def build_sgan_models(args):
long_dtype, float_dtype = get_dtypes(args)
generator = TrajectoryGenerator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
encoder_h_dim=args.encoder_h_dim_g,
decoder_h_dim=args.decoder_h_dim_g,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
noise_dim=args.noise_dim,
noise_type=args.noise_type,
noise_mix_type=args.noise_mix_type,
pooling_type=args.pooling_type,
pool_every_timestep=args.pool_every_timestep,
dropout=args.dropout,
bottleneck_dim=args.bottleneck_dim,
neighborhood_size=args.neighborhood_size,
grid_size=args.grid_size,
interaction_activation=args.interaction_activation,
batch_norm=args.batch_norm)
generator.apply(init_weights)
discriminator = TrajectoryDiscriminator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
h_dim=args.encoder_h_dim_d,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
dropout=args.dropout,
batch_norm=args.batch_norm,
d_type=args.d_type,
interaction_activation=args.interaction_activation,
activation=args.d_activation # default: relu
)
discriminator.apply(init_weights)
return generator, discriminator
def main(args):
print(args)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_num
# train_path = get_dset_path(args.dataset_name, 'train')
# val_path = get_dset_path(args.dataset_name, 'val')
train_path= os.path.join(data_dir,args.dataset_name,'train_small') # 10 files:0-9
val_path= os.path.join(data_dir,args.dataset_name,'val_small') # 5 files: 10-14
long_dtype, float_dtype = get_dtypes(args)
logger.info("Initializing train dataset")
train_dset, train_loader = data_loader(args, train_path)
logger.info("Initializing val dataset")
_, val_loader = data_loader(args, val_path)
iterations_per_epoch = len(train_dset) / args.batch_size / args.d_steps
if args.num_epochs:
args.num_iterations = int(iterations_per_epoch * args.num_epochs)
logger.info(
'There are {} iterations per epoch'.format(iterations_per_epoch)
)
generator = TrajectoryGenerator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
encoder_h_dim=args.encoder_h_dim_g,
decoder_h_dim=args.decoder_h_dim_g,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
noise_dim=args.noise_dim,
noise_type=args.noise_type,
noise_mix_type=args.noise_mix_type,
pooling_type=args.pooling_type,
pool_every_timestep=args.pool_every_timestep,
dropout=args.dropout,
bottleneck_dim=args.bottleneck_dim,
neighborhood_size=args.neighborhood_size,
grid_size=args.grid_size,
batch_norm=args.batch_norm)
generator.apply(init_weights)
generator.type(float_dtype).train()
logger.info('Here is the generator:')
logger.info(generator)
discriminator = TrajectoryDiscriminator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
h_dim=args.encoder_h_dim_d,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
dropout=args.dropout,
batch_norm=args.batch_norm,
d_type=args.d_type,
activation=args.d_activation # default: relu
)
discriminator.apply(init_weights)
discriminator.type(float_dtype).train()
logger.info('Here is the discriminator:')
logger.info(discriminator)
g_loss_fn = gan_g_loss
d_loss_fn = gan_d_loss
optimizer_g = optim.Adam(generator.parameters(), lr=args.g_learning_rate)
optimizer_d = optim.Adam(
discriminator.parameters(), lr=args.d_learning_rate
)
# Maybe restore from checkpoint
restore_path = None
if args.checkpoint_start_from is not None:
restore_path = args.checkpoint_start_from
elif args.restore_from_checkpoint == 1:
restore_path = os.path.join(args.output_dir,
'%s_with_model.pt' % args.checkpoint_name)
if restore_path is not None and os.path.isfile(restore_path):
logger.info('Restoring from checkpoint {}'.format(restore_path))
checkpoint = torch.load(restore_path)
generator.load_state_dict(checkpoint['g_state'])
discriminator.load_state_dict(checkpoint['d_state'])
optimizer_g.load_state_dict(checkpoint['g_optim_state'])
optimizer_d.load_state_dict(checkpoint['d_optim_state'])
t = checkpoint['counters']['t']
epoch = checkpoint['counters']['epoch']
checkpoint['restore_ts'].append(t)
else:
# Starting from scratch, so initialize checkpoint data structure
t, epoch = 0, 0
checkpoint = {
'args': args.__dict__,
'G_losses': defaultdict(list),
'D_losses': defaultdict(list),
'losses_ts': [],
'metrics_val': defaultdict(list),
'metrics_train': defaultdict(list),
'sample_ts': [],
'restore_ts': [],
'norm_g': [],
'norm_d': [],
'counters': {
't': None,
'epoch': None,
},
'g_state': None,
'g_optim_state': None,
'd_state': None,
'd_optim_state': None,
'g_best_state': None,
'd_best_state': None,
'best_t': None,
'g_best_nl_state': None,
'd_best_state_nl': None,
'best_t_nl': None,
}
t0 = None
while t < args.num_iterations:
gc.collect()
d_steps_left = args.d_steps
g_steps_left = args.g_steps
epoch += 1
logger.info('Starting epoch {}'.format(epoch))
for batch in train_loader:
if args.timing == 1:
torch.cuda.synchronize()
t1 = time.time()
