def get_hparam_config(args): config = { 'unit_type': CategoricalParam(choices=['gru', 'lstm']), 'num_layers': DiscreteParam(min=1, max=10), "d_model": DiscreteParam(min=32, max=1024), "stack_width": DiscreteParam(min=10, max=128), "stack_depth": DiscreteParam(min=10, max=64), "dropout": RealParam(0.0, max=0.3), "batch_size": CategoricalParam(choices=[32, 64, 128]), # optimizer params "optimizer": CategoricalParam(choices=[ "sgd", "adam", "adadelta", "adagrad", "adamax", "rmsprop" ]), "optimizer__global__weight_decay": LogRealParam(), "optimizer__global__lr": LogRealParam(), } return config
def hparams_config(): return {'batch': CategoricalParam(choices=[32, 64, 128]), 'd_model': DiscreteParam(min=32, max=256), 'rnn_num_layers': DiscreteParam(min=1, max=3), 'dropout': RealParam(min=0., max=0.8), 'is_bidirectional': CategoricalParam(choices=[True, False]), 'unit_type': CategoricalParam(choices=['gru', 'lstm']), 'optimizer': CategoricalParam(choices=['sgd', 'adam', 'adadelta', 'adagrad', 'adamax', 'rmsprop']), 'optimizer__global__weight_decay': LogRealParam(), 'optimizer__global__lr': LogRealParam()}
def get_hparam_config(args): return {'seed': ConstantParam(args.seed), 'd_model': ConstantParam(1500), 'dropout': RealParam(), 'monte_carlo_N': ConstantParam(5), 'use_monte_carlo_sim': ConstantParam(True), 'no_mc_fill_val': ConstantParam(0.0), 'gamma': ConstantParam(0.97), 'episodes_to_train': ConstantParam(6), 'gae_lambda': RealParam(0.9, max=0.999), 'ppo_eps': ConstantParam(0.2), 'ppo_batch': ConstantParam(1), 'ppo_epochs': ConstantParam(3), 'entropy_beta': ConstantParam(0.05), 'bias_mode': ConstantParam(args.bias_mode), 'use_true_reward': ConstantParam(args.use_true_reward), 'baseline_reward': ConstantParam(args.baseline_reward), 'reward_params': DictParam({'num_layers': ConstantParam(2), 'd_model': DiscreteParam(min=128, max=500), 'unit_type': ConstantParam('gru'), 'demo_batch_size': ConstantParam(128), 'irl_alg_num_iter': DiscreteParam(2, max=10), 'use_attention': ConstantParam(False), 'bidirectional': ConstantParam(True), 'dropout': RealParam(), 'use_validity_flag': ConstantParam(not args.no_smiles_validity_flag), 'optimizer': CategoricalParam( choices=['sgd', 'adam', 'adadelta', 'adagrad', 'adamax', 'rmsprop']), 'optimizer__global__weight_decay': LogRealParam(), 'optimizer__global__lr': LogRealParam()}), 'agent_params': DictParam({'unit_type': ConstantParam('gru'), 'num_layers': ConstantParam(2), 'stack_width': ConstantParam(1500), 'stack_depth': ConstantParam(200), 'optimizer': ConstantParam('adadelta'), 'optimizer__global__weight_decay': ConstantParam(0.005), 'optimizer__global__lr': ConstantParam(0.001)}), 'critic_params': DictParam({'num_layers': ConstantParam(2), 'd_model': ConstantParam(128), 'unit_type': ConstantParam('gru'), 'optimizer': ConstantParam('adam'), 'optimizer__global__weight_decay': ConstantParam(0.0005), 'optimizer__global__lr': ConstantParam(0.001)}), 'expert_model_dir': ConstantParam('./model_dir/expert_xgb_reg') }
def get_hparam_config(args): config = { "attn_heads": CategoricalParam([1, 2, 4, 8]), "attn_layers": DiscreteParam(min=1, max=4), "lin_dims": DiscreteParam(min=64, max=2048, size=DiscreteParam(min=1, max=3)), "d_model": CategoricalParam(choices=[128, 256, 512, 1024]), "d_hidden": DiscreteParam(min=10, max=64), "stack_width": DiscreteParam(min=10, max=64), "stack_depth": DiscreteParam(min=10, max=64), "d_ff": DiscreteParam(min=128, max=2048), "d_ss": DiscreteParam(min=128, max=2024), "dropout": RealParam(0.0, max=0.5), "batch_size": CategoricalParam(choices=[32, 64, 128]), # optimizer params "optimizer": CategoricalParam(choices=[ "sgd", "adam", "adadelta", "adagrad", "adamax", "rmsprop" ]), "optimizer__global__weight_decay": LogRealParam(), "optimizer__global__lr": LogRealParam(), } return config
