huangfu/InteractionImitation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

getting best shail/hail configs outsave up and running, adding shell for test runs

Browse Source
This commit is contained in:
Arec Jamgochian
2022-02-27 01:22:39 -08:00
parent acb4fdc518
commit 21cbd1c956
5 changed files with 207 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
best_configs/hail_expA.json Normal file
View File

@@ -0,0 +1,33 @@
{
"experiment": "A",
"trainenv": {
"stop_on_collision": false,
"safe_actions_collision_method": null,
"abort_unsafe_collision_method": null
},
"policy": {
"learning_rate": 0.0003,
"learning_rate_decay": 1.0,
"clip_ratio": 0.2,
"iterations_per_epoch": 100,
"hidden_layer_size": 20,
"n_hidden_layers": 2,
"activation": 0,
"option": 0
},
"value": {
"learning_rate": 0.001,
"iterations_per_epoch": 1000
},
"discriminator": {
"learning_rate": 0.001,
"weight_decay": 0.0001,
"iterations_per_epoch": 100,
"n_hidden_layers_element": 3,
"n_hidden_layers_global": 1,
"hidden_layer_size": 10,
"activation": 0
},
"train_epochs": 2,
"seed": 0
}

33
best_configs/hail_expB.json Normal file
View File

@@ -0,0 +1,33 @@
{
"experiment": "B",
"trainenv": {
"stop_on_collision": false,
"safe_actions_collision_method": null,
"abort_unsafe_collision_method": null
},
"policy": {
"learning_rate": 0.0003,
"learning_rate_decay": 1.0,
"clip_ratio": 0.2,
"iterations_per_epoch": 100,
"hidden_layer_size": 20,
"n_hidden_layers": 2,
"activation": 0,
"option": 0
},
"value": {
"learning_rate": 0.001,
"iterations_per_epoch": 1000
},
"discriminator": {
"learning_rate": 0.001,
"weight_decay": 0.0001,
"iterations_per_epoch": 100,
"n_hidden_layers_element": 3,
"n_hidden_layers_global": 1,
"hidden_layer_size": 10,
"activation": 0
},
"train_epochs": 2,
"seed": 0
}

33
best_configs/shail_expA.json Normal file
View File

@@ -0,0 +1,33 @@
{
"experiment": "A",
"trainenv": {
"stop_on_collision": false,
"safe_actions_collision_method": "circle",
"abort_unsafe_collision_method": "circle"
},
"policy": {
"learning_rate": 0.0003,
"learning_rate_decay": 1.0,
"clip_ratio": 0.2,
"iterations_per_epoch": 100,
"hidden_layer_size": 20,
"n_hidden_layers": 2,
"activation": 0,
"option": 0
},
"value": {
"learning_rate": 0.001,
"iterations_per_epoch": 1000
},
"discriminator": {
"learning_rate": 0.001,
"weight_decay": 0.0001,
"iterations_per_epoch": 100,
"n_hidden_layers_element": 3,
"n_hidden_layers_global": 1,
"hidden_layer_size": 10,
"activation": 0
},
"train_epochs": 2,
"seed": 0
}

33
best_configs/shail_expB.json Normal file
View File

@@ -0,0 +1,33 @@
{
"experiment": "B",
"trainenv": {
"stop_on_collision": false,
"safe_actions_collision_method": "circle",
"abort_unsafe_collision_method": "circle"
},
"policy": {
"learning_rate": 0.0003,
"learning_rate_decay": 1.0,
"clip_ratio": 0.2,
"iterations_per_epoch": 100,
"hidden_layer_size": 20,
"n_hidden_layers": 2,
"activation": 0,
"option": 0
},
"value": {
"learning_rate": 0.001,
"iterations_per_epoch": 1000
},
"discriminator": {
"learning_rate": 0.001,
"weight_decay": 0.0001,
"iterations_per_epoch": 100,
"n_hidden_layers_element": 3,
"n_hidden_layers_global": 1,
"hidden_layer_size": 10,
"activation": 0
},
"train_epochs": 2,
"seed": 0
}