# Decide whether to use the batch for stepping on discriminator or
# generator; an iteration consists of args.d_steps steps on the
# discriminator followed by args.g_steps steps on the generator.
if d_steps_left > 0:
step_type = 'd'
losses_d = discriminator_step(args, batch, generator,
discriminator, d_loss_fn,
optimizer_d)
checkpoint['norm_d'].append(
get_total_norm(discriminator.parameters()))
d_steps_left -= 1
elif g_steps_left > 0:
step_type = 'g'
losses_g = generator_step(args, batch, generator,
discriminator, g_loss_fn,
optimizer_g)
checkpoint['norm_g'].append(
get_total_norm(generator.parameters())
)
g_steps_left -= 1
if args.timing == 1:
torch.cuda.synchronize()
t2 = time.time()
logger.info('{} step took {}'.format(step_type, t2 - t1))
# Skip the rest if we are not at the end of an iteration
if d_steps_left > 0 or g_steps_left > 0:
continue
if args.timing == 1:
if t0 is not None:
logger.info('Interation {} took {}'.format(
t - 1, time.time() - t0
))
t0 = time.time()
# Maybe save loss
if t % args.print_every == 0:
logger.info('t = {} / {}'.format(t + 1, args.num_iterations))
for k, v in sorted(losses_d.items()):
# logger.info(' [D] {}: {:.3f}'.format(k, v))
checkpoint['D_losses'][k].append(v)
for k, v in sorted(losses_g.items()):
# logger.info(' [G] {}: {:.3f}'.format(k, v))
checkpoint['G_losses'][k].append(v)
checkpoint['losses_ts'].append(t)
## log scalars
for k, v in sorted(losses_d.items()):
writer.add_scalar("loss/{}".format(k), v, t)
for k, v in sorted(losses_g.items()):
writer.add_scalar("loss/{}".format(k), v, t)
# Maybe save a checkpoint
if t > 0 and t % args.checkpoint_every == 0:
checkpoint['counters']['t'] = t
checkpoint['counters']['epoch'] = epoch
checkpoint['sample_ts'].append(t)
# Check stats on the validation set
logger.info('Checking stats on val ...')
metrics_val = check_accuracy(
args, val_loader, generator, discriminator, d_loss_fn
)
logger.info('Checking stats on train ...')
metrics_train = check_accuracy(
args, train_loader, generator, discriminator,
d_loss_fn, limit=True
)
for k, v in sorted(metrics_val.items()):
# logger.info(' [val] {}: {:.3f}'.format(k, v))
checkpoint['metrics_val'][k].append(v)
for k, v in sorted(metrics_train.items()):
# logger.info(' [train] {}: {:.3f}'.format(k, v))
checkpoint['metrics_train'][k].append(v)
## log scalars
for k, v in sorted(metrics_val.items()):
writer.add_scalar("val/{}".format(k), v, t)
for k, v in sorted(metrics_train.items()):
writer.add_scalar("train/{}".format(k), v, t)
min_ade = min(checkpoint['metrics_val']['ade'])
min_ade_nl = min(checkpoint['metrics_val']['ade_nl'])
if metrics_val['ade'] == min_ade:
logger.info('New low for avg_disp_error')
checkpoint['best_t'] = t
checkpoint['g_best_state'] = generator.state_dict()
checkpoint['d_best_state'] = discriminator.state_dict()
if metrics_val['ade_nl'] == min_ade_nl:
logger.info('New low for avg_disp_error_nl')
checkpoint['best_t_nl'] = t
checkpoint['g_best_nl_state'] = generator.state_dict()
checkpoint['d_best_nl_state'] = discriminator.state_dict()
# Save another checkpoint with model weights and
# optimizer state
checkpoint['g_state'] = generator.state_dict()
checkpoint['g_optim_state'] = optimizer_g.state_dict()
checkpoint['d_state'] = discriminator.state_dict()
checkpoint['d_optim_state'] = optimizer_d.state_dict()
# checkpoint_path = os.path.join(
# args.output_dir, '{}_with_model_{:06d}.pt'.format(args.checkpoint_name,t)
# )
checkpoint_path = os.path.join(args.output_dir, '{}_with_mode.pt'.format(args.checkpoint_name))
logger.info('Saving checkpoint to {}'.format(checkpoint_path))
torch.save(checkpoint, checkpoint_path)
logger.info('Done.')
