add conv filter to EpisodicDataset in utils.py, save ckpt every 1000 steps
这个提交包含在:
@ -403,7 +403,7 @@ def train_bc(train_dataloader, val_dataloader, config):
|
|||||||
summary_string += f'{k}: {v.item():.3f} '
|
summary_string += f'{k}: {v.item():.3f} '
|
||||||
print(summary_string)
|
print(summary_string)
|
||||||
|
|
||||||
if epoch % 100 == 0:
|
if epoch % 1000 == 0:
|
||||||
ckpt_path = os.path.join(ckpt_dir, f'policy_epoch_{epoch}_seed_{seed}.ckpt')
|
ckpt_path = os.path.join(ckpt_dir, f'policy_epoch_{epoch}_seed_{seed}.ckpt')
|
||||||
torch.save(policy.state_dict(), ckpt_path)
|
torch.save(policy.state_dict(), ckpt_path)
|
||||||
plot_history(train_history, validation_history, epoch, ckpt_dir, seed)
|
plot_history(train_history, validation_history, epoch, ckpt_dir, seed)
|
||||||
|
|||||||
12
utils.py
12
utils.py
@ -28,7 +28,11 @@ class EpisodicDataset(torch.utils.data.Dataset):
|
|||||||
with h5py.File(dataset_path, 'r') as root:
|
with h5py.File(dataset_path, 'r') as root:
|
||||||
is_sim = root.attrs['sim']
|
is_sim = root.attrs['sim']
|
||||||
if '/base_action' in root:
|
if '/base_action' in root:
|
||||||
action = np.concatenate([root['/action'][()], root['/base_action'][()]], axis=-1)
|
base_action = root['/base_action'][()]
|
||||||
|
smoothed_base_action = np.stack([
|
||||||
|
np.convolve(base_action[:, i], np.ones(20)/20, mode='same') for i in range(base_action.shape[1])
|
||||||
|
], axis=-1, dtype=np.float32)
|
||||||
|
action = np.concatenate([root['/action'][()], smoothed_base_action], axis=-1)
|
||||||
else:
|
else:
|
||||||
action = root['/action'][()]
|
action = root['/action'][()]
|
||||||
original_action_shape = action.shape
|
original_action_shape = action.shape
|
||||||
@ -89,7 +93,11 @@ def get_norm_stats(dataset_dir, num_episodes):
|
|||||||
with h5py.File(dataset_path, 'r') as root:
|
with h5py.File(dataset_path, 'r') as root:
|
||||||
qpos = root['/observations/qpos'][()]
|
qpos = root['/observations/qpos'][()]
|
||||||
qvel = root['/observations/qvel'][()]
|
qvel = root['/observations/qvel'][()]
|
||||||
action = np.concatenate([root['/action'][()], root['/base_action'][()]], axis=-1)
|
base_action = root['/base_action'][()]
|
||||||
|
smoothed_base_action = np.stack([
|
||||||
|
np.convolve(base_action[:, i], np.ones(20)/20, mode='same') for i in range(base_action.shape[1])
|
||||||
|
], axis=-1, dtype=np.float32)
|
||||||
|
action = np.concatenate([root['/action'][()], smoothed_base_action], axis=-1)
|
||||||
except:
|
except:
|
||||||
print(f'Error loading {dataset_path}')
|
print(f'Error loading {dataset_path}')
|
||||||
quit()
|
quit()
|
||||||
|
|||||||
在新工单中引用
屏蔽一个用户