gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2026-05-01 21:40:56 +02:00
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
rl_coach/presets/CARLA_Dueling_DDQN.py
+4 -10
View File
@@ -3,9 +3,8 @@ import math
from rl_coach.agents.ddqn_agent import DDQNAgentParameters
from rl_coach.architectures.tensorflow_components.heads.dueling_q_head import DuelingQHeadParameters
from rl_coach.base_parameters import VisualizationParameters, MiddlewareScheme
from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps, RunPhase
from rl_coach.core_types import TrainingSteps, EnvironmentEpisodes, EnvironmentSteps
from rl_coach.environments.carla_environment import CarlaEnvironmentParameters
from rl_coach.environments.environment import MaxDumpMethod, SelectedPhaseOnlyDumpMethod
from rl_coach.filters.action.box_discretization import BoxDiscretization
from rl_coach.filters.filter import OutputFilter
from rl_coach.graph_managers.basic_rl_graph_manager import BasicRLGraphManager
@@ -26,9 +25,9 @@ schedule_params.heatup_steps = EnvironmentSteps(1000)
#########
agent_params = DDQNAgentParameters()
agent_params.network_wrappers['main'].learning_rate = 0.00025
agent_params.network_wrappers['main'].heads_parameters = [DuelingQHeadParameters()]
agent_params.network_wrappers['main'].heads_parameters = \
[DuelingQHeadParameters(rescale_gradient_from_head_by_factor=1/math.sqrt(2))]
agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Empty
agent_params.network_wrappers['main'].rescale_gradient_from_head_by_factor = [1/math.sqrt(2), 1/math.sqrt(2)]
agent_params.network_wrappers['main'].clip_gradients = 10
agent_params.algorithm.num_consecutive_playing_steps = EnvironmentSteps(4)
agent_params.network_wrappers['main'].input_embedders_parameters['forward_camera'] = \
@@ -40,11 +39,6 @@ agent_params.output_filter.add_action_filter('discretization', BoxDiscretization
# Environment #
###############
env_params = CarlaEnvironmentParameters()
env_params.level = 'town1'
vis_params = VisualizationParameters()
vis_params.video_dump_methods = [SelectedPhaseOnlyDumpMethod(RunPhase.TEST), MaxDumpMethod()]
vis_params.dump_mp4 = False
graph_manager = BasicRLGraphManager(agent_params=agent_params, env_params=env_params,
schedule_params=schedule_params, vis_params=vis_params)
schedule_params=schedule_params, vis_params=VisualizationParameters())