# Save a checkpoint with no model weights by making a shallow
# copy of the checkpoint excluding some items
# checkpoint_path = os.path.join(
# args.output_dir, '{}_no_model_{:06d}.pt' .format(args.checkpoint_name,t))
checkpoint_path = os.path.join(args.output_dir, '{}_no_model.pt' .format(args.checkpoint_name))
logger.info('Saving checkpoint to {}'.format(checkpoint_path))
key_blacklist = [
'g_state', 'd_state', 'g_best_state', 'g_best_nl_state',
'g_optim_state', 'd_optim_state', 'd_best_state',
'd_best_nl_state'
]
small_checkpoint = {}
for k, v in checkpoint.items():
if k not in key_blacklist:
small_checkpoint[k] = v
torch.save(small_checkpoint, checkpoint_path)
logger.info('Done.')
t += 1
d_steps_left = args.d_steps
g_steps_left = args.g_steps
if t >= args.num_iterations:
break
class TLSGAN(MultiHumanRL):
def __init__(self, device, logger):
super().__init__()
self.name = 'SGAN'
self.device = device
self.policy_learning = False
self.logger = logger
def set_generator_parameters(self, config):
self.obs_len = config.getint('frozen', 'obs_len')
self.pred_len = config.getint('frozen', 'pred_len')
self.mlp_dim = config.getint('frozen', 'mlp_dim')
self.num_layers = config.getint('frozen', 'num_layers')
self.pooling_type = config.get('frozen', 'pooling_type')
self.pool_every_timestep = config.getboolean('frozen',
'pool_every_timestep')
self.dropout = config.getfloat('frozen', 'dropout')
self.bottleneck_dim = config.getint('frozen', 'bottleneck_dim')
self.neighborhood_size = config.getfloat('frozen', 'neighborhood_size')
self.grid_size = config.getint('frozen', 'grid_size')
self.batch_norm = config.getboolean('frozen', 'batch_norm')
self.multiagent_training = config.getboolean('frozen',
'multiagent_training')
def configure(self, config):
self.set_common_parameters(config)
self.set_generator_parameters(config)
checkpoint = torch.load(config.get('frozen', 'model_file'),
map_location='cpu')
self.with_pretrained = config.getboolean('frozen', 'with_pretrained')
self.frozen_training = config.getboolean('frozen', 'frozen_training')
self.policy_learning = config.getboolean('frozen', 'policy_learning')
self.generator = TrajectoryGenerator(
obs_len=self.obs_len,
pred_len=self.pred_len,
embedding_dim=checkpoint['hyperparams']['embedding_dim'],
encoder_h_dim=checkpoint['hyperparams']['encoder_h_dim_g'],
decoder_h_dim=checkpoint['hyperparams']['decoder_h_dim_g'],
mlp_dim=self.mlp_dim,
num_layers=self.num_layers,
noise_dim=(0, ),
noise_type='gaussian',
noise_mix_type='ped',
pooling_type=self.pooling_type,
pool_every_timestep=self.pool_every_timestep,
dropout=self.dropout,
bottleneck_dim=self.bottleneck_dim,
neighborhood_size=self.neighborhood_size,
grid_size=self.grid_size,
batch_norm=self.batch_norm)
if self.with_pretrained:
self.generator.load_state_dict(checkpoint['g_best_state'])
else:
self.generator.apply(init_weights)
self.model = ValueNetwork(self.generator, self.obs_len, self.device,
self.policy_learning).to(self.device)
for p in self.model.parameters():
p.requires_grad = True
for p in self.model.generator.parameters():
p.requires_grad = not self.frozen_training
self.logger.info('Policy: {} '.format(self.name))
def predict(self, states):
"""
A base class for all methods that takes pairwise joint state as input to value network.
