to the screen, rendering, and saving videos.
"""
super().__init__()
self.visualization = visualization
@@ -278,13 +414,14 @@ class AgentParameters(Parameters):
class TaskParameters(Parameters):
def __init__(self, framework_type: str, evaluate_only: bool=False, use_cpu: bool=False, experiment_path=None,
seed=None):
def __init__(self, framework_type: str="tensorflow", evaluate_only: bool=False, use_cpu: bool=False,
experiment_path="./experiments/test/", seed=None, save_checkpoint_secs=None):
"""
:param framework_type: deep learning framework type. currently only tensorflow is supported
:param evaluate_only: the task will be used only for evaluating the model
:param use_cpu: use the cpu for this task
:param experiment_path: the path to the directory which will store all the experiment outputs
:param save_checkpoint_secs: the number of seconds between each checkpoint saving
:param seed: a seed to use for the random numbers generator
"""
self.framework_type = framework_type
@@ -292,6 +429,7 @@ class TaskParameters(Parameters):
self.evaluate_only = evaluate_only
self.use_cpu = use_cpu
self.experiment_path = experiment_path
self.save_checkpoint_secs = save_checkpoint_secs
self.seed = seed