| import gin.tf.external_configurables |
| import compiler_opt.rl.constant |
| import compiler_opt.rl.constant_value_network |
| import compiler_opt.rl.gin_external_configurables |
| import compiler_opt.rl.regalloc.config |
| import compiler_opt.rl.regalloc_network |
| import tf_agents.agents.ppo.ppo_agent |
| import tf_agents.networks.actor_distribution_network |
| |
| train_eval.agent_name=%constant.AgentName.PPO |
| train_eval.problem_type='regalloc' |
| train_eval.warmstart_policy_dir='' |
| train_eval.num_policy_iterations=3000 |
| train_eval.num_iterations=200 |
| train_eval.batch_size=256 |
| train_eval.train_sequence_length=16 |
| train_eval.deploy_policy_name='saved_collect_policy' |
| train_eval.moving_average_decay_rate=0.8 |
| train_eval.use_random_network_distillation=False |
| |
| ####################################### |
| # Turn on if using train_locally.py |
| train_eval.use_stale_results=False |
| |
| # Turn on if using train_with_rpc.py |
| # train_eval.additional_compilation_flags=() |
| ####################################### |
| |
| # RandomNetworkDistillation configs, off if train_eval.use_random_network_distillation=False. |
| RandomNetworkDistillation.encoding_network = @regalloc_network.RegAllocRNDEncodingNetwork |
| RandomNetworkDistillation.learning_rate = 1e-4 |
| RandomNetworkDistillation.update_frequency = 2 |
| RandomNetworkDistillation.fc_layer_params = [32, 128] |
| RandomNetworkDistillation.initial_intrinsic_reward_scale = 1.0 |
| RandomNetworkDistillation.half_decay_steps = 10000 |
| |
| regalloc.config.get_observation_processing_layer_creator.quantile_file_dir='compiler_opt/rl/regalloc/vocab' |
| regalloc.config.get_observation_processing_layer_creator.with_sqrt = False |
| regalloc.config.get_observation_processing_layer_creator.with_z_score_normalization = False |
| |
| create_agent.policy_network = @regalloc_network.RegAllocNetwork |
| |
| RegAllocNetwork.preprocessing_combiner=@tf.keras.layers.Concatenate() |
| RegAllocNetwork.fc_layer_params=(80, 40) |
| RegAllocNetwork.dropout_layer_params=None |
| RegAllocNetwork.activation_fn=@tf.keras.activations.relu |
| |
| ConstantValueNetwork.constant_output_val=0 |
| |
| tf.train.AdamOptimizer.learning_rate = 0.0003 |
| tf.train.AdamOptimizer.epsilon = 0.0003125 |
| |
| PPOAgent.optimizer = @tf.train.AdamOptimizer() |
| PPOAgent.importance_ratio_clipping = 0.2 |
| PPOAgent.lambda_value = 0.0 |
| PPOAgent.discount_factor = 0.0 |
| PPOAgent.entropy_regularization = 0.005 |
| PPOAgent.policy_l2_reg = 0.00001 |
| PPOAgent.value_function_l2_reg = 0.0 |
| PPOAgent.shared_vars_l2_reg = 0.0 |
| PPOAgent.value_pred_loss_coef = 0.0 |
| PPOAgent.num_epochs = 1 |
| PPOAgent.use_gae = False |
| PPOAgent.use_td_lambda_return = False |
| PPOAgent.normalize_rewards = False |
| PPOAgent.reward_norm_clipping = 10.0 |
| PPOAgent.normalize_observations = False |
| PPOAgent.log_prob_clipping = 0.0 |
| PPOAgent.kl_cutoff_factor = 2.0 |
| PPOAgent.kl_cutoff_coef = 1000.0 |
| PPOAgent.initial_adaptive_kl_beta = 1.0 |
| PPOAgent.adaptive_kl_target = 0.01 |
| PPOAgent.adaptive_kl_tolerance = 0.3 |
| PPOAgent.gradient_clipping = None |
| PPOAgent.value_clipping = None |
| PPOAgent.check_numerics = False |
| PPOAgent.compute_value_and_advantage_in_train = True |
| PPOAgent.update_normalizers_in_train=True |
| PPOAgent.debug_summaries = True |
| PPOAgent.summarize_grads_and_vars = True |