The input to the value network is always of shape (batch_size, # humans, rotated joint state length)
states is a list of joint states
"""
if self.phase is None or self.device is None:
raise AttributeError('Phase, device attributes have to be set!')
if self.phase == 'train' and self.epsilon is None:
raise AttributeError(
'Epsilon attribute has to be set in training phase')
if self.reach_destination(states[-1]):
return ActionXY(
0, 0) if self.kinematics == 'holonomic' else ActionRot(0, 0)
if self.action_space is None:
self.build_action_space(states[-1].self_state.v_pref)
if self.policy_learning:
trajs, rel_trajs, self_info = stateToTrajTensors([states])
next_position = self.model(
(trajs, rel_trajs, self_info)).data.squeeze(0)
action = ActionXY(next_position[0], next_position[1])
return action
else:
occupancy_maps = None
probability = np.random.random()
if self.phase == 'train' and probability < self.epsilon:
max_action = self.action_space[np.random.choice(
len(self.action_space))]
else:
self.action_values = list()
max_value = float('-inf')
max_action = None
for action in self.action_space:
next_self_state = self.propagate(
states[-1].self_state, action
) # 'px', 'py', 'vx', 'vy', 'radius', 'gx', 'gy', 'v_pref', 'theta'
if self.query_env:
next_human_states, reward, done, info = self.env.onestep_lookahead(
action)
else:
next_human_states = [
self.propagate(
human_state,
ActionXY(human_state.vx, human_state.vy))
for human_state in states.human_states
] # 'px1', 'py1', 'vx1', 'vy1', 'radius1'
reward = self.compute_reward(next_self_state,
next_human_states)
# VALUE UPDATE
trajs, rel_trajs, self_info = stateToTrajTensors([
states +
[JointState(next_self_state, next_human_states)]
])
next_state_value = self.model(
(trajs, rel_trajs, self_info)).data.item()
value = reward + pow(
self.gamma, self.time_step *
states[-1].self_state.v_pref) * next_state_value
self.action_values.append(value)
if value > max_value:
max_value = value
max_action = action
if max_action is None:
raise ValueError('Value network is not well trained. ')
if self.phase == 'train':
self.last_state = states
return max_action
def main(args):
if args.mode == 'training':
args.checkpoint_every = 100
args.teacher_name = "default"
args.restore_from_checkpoint = 0
#args.l2_loss_weight = 0.0
args.rollout_steps = 1
args.rollout_rate = 1
args.rollout_method = 'sgd'
#print("HHHH"+str(args.l2_loss_weight))
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_num
train_path = get_dset_path(args.dataset_name, 'train')
val_path = get_dset_path(args.dataset_name, 'val')
long_dtype, float_dtype = get_dtypes(args)
logger.info("Initializing train dataset")
train_dset, train_loader = data_loader(args, train_path)
logger.info("Initializing val dataset")
_, val_loader = data_loader(args, val_path)
iterations_per_epoch = len(train_dset) / args.batch_size / args.d_steps
if args.num_epochs:
args.num_iterations = int(iterations_per_epoch * args.num_epochs)
logger.info(
'There are {} iterations per epoch'.format(iterations_per_epoch))
global TrajectoryGenerator, TrajectoryDiscriminator
if args.GAN_type == 'rnn':
print("Default Social GAN")
from sgan.models import TrajectoryGenerator, TrajectoryDiscriminator
elif args.GAN_type == 'simple_rnn':
print("Default Social GAN")
from sgan.rnn_models import TrajectoryGenerator, TrajectoryDiscriminator
