def train_dynamics(config, data_path, train_dir):
# access dict values as attributes
config = edict(config)
# set random seed for reproduction
set_seed(config.train.random_seed)
st_epoch = config.train.resume_epoch if config.train.resume_epoch > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
print(config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(config, data_path, phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase], batch_size=config.train.batch_size,
shuffle=True if phase == 'train' else False,
num_workers=config.train.num_workers)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
'''
define model for dynamics prediction
'''
model_dy = DynaNetMLP(config)
print("model_dy #params: %d" % count_trainable_parameters(model_dy))
if config.train.resume_epoch >= 0:
# if resume from a pretrained checkpoint
model_dy_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d.pth' % (
config.train.resume_epoch, config.train.resume_iter))
print("Loading saved ckp from %s" % model_dy_path)
model_dy.load_state_dict(torch.load(model_dy_path))
# criterion
criterionMSE = nn.MSELoss()
# optimizer
params = model_dy.parameters()
optimizer = optim.Adam(params, lr=config.train.lr, betas=(config.train.adam_beta1, 0.999))
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.9, patience=10, verbose=True)
if use_gpu:
model_dy = model_dy.cuda()
best_valid_loss = np.inf
for epoch in range(st_epoch, config.train.n_epoch):
phases = ['train', 'valid']
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in bar(enumerate(loader)):
if use_gpu:
if isinstance(data, list):
data = [d.cuda() for d in data]
else:
data = data.cuda()
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config.train.n_history, config.train.n_rollout
n_samples = n_his + n_roll
if config.env.type in ['PusherSlider']:
states, actions = data
assert states.size(1) == n_samples
B = states.size(0)
loss_mse = 0.
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
for j in range(n_roll):
state_des = states[:, n_his + j]
# action_cur: B x n_his x action_dim
action_cur = actions[:, j : j + n_his] if actions is not None else None
# state_pred: B x state_dim
state_pred = model_dy(state_cur, action_cur)
loss_mse_cur = criterionMSE(state_pred, state_des)
loss_mse += loss_mse_cur / config.train.n_rollout
# update state_cur
state_cur = torch.cat([state_cur[:, 1:], state_pred.unsqueeze(1)], 1)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if i % config.train.log_per_iter == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config.train.n_epoch, i, data_n_batches[phase],
get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (
np.sqrt(loss_mse.item()), meter_loss_rmse.avg)
print(log)
if phase == 'train' and i % config.train.ckp_per_iter == 0:
torch.save(model_dy.state_dict(), '%s/net_dy_epoch_%d_iter_%d.pth' % (train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config.train.n_epoch, meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == 'valid':
scheduler.step(meter_loss_rmse.avg)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
torch.save(model_dy.state_dict(), '%s/net_best_dy.pth' % (train_dir))
def train_dynamics(
config,
train_dir, # str: directory to save output
):
# set random seed for reproduction
set_seed(config['train']['random_seed'])
st_epoch = config['train'][
'resume_epoch'] if config['train']['resume_epoch'] > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
print(config)
# load the data
episodes = load_episodes_from_config(config)
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(
config,
action_function=action_function,
observation_function=observation_function,
episodes=episodes,
phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'])
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
# compute normalization parameters if not starting from pre-trained network . . .
'''
define model for dynamics prediction
'''
model_dy = None
if config['train']['resume_epoch'] >= 0:
# if resume from a pretrained checkpoint
state_dict_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d_state_dict.pth' %
(config['train']['resume_epoch'], config['train']['resume_iter']))
print("Loading saved ckp from %s" % state_dict_path)
# why is this needed if we already do torch.load???
model_dy.load_state_dict(torch.load(state_dict_path))
