def train():
torch.backends.cudnn.benchmark = True
_, dataloader = create_dataloader(config.IMG_DIR + "/train", config.MESH_DIR + "/train",
batch_size=config.BATCH_SIZE, used_layers=config.USED_LAYERS,
img_size=config.IMAGE_SIZE, map_size=config.MAP_SIZE,
augment=config.AUGMENT, workers=config.NUM_WORKERS,
pin_memory=config.PIN_MEMORY, shuffle=True)
in_channels = num_channels(config.USED_LAYERS)
encoder = Encoder(in_channels=in_channels)
decoder = Decoder(num_classes=config.NUM_CLASSES+1)
encoder.apply(init_weights)
decoder.apply(init_weights)
encoder_solver = torch.optim.Adam(filter(lambda p: p.requires_grad, encoder.parameters()),
lr=config.ENCODER_LEARNING_RATE,
betas=config.BETAS)
decoder_solver = torch.optim.Adam(decoder.parameters(),
lr=config.DECODER_LEARNING_RATE,
betas=config.BETAS)
encoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(encoder_solver,
milestones=config.ENCODER_LR_MILESTONES,
gamma=config.GAMMA)
decoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(decoder_solver,
milestones=config.DECODER_LR_MILESTONES,
gamma=config.GAMMA)
encoder = encoder.to(config.DEVICE)
decoder = decoder.to(config.DEVICE)
loss_fn = LossFunction()
init_epoch = 0
if config.CHECKPOINT_FILE and config.LOAD_MODEL:
init_epoch, encoder, decoder = load_checkpoint(encoder, decoder, config.CHECKPOINT_FILE, config.DEVICE)
output_dir = os.path.join(config.OUT_PATH, re.sub("[^0-9a-zA-Z]+", "-", dt.now().isoformat()))
for epoch_idx in range(init_epoch, config.NUM_EPOCHS):
encoder.train()
decoder.train()
train_one_epoch(encoder, decoder, dataloader, loss_fn, encoder_solver, decoder_solver, epoch_idx)
encoder_lr_scheduler.step()
decoder_lr_scheduler.step()
if config.TEST:
test(encoder, decoder)
if config.SAVE_MODEL:
save_checkpoint(epoch_idx, encoder, decoder, output_dir)
if not config.TEST:
test(encoder, decoder)
if not config.SAVE_MODEL:
save_checkpoint(config.NUM_EPOCHS - 1, encoder, decoder, output_dir)
def train(model_config, train_config):
mode = 'train'
dataset = ShakespeareModern(train_shakespeare_path,
test_shakespeare_path,
train_modern_path,
test_modern_path,
mode=mode)
dataloader = DataLoader(dataset,
batch_size=train_config['batch_size'],
shuffle=False)
vocab = dataset.vocab
max_length = dataset.domain_A_max_len
encoder = Encoder(model_config['embedding_size'],
model_config['hidden_dim'],
dataset.vocab.num_words,
batch_size=train_config['batch_size']).cuda()
# print(dataset.domain_A_max_len)
decoder = Decoder(model_config['embedding_size'],
model_config['hidden_dim'],
dataset.vocab.num_words,
max_length,
batch_size=train_config['batch_size']).cuda()
criterion = nn.NLLLoss().cuda()
encoder_optimizer = torch.optim.SGD(encoder.parameters(),
lr=train_config['base_lr'])
decoder_optimizer = torch.optim.SGD(decoder.parameters(),
lr=train_config['base_lr'])
for epoch in range(train_config['num_epochs']):
for idx, (s, s_addn_feats, m,
m_addn_feats) in tqdm(enumerate(dataloader)):
input_tensor = s.transpose(0, 1).cuda()
target_tensor = m.transpose(0, 1).cuda()
encoder_optimizer.zero_grad()
decoder_optimizer.zero_grad()
input_length = input_tensor.size(0)
target_length = target_tensor.size(0)
encoder_outputs = torch.zeros(max_length, encoder.hidden_size)
loss = 0
print('ip', input_tensor.size())
encoder_output, encoder_hidden = encoder(input_tensor)
# encoder_outputs = encoder_output[0, 0]
decoder_input = torch.empty(
(train_config['batch_size'],
1)).fill_(SOS_token).type(torch.LongTensor).cuda()
print(decoder_input.size())
decoder_hidden = encoder_output[-1]
print('dec hid', decoder_hidden.size(), type(decoder_hidden))
while decoder_input:
decoder.hidden = decoder_hidden
decoder_input, decoder_hidden = decoder(
decoder_input, encoder_output)
loss += criterion(decoder_output, target_tensor[di])
loss.backward()
encoder_optimizer.step()
decoder_optimizer.step()
if idx % 100 == 0:
print(