else:
print("Feedforward GAN")
if (args.Encoder_type == 'MLP' and args.Decoder_type == 'MLP'):
from sgan.cgs_integrated_model.cgs_ffd_models_E_MLP_D_MLP import TrajectoryGenerator, TrajectoryDiscriminator
if (args.Encoder_type == 'MLP' and args.Decoder_type == 'LSTM'):
from sgan.cgs_integrated_model.cgs_ffd_models_E_MLP_D_LSTM import TrajectoryGenerator, TrajectoryDiscriminator
if (args.Encoder_type == 'LSTM' and args.Decoder_type == 'MLP'):
from sgan.cgs_integrated_model.cgs_ffd_models_E_LSTM_D_MLP import TrajectoryGenerator, TrajectoryDiscriminator
if (args.Encoder_type == 'LSTM' and args.Decoder_type == 'LSTM'):
from sgan.cgs_integrated_model.cgs_ffd_models_E_LSTM_D_LSTM import TrajectoryGenerator, TrajectoryDiscriminator
#image_dir = 'images/' + 'curve_5_traj_l2_0.5'
#image_dir = 'images/5trajectory/' + 'havingplots'+ '2-layers-EN-' + args.Encoder_type + '-DE-20-layers-' + args.Decoder_type + '-L2_' + str(args.l2_loss_weight)
image_dir = 'images/' + str(args.dataset_name) + \
'_EN_' + args.Encoder_type + '(' + str(*[args.mlp_encoder_layers if args.Encoder_type == 'MLP' else 1]) + ')' + \
'_DE_' + args.Decoder_type + '(' + str(*[args.mlp_decoder_layers if args.Decoder_type == 'MLP' else 1]) + ')' + \
'_DIS_' + args.GAN_type.upper() + '(' + str(args.mlp_discriminator_layers) + ')' + \
'_L2_Weight' + '(' + str(args.l2_loss_weight) + ')'
print("Image Dir: ", image_dir)
if not os.path.exists(image_dir):
os.makedirs(image_dir)
generator = TrajectoryGenerator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
encoder_h_dim=args.encoder_h_dim_g,
decoder_h_dim=args.decoder_h_dim_g,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
noise_dim=args.noise_dim,
noise_type=args.noise_type,
noise_mix_type=args.noise_mix_type,
pooling_type=args.pooling_type,
pool_every_timestep=args.pool_every_timestep,
dropout=args.dropout,
bottleneck_dim=args.bottleneck_dim,
neighborhood_size=args.neighborhood_size,
grid_size=args.grid_size,
batch_norm=args.batch_norm,
num_mlp_decoder_layers=args.mlp_decoder_layers,
num_mlp_encoder_layers=args.mlp_encoder_layers)
generator.apply(init_weights)
generator.type(float_dtype).train()
logger.info('Here is the generator:')
logger.info(generator)
discriminator = TrajectoryDiscriminator(
obs_len=args.obs_len,
pred_len=args.pred_len,
embedding_dim=args.embedding_dim,
h_dim=args.encoder_h_dim_d,
mlp_dim=args.mlp_dim,
num_layers=args.num_layers,
dropout=args.dropout,
batch_norm=args.batch_norm,
d_type=args.d_type,
mlp_discriminator_layers=args.mlp_discriminator_layers,
num_mlp_encoder_layers=args.mlp_encoder_layers)
discriminator.apply(init_weights)
discriminator.type(float_dtype).train()
logger.info('Here is the discriminator:')
logger.info(discriminator)
# build teacher
print("[!] teacher_name: ", args.teacher_name)
if args.teacher_name == 'default':
teacher = None
elif args.teacher_name == 'gpurollout':
from teacher_gpu_rollout_torch import TeacherGPURollout
teacher = TeacherGPURollout(args)
teacher.set_env(discriminator, generator)
print("GPU Rollout Teacher")
else:
raise NotImplementedError
g_loss_fn = gan_g_loss
d_loss_fn = gan_d_loss
optimizer_g = optim.Adam(generator.parameters(), lr=args.g_learning_rate)
optimizer_d = optim.Adam(discriminator.parameters(),
lr=args.d_learning_rate)