# don't we also need to load optimizer state from pre-trained???
else:
# not starting from pre-trained create the network and compute the
# normalization parameters
model_dy = DynaNetMLP(config)
# compute normalization params
stats = datasets["train"].compute_dataset_statistics()
obs_mean = stats['observations']['mean']
obs_std = stats['observations']['std']
observations_normalizer = DataNormalizer(obs_mean, obs_std)
action_mean = stats['actions']['mean']
action_std = stats['actions']['std']
actions_normalizer = DataNormalizer(action_mean, action_std)
model_dy.action_normalizer = actions_normalizer
model_dy.state_normalizer = observations_normalizer
print("model_dy #params: %d" % count_trainable_parameters(model_dy))
# criterion
criterionMSE = nn.MSELoss()
# optimizer
params = model_dy.parameters()
optimizer = optim.Adam(params,
lr=config['train']['lr'],
betas=(config['train']['adam_beta1'], 0.999))
scheduler = ReduceLROnPlateau(optimizer,
'min',
factor=0.9,
patience=10,
verbose=True)
if use_gpu:
model_dy = model_dy.cuda()
best_valid_loss = np.inf
global_iteration = 0
epoch_counter_external = 0
try:
for epoch in range(st_epoch, config['train']['n_epoch']):
phases = ['train', 'valid']
epoch_counter_external = epoch
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
global_iteration += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config[
'train']['n_rollout']
n_samples = n_his + n_roll
if config['env']['type'] in ['PusherSlider']:
states = data['observations']
actions = data['actions']
if use_gpu:
states = states.cuda()
actions = actions.cuda()
# states, actions = data
assert states.size(1) == n_samples
# normalize states and actions once for entire rollout
states = model_dy.state_normalizer.normalize(
states)
actions = model_dy.action_normalizer.normalize(
actions)
B = states.size(0)
loss_mse = 0.
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
for j in range(n_roll):
state_des = states[:, n_his + j]
# action_cur: B x n_his x action_dim
action_cur = actions[:, j:j +
n_his] if actions is not None else None
# state_pred: B x state_dim
# state_cur: B x n_his x state_dim
# state_pred: B x state_dim
state_pred = model_dy(state_cur, action_cur)
loss_mse_cur = criterionMSE(
state_pred, state_des)
loss_mse += loss_mse_cur / n_roll
# update state_cur
# state_pred.unsqueeze(1): B x 1 x state_dim
state_cur = torch.cat([
state_cur[:, 1:],
state_pred.unsqueeze(1)
], 1)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if i % config['train']['log_per_iter'] == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config['train']['n_epoch'], i,
data_n_batches[phase], get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (np.sqrt(
loss_mse.item()), meter_loss_rmse.avg)
print(log)
# log data to tensorboard
# only do it once we have reached 500 iterations
if global_iteration > 500:
writer.add_scalar("Params/learning rate",
get_lr(optimizer),
global_iteration)
writer.add_scalar("Loss/train", loss_mse.item(),
global_iteration)
writer.add_scalar("RMSE average loss/train",
meter_loss_rmse.avg,
global_iteration)
if phase == 'train' and i % config['train'][
'ckp_per_iter'] == 0:
save_model(
model_dy, '%s/net_dy_epoch_%d_iter_%d' %
(train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config['train']['n_epoch'],
meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == 'valid':
scheduler.step(meter_loss_rmse.avg)
writer.add_scalar("RMSE average loss/valid",
meter_loss_rmse.avg, global_iteration)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(
model_dy, '%s/net_dy_epoch_%d_keyboard_interrupt' %
(train_dir, epoch_counter_external))
writer.flush() # flush SummaryWriter events to disk
def train_dynamics(config, train_dir, data_dir, model_dy, global_iteration,
writer):
# load the data
multi_episode_dict = DrakeSimEpisodeReader.load_dataset(
data_dir, load_image_data=False)
'''
for episode_name in list(multi_episode_dict.keys()):
print("episode name", episode_name)
episode = multi_episode_dict[episode_name]
obs = episode.get_observation(34)
print(obs)
'''
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(
config,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase=phase)
# print(config['train'])
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'],
drop_last=True)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
'''
define model for dynamics prediction
'''
if model_dy is None:
model_dy = DynaNetMLP(config)
# criterion
MSELoss = nn.MSELoss()
L1Loss = nn.L1Loss()
# optimizer
params = model_dy.parameters()
lr = float(config['train']['lr'])
optimizer = optim.Adam(params,
lr=lr,
betas=(config['train']['adam_beta1'], 0.999))
# setup scheduler
sc = config['train']['lr_scheduler']
scheduler = None
if config['train']['lr_scheduler']['enabled']:
if config['train']['lr_scheduler']['type'] == "ReduceLROnPlateau":
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=sc['factor'],
patience=sc['patience'],
threshold_mode=sc['threshold_mode'],
cooldown=sc['cooldown'],
verbose=True)
elif config['train']['lr_scheduler']['type'] == "StepLR":
step_size = config['train']['lr_scheduler']['step_size']
gamma = config['train']['lr_scheduler']['gamma']
scheduler = StepLR(optimizer, step_size=step_size, gamma=gamma)
else:
raise ValueError("unknown scheduler type: %s" %
(config['train']['lr_scheduler']['type']))
if use_gpu:
print("using gpu")
model_dy = model_dy.cuda()
best_valid_loss = np.inf
counters = {'train': 0, 'valid': 0}
try:
for epoch in range(config['train']['n_epoch']):
phases = ['train', 'valid']
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
step_duration_meter = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
loss_container = dict() # store the losses for this step
step_start_time = time.time()
global_iteration += 1
counters[phase] += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config[
'train']['n_rollout']
n_samples = n_his + n_roll
if DEBUG:
print("global iteration: %d" % global_iteration)
print("n_samples", n_samples)
# [B, n_samples, obs_dim]
observations = data['observations']
# [B, n_samples, action_dim]
actions = data['actions']
B = actions.shape[0]
if use_gpu:
observations = observations.cuda()
actions = actions.cuda()
# states, actions = data
assert actions.shape[1] == n_samples
loss_mse = 0.