'\tepoch [{}/{}], iter: {}, s_loss: {:.4f}, m_loss: {:.4f}, preds: s: {}, {}, m: {}, {}'
.format(epoch + 1, train_config['num_epochs'], idx,
s_loss.item(), m_loss.item(), s_output.item(),
round(s_output.item()), m_output.item(),
round(m_output.item())))
print('\tepoch [{}/{}]'.format(epoch + 1, train_config['num_epochs']))
return loss.item() / target_length
def train_net(cfg):
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
train_transforms = utils.data_transforms.Compose([
utils.data_transforms.RandomCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(
cfg.TRAIN.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.ColorJitter(cfg.TRAIN.BRIGHTNESS,
cfg.TRAIN.CONTRAST,
cfg.TRAIN.SATURATION),
utils.data_transforms.RandomNoise(cfg.TRAIN.NOISE_STD),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.RandomFlip(),
utils.data_transforms.RandomPermuteRGB(),
utils.data_transforms.ToTensor(),
])
val_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
# Set up data loader
train_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TRAIN_DATASET](cfg)
val_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TEST_DATASET](cfg)
train_data_loader = torch.utils.data.DataLoader(
dataset=train_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.TRAIN, cfg.CONST.N_VIEWS_RENDERING,
train_transforms),
batch_size=cfg.CONST.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKER,
pin_memory=True,
shuffle=True,
drop_last=True)
val_data_loader = torch.utils.data.DataLoader(
dataset=val_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.VAL, cfg.CONST.N_VIEWS_RENDERING,
val_transforms),
batch_size=1,
num_workers=1,
pin_memory=True,
shuffle=False)
# Set up networks
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
print('[DEBUG] %s Parameters in Encoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(encoder)))
print('[DEBUG] %s Parameters in Decoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(decoder)))
print('[DEBUG] %s Parameters in Refiner: %d.' %
(dt.now(), utils.network_utils.count_parameters(refiner)))
print('[DEBUG] %s Parameters in Merger: %d.' %
(dt.now(), utils.network_utils.count_parameters(merger)))
# Initialize weights of networks
encoder.apply(utils.network_utils.init_weights)
decoder.apply(utils.network_utils.init_weights)
refiner.apply(utils.network_utils.init_weights)
merger.apply(utils.network_utils.init_weights)
# Set up solver
if cfg.TRAIN.POLICY == 'adam':
encoder_solver = torch.optim.Adam(filter(lambda p: p.requires_grad,
encoder.parameters()),
lr=cfg.TRAIN.ENCODER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
decoder_solver = torch.optim.Adam(decoder.parameters(),
lr=cfg.TRAIN.DECODER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
refiner_solver = torch.optim.Adam(refiner.parameters(),
lr=cfg.TRAIN.REFINER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
merger_solver = torch.optim.Adam(merger.parameters(),
lr=cfg.TRAIN.MERGER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
elif cfg.TRAIN.POLICY == 'sgd':
encoder_solver = torch.optim.SGD(filter(lambda p: p.requires_grad,
encoder.parameters()),
lr=cfg.TRAIN.ENCODER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
decoder_solver = torch.optim.SGD(decoder.parameters(),
lr=cfg.TRAIN.DECODER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
refiner_solver = torch.optim.SGD(refiner.parameters(),
lr=cfg.TRAIN.REFINER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
merger_solver = torch.optim.SGD(merger.parameters(),
lr=cfg.TRAIN.MERGER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
else:
raise Exception('[FATAL] %s Unknown optimizer %s.' %
(dt.now(), cfg.TRAIN.POLICY))
# Set up learning rate scheduler to decay learning rates dynamically
encoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
encoder_solver,
milestones=cfg.TRAIN.ENCODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
decoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
decoder_solver,
milestones=cfg.TRAIN.DECODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
refiner_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
refiner_solver,
milestones=cfg.TRAIN.REFINER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
merger_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
merger_solver,
milestones=cfg.TRAIN.MERGER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
if torch.cuda.is_available():
encoder = torch.nn.DataParallel(encoder).cuda()
decoder = torch.nn.DataParallel(decoder).cuda()
refiner = torch.nn.DataParallel(refiner).cuda()
merger = torch.nn.DataParallel(merger).cuda()
# Set up loss functions
bce_loss = torch.nn.BCELoss()
# Load pretrained model if exists
init_epoch = 0
best_iou = -1
best_epoch = -1
if 'WEIGHTS' in cfg.CONST and cfg.TRAIN.RESUME_TRAIN:
print('[INFO] %s Recovering from %s ...' %
(dt.now(), cfg.CONST.WEIGHTS))
checkpoint = torch.load(cfg.CONST.WEIGHTS)
init_epoch = checkpoint['epoch_idx']
best_iou = checkpoint['best_iou']
best_epoch = checkpoint['best_epoch']
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
merger.load_state_dict(checkpoint['merger_state_dict'])
print('[INFO] %s Recover complete. Current epoch #%d, Best IoU = %.4f at epoch #%d.' \
% (dt.now(), init_epoch, best_iou, best_epoch))
# Summary writer for TensorBoard
output_dir = os.path.join(cfg.DIR.OUT_PATH, '%s', dt.now().isoformat())
log_dir = output_dir % 'logs'
ckpt_dir = output_dir % 'checkpoints'
train_writer = SummaryWriter(os.path.join(log_dir, 'train'))
val_writer = SummaryWriter(os.path.join(log_dir, 'test'))
# Training loop
for epoch_idx in range(init_epoch, cfg.TRAIN.NUM_EPOCHES):
# Tick / tock
epoch_start_time = time()
# Batch average meterics
batch_time = utils.network_utils.AverageMeter()
data_time = utils.network_utils.AverageMeter()
encoder_losses = utils.network_utils.AverageMeter()
refiner_losses = utils.network_utils.AverageMeter()
# Adjust learning rate
encoder_lr_scheduler.step()
decoder_lr_scheduler.step()
refiner_lr_scheduler.step()
merger_lr_scheduler.step()
# switch models to training mode
encoder.train()
decoder.train()
merger.train()
refiner.train()
batch_end_time = time()
n_batches = len(train_data_loader)
for batch_idx, (taxonomy_names, sample_names, rendering_images,
ground_truth_volumes) in enumerate(train_data_loader):
# Measure data time
data_time.update(time() - batch_end_time)
# Get data from data loader
rendering_images = utils.network_utils.var_or_cuda(
rendering_images)
ground_truth_volumes = utils.network_utils.var_or_cuda(
ground_truth_volumes)
# Train the encoder, decoder, refiner, and merger
image_features = encoder(rendering_images)
raw_features, generated_volumes = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volumes = merger(raw_features, generated_volumes)
else:
generated_volumes = torch.mean(generated_volumes, dim=1)
encoder_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volumes = refiner(generated_volumes)
refiner_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
else:
refiner_loss = encoder_loss
# Gradient decent
encoder.zero_grad()
decoder.zero_grad()
refiner.zero_grad()
merger.zero_grad()
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
encoder_loss.backward(retain_graph=True)
refiner_loss.backward()
else:
encoder_loss.backward()
encoder_solver.step()
decoder_solver.step()
refiner_solver.step()
merger_solver.step()
# Append loss to average metrics
encoder_losses.update(encoder_loss.item())
refiner_losses.update(refiner_loss.item())
# Append loss to TensorBoard
n_itr = epoch_idx * n_batches + batch_idx
train_writer.add_scalar('EncoderDecoder/BatchLoss',
encoder_loss.item(), n_itr)
train_writer.add_scalar('Refiner/BatchLoss', refiner_loss.item(),
n_itr)
# Tick / tock
batch_time.update(time() - batch_end_time)