# # Create D optimizer.
# self.d_optim = tf.train.AdamOptimizer(self.disc_LR*config.D_LR, beta1=config.beta1)
# # Compute the gradients for a list of variables.
# self.grads_d_and_vars = self.d_optim.compute_gradients(self.d_loss, var_list=self.d_vars)
# self.grad_default_real = self.d_optim.compute_gradients(self.d_loss_real, var_list=inputs)
# # Ask the optimizer to apply the capped gradients.
# self.update_d = self.d_optim.apply_gradients(self.grads_d_and_vars)
# ## Get Saliency Map - Teacher
# self.saliency_map = tf.gradients(self.d_loss, self.inputs)[0]
# ###### G Optimizer ######
# # Create G optimizer.
# self.g_optim = tf.train.AdamOptimizer(config.learning_rate*config.G_LR, beta1=config.beta1)
# # Compute the gradients for a list of variables.
# ## With respect to Generator Weights - AutoLoss
# self.grad_default = self.g_optim.compute_gradients(self.g_loss, var_list=[self.G, self.g_vars])
# ## With Respect to Images given to D - Teacher
# # self.grad_default = g_optim.compute_gradients(self.g_loss, var_list=)
# if config.teacher_name == 'default':
# self.optimal_grad = self.grad_default[0][0]
# self.optimal_batch = self.G - self.optimal_grad
# else:
# self.optimal_grad, self.optimal_batch = self.teacher.build_teacher(self.G, self.D_, self.grad_default[0][0], self.inputs)
# # Ask the optimizer to apply the manipulated gradients.
# grads_collected = tf.gradients(self.G, self.g_vars, self.optimal_grad)
# grads_and_vars_collected = list(zip(grads_collected, self.g_vars))
# self.g_teach = self.g_optim.apply_gradients(grads_and_vars_collected)
# Maybe restore from checkpoint
restore_path = None
if args.checkpoint_start_from is not None:
restore_path = args.checkpoint_start_from
elif args.restore_from_checkpoint == 1:
restore_path = os.path.join(args.output_dir,
'%s_with_model.pt' % args.checkpoint_name)
if restore_path is not None and os.path.isfile(restore_path):
logger.info('Restoring from checkpoint {}'.format(restore_path))
checkpoint = torch.load(restore_path)
generator.load_state_dict(checkpoint['g_state'])
discriminator.load_state_dict(checkpoint['d_state'])
optimizer_g.load_state_dict(checkpoint['g_optim_state'])
optimizer_d.load_state_dict(checkpoint['d_optim_state'])
t = checkpoint['counters']['t']
epoch = checkpoint['counters']['epoch']
checkpoint['restore_ts'].append(t)
else:
# Starting from scratch, so initialize checkpoint data structure
t, epoch = 0, 0
checkpoint = {
'args': args.__dict__,
'G_losses': defaultdict(list),
'D_losses': defaultdict(list),
'losses_ts': [],
'metrics_val': defaultdict(list),
'metrics_train': defaultdict(list),
'sample_ts': [],
'restore_ts': [],
'norm_g': [],
'norm_d': [],
'counters': {
't': None,
'epoch': None,
},
'g_state': None,
'g_optim_state': None,
'd_state': None,
'd_optim_state': None,
'g_best_state': None,
'd_best_state': None,
'best_t': None,
'g_best_nl_state': None,
'd_best_state_nl': None,
'best_t_nl': None,
}
t0 = None
fig = plt.figure()
ax = fig.add_axes([0.1, 0.1, 0.75, 0.75])
while t < args.num_iterations:
gc.collect()
d_steps_left = args.d_steps
g_steps_left = args.g_steps
epoch += 1
logger.info('Starting epoch {}'.format(epoch))
for batch in train_loader:
if args.timing == 1:
torch.cuda.synchronize()
t1 = time.time()