114
shail-experiment.py
View File

@@ -142,8 +142,8 @@ def training_function(config):
v_opt=v_opt, v_opt=v_opt,
v_iters=config['value']['iterations_per_epoch'], v_iters=config['value']['iterations_per_epoch'],
epochs=config['train_epochs'], epochs=config['train_epochs'],
rollout_episodes=60, rollout_episodes=6, #60, FIXME
rollout_steps=60, rollout_steps=6, #60, FIXME
gamma=0.99, gamma=0.99,
gae_lambda=0.9, gae_lambda=0.9,
clip_ratio=config['policy']['clip_ratio'], clip_ratio=config['policy']['clip_ratio'],
@@ -162,44 +162,78 @@ if __name__ == '__main__':
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('experiment', choices=['A', 'B']) parser.add_argument('experiment', choices=['A', 'B'])
parser.add_argument('--epochs', type=int, default=200) parser.add_argument('--epochs', type=int, default=200)
parser.add_argument('--test', type=str, help='path to config file to run final training on')
parser.add_argument('--test_seeds', type=int, default=5)
args = parser.parse_args() args = parser.parse_args()
# if no test config specified, train
if args.test is None:
print('Running Tuning for Experiment %s'%(args.experiment))
analysis = tune.run(
training_function,
config={
'experiment': args.experiment,
'trainenv': {
'stop_on_collision': False,
'safe_actions_collision_method': 'circle',
'abort_unsafe_collision_method': 'circle',
},
'policy': {
'learning_rate': 3e-4,
'learning_rate_decay': 1.0,
'clip_ratio': 0.2,
'iterations_per_epoch': 100,
'hidden_layer_size': tune.grid_search([20, 40]),
'n_hidden_layers': tune.grid_search([2, 3]),
'activation':0,
'option': tune.grid_search(list(range(len(option_list))))
},
'value': {
'learning_rate': 1e-3,
'iterations_per_epoch': 1000,
},
'discriminator': {
'learning_rate': 1e-3,
'weight_decay': 1e-4,
'iterations_per_epoch': 100,
'n_hidden_layers_element': tune.grid_search([3,4]),
'n_hidden_layers_global': tune.grid_search([1,2]),
'hidden_layer_size': 10,
'activation': 0,
},
'train_epochs': args.epochs,
'seed': 0,
}
)
best_config = analysis.get_best_config(metric='gen_collision_rate', mode='min')
print('Best config: ', best_config)
print('Running Tuning for Experiment %s'%(args.experiment)) # safe best_config
analysis = tune.run( if not os.path.isdir(os.path.join(DIR, 'best_configs')):
training_function, os.mkdir(os.path.join(DIR, 'best_configs'))
config={
'experiment': args.experiment, # save shail
'trainenv': { with open(os.path.join(DIR, 'best_configs',f'shail_exp{args.experiment}.json'), 'w', encoding='utf-8') as f:
'stop_on_collision': False, json.dump(best_config, f, ensure_ascii=False, indent=4)
'safe_actions_collision_method': 'circle',
'abort_unsafe_collision_method': 'circle', # save hail
}, best_config['trainenv']['safe_actions_collision_method']=None
'policy': { best_config['trainenv']['abort_unsafe_collision_method']=None
'learning_rate': 3e-4, with open(os.path.join(DIR, 'best_configs',f'hail_exp{args.experiment}.json'), 'w', encoding='utf-8') as f:
'learning_rate_decay': 1.0, json.dump(best_config, f, ensure_ascii=False, indent=4)
'clip_ratio': 0.2,
'iterations_per_epoch': 100, # if config file specified, rerun it with appropriate number of seeds
'hidden_layer_size': tune.grid_search([20, 40]), else:
'n_hidden_layers': tune.grid_search([2, 3]), with open(args.test, 'rb') as f:
'activation':0, config = json.load(f)
'option': tune.grid_search(list(range(len(option_list))))
}, print(f'Retraining {args.test} with {args.test_seed} seeds on experiment {config["experiment"]}')
'value': {
'learning_rate': 1e-3, # rerun with appropriate number of seeds
'iterations_per_epoch': 1000, config['seed'] = tune.grid_search(list(range(1,args.test_seeds+1)))
}, analysis = tune.run(training_function, config)
'discriminator': {
'learning_rate': 1e-3, # move final policies to appropriate directory
'weight_decay': 1e-4, # import pdb
'iterations_per_epoch': 100, # pdb.set_trace()
'n_hidden_layers_element': tune.grid_search([3,4]), # a = 0
'n_hidden_layers_global': tune.grid_search([1,2]),
'hidden_layer_size': 10,
'activation': 0,
},
'train_epochs': args.epochs,
'seed': 0,
}
)
print('Best config: ', analysis.get_best_config(metric='gen_collision_rate', mode='min'))