batch_end_time = time()
print('[INFO] %s [Epoch %d/%d][Batch %d/%d] BatchTime = %.3f (s) DataTime = %.3f (s) EDLoss = %.4f RLoss = %.4f' % \
(dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES, batch_idx + 1, n_batches, \
batch_time.val, data_time.val, encoder_loss.item(), refiner_loss.item()))
# Append epoch loss to TensorBoard
train_writer.add_scalar('EncoderDecoder/EpochLoss', encoder_losses.avg,
epoch_idx + 1)
train_writer.add_scalar('Refiner/EpochLoss', refiner_losses.avg,
epoch_idx + 1)
# Tick / tock
epoch_end_time = time()
print('[INFO] %s Epoch [%d/%d] EpochTime = %.3f (s) EDLoss = %.4f RLoss = %.4f' %
(dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES, epoch_end_time - epoch_start_time, \
encoder_losses.avg, refiner_losses.avg))
# Update Rendering Views
if cfg.TRAIN.UPDATE_N_VIEWS_RENDERING:
n_views_rendering = random.randint(1, cfg.CONST.N_VIEWS_RENDERING)
train_data_loader.dataset.set_n_views_rendering(n_views_rendering)
print('[INFO] %s Epoch [%d/%d] Update #RenderingViews to %d' % \
(dt.now(), epoch_idx + 2, cfg.TRAIN.NUM_EPOCHES, n_views_rendering))
# Validate the training models
iou = test_net(cfg, epoch_idx + 1, output_dir, val_data_loader,
val_writer, encoder, decoder, refiner, merger)
# Save weights to file
if (epoch_idx + 1) % cfg.TRAIN.SAVE_FREQ == 0:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
utils.network_utils.save_checkpoints(cfg, \
os.path.join(ckpt_dir, 'ckpt-epoch-%04d.pth' % (epoch_idx + 1)), \
epoch_idx + 1, encoder, encoder_solver, decoder, decoder_solver, \
refiner, refiner_solver, merger, merger_solver, best_iou, best_epoch)
if iou > best_iou:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
best_iou = iou
best_epoch = epoch_idx + 1
utils.network_utils.save_checkpoints(cfg, \
os.path.join(ckpt_dir, 'best-ckpt.pth'), \
epoch_idx + 1, encoder, encoder_solver, decoder, decoder_solver, \
refiner, refiner_solver, merger, merger_solver, best_iou, best_epoch)
# Close SummaryWriter for TensorBoard
train_writer.close()
val_writer.close()
def train_net(cfg):
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
train_transforms = utils.data_transforms.Compose([
utils.data_transforms.RandomCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(
cfg.TRAIN.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.ColorJitter(cfg.TRAIN.BRIGHTNESS,
cfg.TRAIN.CONTRAST,
cfg.TRAIN.SATURATION),
utils.data_transforms.RandomNoise(cfg.TRAIN.NOISE_STD),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.RandomFlip(),
utils.data_transforms.RandomPermuteRGB(),
utils.data_transforms.ToTensor(),
])
val_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
# Set up data loader
train_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TRAIN_DATASET](cfg)
val_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TEST_DATASET](cfg)
train_data_loader = paddle.io.DataLoader(
dataset=train_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.TRAIN, cfg.CONST.N_VIEWS_RENDERING,
train_transforms),
batch_size=cfg.CONST.BATCH_SIZE,
#num_workers=0 , # cfg.TRAIN.NUM_WORKER>0时报错，因为dev/shm/太小 https://blog.csdn.net/ctypyb2002/article/details/107914643
#pin_memory=True,
use_shared_memory=False,
shuffle=True,
drop_last=True)
val_data_loader = paddle.io.DataLoader(
dataset=val_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.VAL, cfg.CONST.N_VIEWS_RENDERING,
val_transforms),
batch_size=1,
#num_workers=1,
#pin_memory=True,
shuffle=False)
# Set up networks # paddle.Model prepare fit save
encoder = Encoder(cfg)
decoder = Decoder(cfg)
merger = Merger(cfg)
refiner = Refiner(cfg)
print('[DEBUG] %s Parameters in Encoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(encoder)))
print('[DEBUG] %s Parameters in Decoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(decoder)))
print('[DEBUG] %s Parameters in Merger: %d.' %
(dt.now(), utils.network_utils.count_parameters(merger)))