# Decide whether to use the batch for stepping on discriminator or
# generator; an iteration consists of args.d_steps steps on the
# discriminator followed by args.g_steps steps on the generator.
if d_steps_left > 0:
if args.mode != 'testing':
step_type = 'd'
losses_d = discriminator_step(args, batch, generator,
discriminator, d_loss_fn,
optimizer_d, teacher,
args.mode)
checkpoint['norm_d'].append(
get_total_norm(discriminator.parameters()))
d_steps_left -= 1
elif g_steps_left > 0:
if args.mode != 'testing':
step_type = 'g'
losses_g = generator_step(args, batch, generator,
discriminator, g_loss_fn,
optimizer_g, args.mode)
checkpoint['norm_g'].append(
get_total_norm(generator.parameters()))
g_steps_left -= 1
if args.timing == 1:
torch.cuda.synchronize()
t2 = time.time()
logger.info('{} step took {}'.format(step_type, t2 - t1))
# Skip the rest if we are not at the end of an iteration
if d_steps_left > 0 or g_steps_left > 0:
continue
if args.timing == 1:
if t0 is not None:
logger.info('Interation {} took {}'.format(
t - 1,
time.time() - t0))
t0 = time.time()
# Maybe save loss
if t % args.print_every == 0 and args.mode != 'testing':
logger.info('t = {} / {}'.format(t + 1, args.num_iterations))
for k, v in sorted(losses_d.items()):
logger.info(' [D] {}: {:.3f}'.format(k, v))
checkpoint['D_losses'][k].append(v)
for k, v in sorted(losses_g.items()):
logger.info(' [G] {}: {:.3f}'.format(k, v))
checkpoint['G_losses'][k].append(v)
checkpoint['losses_ts'].append(t)
# # Check stats on the validation set
# logger.info('Checking stats on val ...')
# metrics_val = check_accuracy(
# args, val_loader, generator, discriminator, d_loss_fn
# )
# logger.info('Checking stats on train ...')
# metrics_train = check_accuracy(
# args, train_loader, generator, discriminator,
# d_loss_fn, limit=True
# )
# for k, v in sorted(metrics_val.items()):
# logger.info(' [val] {}: {:.3f}'.format(k, v))
# checkpoint['metrics_val'][k].append(v)
# for k, v in sorted(metrics_train.items()):
# logger.info(' [train] {}: {:.3f}'.format(k, v))
# checkpoint['metrics_train'][k].append(v)
# min_ade = min(checkpoint['metrics_val']['ade'])
# min_ade_nl = min(checkpoint['metrics_val']['ade_nl'])
# if metrics_val['ade'] == min_ade:
# logger.info('New low for avg_disp_error')
# checkpoint['best_t'] = t
# checkpoint['g_best_state'] = generator.state_dict()
# checkpoint['d_best_state'] = discriminator.state_dict()
# if metrics_val['ade_nl'] == min_ade_nl:
# logger.info('New low for avg_disp_error_nl')
# checkpoint['best_t_nl'] = t
# checkpoint['g_best_nl_state'] = generator.state_dict()
# checkpoint['d_best_nl_state'] = discriminator.state_dict()
if t % 50 == 0:
# save = False
# if t == 160:
# save = True
# print(t)
plot_trajectory(fig,
ax,
args,
val_loader,
generator,
teacher,
args.mode,
t,
save=True,
image_dir=image_dir)
# Maybe save a checkpoint
if t > 0 and t % args.checkpoint_every == 0:
print("Iteration: ", t)
checkpoint['counters']['t'] = t
checkpoint['counters']['epoch'] = epoch
checkpoint['sample_ts'].append(t)
# Save another checkpoint with model weights and
# optimizer state
checkpoint['g_state'] = generator.state_dict()
checkpoint['g_optim_state'] = optimizer_g.state_dict()
checkpoint['d_state'] = discriminator.state_dict()
checkpoint['d_optim_state'] = optimizer_d.state_dict()
checkpoint_path = os.path.join(
args.output_dir, '%s_with_model.pt' % args.checkpoint_name)
logger.info('Saving checkpoint to {}'.format(checkpoint_path))
torch.save(checkpoint, checkpoint_path)
logger.info('Done.')
# Save a checkpoint with no model weights by making a shallow
# copy of the checkpoint excluding some items
checkpoint_path = os.path.join(
args.output_dir, '%s_no_model.pt' % args.checkpoint_name)
logger.info('Saving checkpoint to {}'.format(checkpoint_path))
key_blacklist = [
'g_state', 'd_state', 'g_best_state', 'g_best_nl_state',
'g_optim_state', 'd_optim_state', 'd_best_state',
'd_best_nl_state'
]
small_checkpoint = {}
for k, v in checkpoint.items():
if k not in key_blacklist:
small_checkpoint[k] = v
torch.save(small_checkpoint, checkpoint_path)
logger.info('Done.')
t += 1
d_steps_left = args.d_steps
g_steps_left = args.g_steps
if t >= args.num_iterations:
break