def get_hparam_config(args): return { 'd_model': ConstantParam(1500), 'dropout': RealParam(min=0.), 'monte_carlo_N': ConstantParam(5), 'use_monte_carlo_sim': ConstantParam(True), 'no_mc_fill_val': ConstantParam(0.0), 'gamma': ConstantParam(0.97), 'episodes_to_train': DiscreteParam(min=5, max=20), 'reinforce_max_norm': ConstantParam(None), 'lr_decay_gamma': RealParam(), 'lr_decay_step_size': DiscreteParam(min=100, max=1000), 'xent_lambda': ConstantParam(0.0), 'use_true_reward': ConstantParam(args.use_true_reward), 'bias_mode': ConstantParam(args.bias_mode), 'reward_params': DictParam({ 'num_layers': ConstantParam(2), 'd_model': ConstantParam(256), 'unit_type': ConstantParam('lstm'), 'demo_batch_size': ConstantParam(32), 'irl_alg_num_iter': ConstantParam(5), 'use_attention': ConstantParam(args.use_attention), 'use_validity_flag': ConstantParam(not args.no_smiles_validity_flag), 'bidirectional': ConstantParam(True), 'optimizer': ConstantParam('adadelta'), 'optimizer__global__weight_decay': ConstantParam(0.0000), 'optimizer__global__lr': ConstantParam(0.001), }), 'agent_params': DictParam({ 'unit_type': ConstantParam('gru'), 'num_layers': ConstantParam(2), 'stack_width': ConstantParam(1500), 'stack_depth': ConstantParam(200), 'optimizer': ConstantParam('adadelta'), 'optimizer__global__weight_decay': LogRealParam(), 'optimizer__global__lr': LogRealParam() }), 'expert_model_dir': ConstantParam('./model_dir/expert_xgb_reg') }
def get_hparam_config(args): return { 'd_model': ConstantParam(1500), 'dropout': RealParam(min=0.), 'monte_carlo_N': ConstantParam(5), 'use_monte_carlo_sim': ConstantParam(True), 'no_mc_fill_val': ConstantParam(0.0), 'gamma': ConstantParam(0.97), 'episodes_to_train': DiscreteParam(min=5, max=20), 'reinforce_max_norm': ConstantParam(None), 'lr_decay_gamma': RealParam(), 'lr_decay_step_size': DiscreteParam(min=100, max=1000), 'xent_lambda': ConstantParam(0.0), 'use_true_reward': ConstantParam(args.use_true_reward), 'reward_params': DictParam({ 'num_layers': DiscreteParam(min=1, max=4), 'd_model': DiscreteParam(min=128, max=1024), 'unit_type': ConstantParam('lstm'), 'demo_batch_size': CategoricalParam([64, 128, 256]), 'irl_alg_num_iter': DiscreteParam(2, max=10), 'use_attention': ConstantParam(False), 'bidirectional': ConstantParam(True), 'dropout': RealParam(), 'use_validity_flag': ConstantParam(not args.no_smiles_validity_flag), 'optimizer': CategoricalParam(choices=[ 'sgd', 'adam', 'adadelta', 'adagrad', 'adamax', 'rmsprop' ]), 'optimizer__global__weight_decay': LogRealParam(), 'optimizer__global__lr': LogRealParam() }), 'agent_params': DictParam({ 'unit_type': ConstantParam('gru'), 'num_layers': ConstantParam(2), 'stack_width': ConstantParam(1500), 'stack_depth': ConstantParam(200), 'optimizer': ConstantParam('adadelta'), 'optimizer__global__weight_decay': LogRealParam(), 'optimizer__global__lr': LogRealParam() }), 'expert_model_params': DictParam({ 'model_dir': ConstantParam('./model_dir/expert_rnn_bin'), 'd_model': ConstantParam(128), 'rnn_num_layers': ConstantParam(2), 'dropout': ConstantParam(0.8), 'is_bidirectional': ConstantParam(True), 'unit_type': ConstantParam('lstm') }) }