# we don't have any visual observations, so states are observations
states = observations
# [B, n_his, state_dim]
state_init = states[:, :n_his]
# We want to rollout n_roll steps
# actions = [B, n_his + n_roll, -1]
# so we want action_seq.shape = [B, n_roll, -1]
action_start_idx = 0
action_end_idx = n_his + n_roll - 1
action_seq = actions[:, action_start_idx:
action_end_idx, :]
if DEBUG:
print("states.shape", states.shape)
print("state_init.shape", state_init.shape)
print("actions.shape", actions.shape)
print("action_seq.shape", action_seq.shape)
# try using models_dy.rollout_model instead of doing this manually
rollout_data = rollout_model(state_init=state_init,
action_seq=action_seq,
dynamics_net=model_dy,
compute_debug_data=False)
# [B, n_roll, state_dim]
state_rollout_pred = rollout_data['state_pred']
# [B, n_roll, state_dim]
state_rollout_gt = states[:, n_his:]
if DEBUG:
print("state_rollout_gt.shape",
state_rollout_gt.shape)
print("state_rollout_pred.shape",
state_rollout_pred.shape)
# the loss function is between
# [B, n_roll, state_dim]
state_pred_err = state_rollout_pred - state_rollout_gt
# everything is in 3D space now so no need to do any scaling
# all the losses would be in meters . . . .
loss_mse = MSELoss(state_rollout_pred,
state_rollout_gt)
loss_l1 = L1Loss(state_rollout_pred, state_rollout_gt)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
# compute losses at final step of the rollout
mse_final_step = MSELoss(state_rollout_pred[:, -1, :],
state_rollout_gt[:, -1, :])
l2_final_step = torch.norm(state_pred_err[:, -1],
dim=-1).mean()
l1_final_step = L1Loss(state_rollout_pred[:, -1, :],
state_rollout_gt[:, -1, :])
loss_container['mse'] = loss_mse
loss_container['l1'] = loss_l1
loss_container['mse_final_step'] = mse_final_step
loss_container['l1_final_step'] = l1_final_step
loss_container['l2_final_step'] = l2_final_step
step_duration_meter.update(time.time() - step_start_time)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if i % config['train']['log_per_iter'] == 0:
log = '%s %d [%d/%d][%d/%d] LR: %.6f' % (
phase, global_iteration, epoch,
config['train']['n_epoch'], i,
data_n_batches[phase], get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (np.sqrt(
loss_mse.item()), meter_loss_rmse.avg)
log += ', step time %.6f' % (step_duration_meter.avg)
step_duration_meter.reset()
print(log)
# log data to tensorboard
# only do it once we have reached 100 iterations
if global_iteration > 100:
writer.add_scalar("Params/learning rate",
get_lr(optimizer),
global_iteration)
writer.add_scalar("Loss_MSE/%s" % (phase),
loss_mse.item(),
global_iteration)
writer.add_scalar("L1/%s" % (phase),
loss_l1.item(), global_iteration)
writer.add_scalar("RMSE average loss/%s" % (phase),
meter_loss_rmse.avg,
global_iteration)
writer.add_scalar("n_taj", len(multi_episode_dict),
global_iteration)
for loss_type, loss_obj in loss_container.items():
plot_name = "Loss/%s/%s" % (loss_type, phase)
writer.add_scalar(plot_name, loss_obj.item(),
global_iteration)
if phase == 'train' and global_iteration % config['train'][
'ckp_per_iter'] == 0:
save_model(
model_dy, '%s/net_dy_iter_%d' %
(train_dir, global_iteration))
log = '%s %d [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, global_iteration, epoch, config['train']['n_epoch'],
meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == "train":
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "StepLR"):
scheduler.step()
if phase == 'valid':
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "ReduceLROnPlateau"):
scheduler.step(meter_loss_rmse.avg)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(
model_dy, '%s/net_dy_iter_%d_keyboard_interrupt' %
(train_dir, global_iteration))
writer.flush() # flush SummaryWriter events to disk
return model_dy, global_iteration