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

changing how tune reporting works so the scheduler doesnt break if itcant find cv loss. also fixed bug in config structure that was rendering impossible policies

Browse Source
This commit is contained in:
Arec
2021-07-28 09:55:11 -07:00
parent 5a7090a21c
commit 09f77e0587
2 changed files with 32 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
src/bc/bc.py
View File

@@ -20,8 +20,11 @@ def bc_config(ray_config):
'hidden_dim': ray_config['deepsets_phi_hidden_dim']
},
'latent_dim': ray_config['deepsets_latent_dim'],
'rho': {'hidden_n': 0, 'hidden_dim': 10},
'output_dim': 0
'rho': {
'hidden_n': ray_config['deepsets_rho_hidden_n'],
'hidden_dim': ray_config['deepsets_rho_hidden_dim']
},
'output_dim': ray_config['deepsets_output_dim']
},
'path_encoder': {'input_dim': 40, 'hidden_n': 0, 'hidden_dim': 0, 'output_dim': 0},
'head': {
@@ -36,14 +39,13 @@ def bc_config(ray_config):
'lr':ray_config['lr'],
'weight_decay':ray_config['weight_decay']
},
'train_epochs': 100,
'train_epochs': 50,
'train_batch_size': ray_config['train_batch_size'],
'loss': ray_config['loss'],
}
return config
class BehaviorCloningPolicy():
"""
Class for (continuous) behavior cloning policy
@@ -160,6 +162,8 @@ def generate_transforms(dataset):
def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
using_ray = kwargs.get('ray', False)
if using_ray:
print('using ray')
# hyperparams
loss_type = config['loss']
@@ -169,9 +173,9 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
learning_rate = config['optim']['lr']
weight_decay = config['optim']['weight_decay']
cv_every = 5
cv_every = 1
print_epoch_every = 1000
print_cv_every = 1000
print_cv_every = 5
checkpoint_every = 100
cv_batch_size = 256 # doesn't matter
@@ -207,6 +211,10 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
for i in tqdm(range(train_epochs)):
# save model checkpoints
if i % checkpoint_every == 0:
policy.save_model(filestr + '_epoch%04i'%(i) )
# train
epoch_loss = 0
for (batch_idx, batch) in enumerate(training_loader):
@@ -222,12 +230,6 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
epoch_loss += loss.item() / len(train_dataset)
# Write epoch loss
if using_ray:
tune.report(training_loss=epoch_loss, training_iteration=i)
else:
writer.add_scalar('training loss',epoch_loss, i)
# if i % print_epoch_every == 0:
# print('Epoch: {}, Training Loss: {}'.format(i, epoch_loss))
@@ -241,16 +243,20 @@ def train(config, policy, train_dataset, cv_dataset, filestr, **kwargs):
loss = cv_loss_fn(pred_action, batch['action'])
cv_loss += loss.item() / len(cv_dataset)
if using_ray:
tune.report(cv_loss=cv_loss, cv_epoch=i)
# Write epoch loss
if using_ray:
if i % cv_every == 0:
tune.report(training_loss=epoch_loss, cv_loss=cv_loss, training_iteration=i+1)
else:
tune.report(training_loss=epoch_loss, training_iteration=i+1)
else:
writer.add_scalar('training loss',epoch_loss, i)
if i % cv_every == 0:
writer.add_scalar('cv loss', cv_loss, i)
# if i % print_cv_every == 0:
# print('Epoch: {}, CV Loss: {}'.format(i, cv_loss))
# save model checkpoints
if i % checkpoint_every == 0:
policy.save_model(filestr + '_epoch%04i'%(i) )
policy.save_model(filestr)