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

Move hyperparameters to config object

Browse Source 
ToDo: parameterize network architectures
This commit is contained in:
ebuehrle
2022-02-23 17:50:21 +01:00
parent 406c4ad9ee
commit 7f64ec7bb0
1 changed files with 23 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
sgail-ppo-options-setobs2.py
View File

@@ -49,14 +49,14 @@ def training_function(config):
env_fn = lambda i: envs[i] env_fn = lambda i: envs[i]
policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n) # config net architecture policy = SetMaskedDiscretePolicy(env_fn(0).action_space.n) # config net architecture
pi_opt = torch.optim.Adam(policy.parameters(), lr=3e-5) # config learning rate pi_opt = torch.optim.Adam(policy.parameters(), lr=config['policy']['learning_rate'])
pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=0.98) # config lr decay pi_lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(pi_opt, gamma=config['policy']['learning_rate_decay'])
value = SetValue() # config net architecture value = SetValue() # config net architecture
v_opt = torch.optim.Adam(value.parameters(), lr=1e-3) # config lr v_opt = torch.optim.Adam(value.parameters(), lr=config['value']['learning_rate'])
discriminator = DeepsetDiscriminator() # config net architecture discriminator = DeepsetDiscriminator() # config net architecture
disc_opt = torch.optim.Adam(discriminator.parameters(), lr=1e-3, weight_decay=1e-3) # config lr, weight decay disc_opt = torch.optim.Adam(discriminator.parameters(), lr=config['discriminator']['learning_rate'], weight_decay=config['discriminator']['weight_decay'])
expert_data = torch.load(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'intersimple-expert-data-setobs2.pt')) expert_data = torch.load(os.path.join(os.path.dirname(os.path.abspath(__file__)), 'intersimple-expert-data-setobs2.pt'))
expert_data = Buffer(*expert_data) expert_data = Buffer(*expert_data)
@@ -69,19 +69,19 @@ def training_function(config):
expert_data=expert_data, expert_data=expert_data,
discriminator=discriminator, discriminator=discriminator,
disc_opt=disc_opt, disc_opt=disc_opt,
disc_iters=100, # config disc_iters=config['discriminator']['iterations_per_epoch'],
policy=policy, policy=policy,
value=value, value=value,
v_opt=v_opt, v_opt=v_opt,
v_iters=1000, # config v_iters=config['value']['iterations_per_epoch'],
epochs=200, epochs=200,
rollout_episodes=60, rollout_episodes=60,
rollout_steps=60, rollout_steps=60,
gamma=0.99, gamma=0.99,
gae_lambda=0.9, gae_lambda=0.9,
clip_ratio=0.2, # config clip_ratio=config['policy']['clip_ratio'],
pi_opt=pi_opt, pi_opt=pi_opt,
pi_iters=100, # config pi_iters=config['policy']['iterations_per_epoch'],
logger=SummaryWriter(comment='sgail-ppo-options-setobs2'), logger=SummaryWriter(comment='sgail-ppo-options-setobs2'),
callback=callback, callback=callback,
lr_schedulers=[pi_lr_scheduler], lr_schedulers=[pi_lr_scheduler],
@@ -90,7 +90,21 @@ def training_function(config):
analysis = tune.run( analysis = tune.run(
training_function, training_function,
config={ config={
'dummy': tune.grid_search([0.001, 0.01, 0.1]), 'policy': {
'learning_rate': tune.grid_search([3e-4]),
'learning_rate_decay': tune.grid_search([1.0]),
'clip_ratio': tune.grid_search([0.2]),
'iterations_per_epoch': tune.grid_search([100]),
},
'value': {
'learning_rate': tune.grid_search([1e-3]),
'iterations_per_epoch': tune.grid_search([1000]),
},
'discriminator': {
'learning_rate': tune.grid_search([1e-3]),
'weight_decay': tune.grid_search([1e-4]),
'iterations_per_epoch': tune.grid_search([100]),
}
} }
) )