gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-18 11:40:18 +01:00
Code Issues Projects Releases Wiki Activity

updated gifs in README + fix for multiworker crashes + improved Atari DQN and Dueling DDQN presets

Browse Source
This commit is contained in:
Itai Caspi
2018-08-16 18:23:32 +03:00
parent 8f99409387
commit 1de04d6fee
11 changed files with 19 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
rl_coach/graph_managers/graph_manager.py
View File

@@ -142,6 +142,9 @@ class GraphManager(object):
config = tf.ConfigProto()
config.allow_soft_placement = True # allow placing ops on cpu if they are not fit for gpu
config.gpu_options.allow_growth = True # allow the gpu memory allocated for the worker to grow if needed
config.gpu_options.per_process_gpu_memory_fraction = 0.2
config.intra_op_parallelism_threads = 1
config.inter_op_parallelism_threads = 1
from rl_coach.architectures.tensorflow_components.distributed_tf_utils import create_and_start_parameters_server, \
create_cluster_spec, create_worker_server_and_device
@@ -169,6 +172,8 @@ class GraphManager(object):
config.allow_soft_placement = True # allow placing ops on cpu if they are not fit for gpu
config.gpu_options.allow_growth = True # allow the gpu memory allocated for the worker to grow if needed
# config.gpu_options.per_process_gpu_memory_fraction = 0.2
config.intra_op_parallelism_threads = 1
config.inter_op_parallelism_threads = 1
if isinstance(task_parameters, DistributedTaskParameters):
# the distributed tensorflow setting

3
rl_coach/presets/Atari_DQN.py
View File

@@ -20,7 +20,8 @@ schedule_params.heatup_steps = EnvironmentSteps(50000)
# Agent #
#########
agent_params = DQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
# since we are using Adam instead of RMSProp, we adjust the learning rate as well
agent_params.network_wrappers['main'].learning_rate = 0.0001
###############
# Environment #

4
rl_coach/presets/Atari_Dueling_DDQN.py
View File

@@ -23,7 +23,9 @@ schedule_params.heatup_steps = EnvironmentSteps(50000)
# Agent #
#########
agent_params = DDQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
# since we are using Adam instead of RMSProp, we adjust the learning rate as well
agent_params.network_wrappers['main'].learning_rate = 0.0001
agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Empty
agent_params.network_wrappers['main'].heads_parameters = [DuelingQHeadParameters()]
agent_params.network_wrappers['main'].rescale_gradient_from_head_by_factor = [1/math.sqrt(2)]