print('[DEBUG] %s Parameters in Refiner: %d.' %
(dt.now(), utils.network_utils.count_parameters(refiner)))
# # Initialize weights of networks # paddle的参数化不同，参见API
# encoder.apply(utils.network_utils.init_weights)
# decoder.apply(utils.network_utils.init_weights)
# merger.apply(utils.network_utils.init_weights)
# Set up learning rate scheduler to decay learning rates dynamically
encoder_lr_scheduler = paddle.optimizer.lr.MultiStepDecay(
learning_rate=cfg.TRAIN.ENCODER_LEARNING_RATE,
milestones=cfg.TRAIN.ENCODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA,
verbose=True)
decoder_lr_scheduler = paddle.optimizer.lr.MultiStepDecay(
learning_rate=cfg.TRAIN.DECODER_LEARNING_RATE,
milestones=cfg.TRAIN.DECODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA,
verbose=True)
merger_lr_scheduler = paddle.optimizer.lr.MultiStepDecay(
learning_rate=cfg.TRAIN.MERGER_LEARNING_RATE,
milestones=cfg.TRAIN.MERGER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA,
verbose=True)
refiner_lr_scheduler = paddle.optimizer.lr.MultiStepDecay(
learning_rate=cfg.TRAIN.REFINER_LEARNING_RATE,
milestones=cfg.TRAIN.REFINER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA,
verbose=True)
# Set up solver
# if cfg.TRAIN.POLICY == 'adam':
encoder_solver = paddle.optimizer.Adam(learning_rate=encoder_lr_scheduler,
parameters=encoder.parameters())
decoder_solver = paddle.optimizer.Adam(learning_rate=decoder_lr_scheduler,
parameters=decoder.parameters())
merger_solver = paddle.optimizer.Adam(learning_rate=merger_lr_scheduler,
parameters=merger.parameters())
refiner_solver = paddle.optimizer.Adam(learning_rate=refiner_lr_scheduler,
parameters=refiner.parameters())
# if torch.cuda.is_available():
# encoder = torch.nn.DataParallel(encoder).cuda()
# decoder = torch.nn.DataParallel(decoder).cuda()
# merger = torch.nn.DataParallel(merger).cuda()
# Set up loss functions
bce_loss = paddle.nn.BCELoss()
# Load pretrained model if exists
init_epoch = 0
best_iou = -1
best_epoch = -1
if 'WEIGHTS' in cfg.CONST and cfg.TRAIN.RESUME_TRAIN:
print('[INFO] %s Recovering from %s ...' %
(dt.now(), cfg.CONST.WEIGHTS))
# load
encoder_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "encoder.pdparams"))
encoder_solver_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "encoder_solver.pdopt"))
encoder.set_state_dict(encoder_state_dict)
encoder_solver.set_state_dict(encoder_solver_state_dict)
decoder_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "decoder.pdparams"))
decoder_solver_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "decoder_solver.pdopt"))
decoder.set_state_dict(decoder_state_dict)
decoder_solver.set_state_dict(decoder_solver_state_dict)
if cfg.NETWORK.USE_MERGER:
merger_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "merger.pdparams"))
merger_solver_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "merger_solver.pdopt"))
merger.set_state_dict(merger_state_dict)
merger_solver.set_state_dict(merger_solver_state_dict)
if cfg.NETWORK.USE_REFINER:
refiner_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "refiner.pdparams"))
refiner_solver_state_dict = paddle.load(
os.path.join(cfg.CONST.WEIGHTS, "refiner_solver.pdopt"))
refiner.set_state_dict(refiner_state_dict)
refiner_solver.set_state_dict(refiner_solver_state_dict)
print(
'[INFO] %s Recover complete. Current epoch #%d, Best IoU = %.4f at epoch #%d.'
% (dt.now(), init_epoch, best_iou, best_epoch))
# Summary writer for TensorBoard
output_dir = os.path.join(cfg.DIR.OUT_PATH, '%s', dt.now().isoformat())
log_dir = output_dir % 'logs'
ckpt_dir = output_dir % 'checkpoints'
# train_writer = SummaryWriter()
# val_writer = SummaryWriter(os.path.join(log_dir, 'test'))
train_writer = LogWriter(os.path.join(log_dir, 'train'))
val_writer = LogWriter(os.path.join(log_dir, 'val'))
# Training loop
for epoch_idx in range(init_epoch, cfg.TRAIN.NUM_EPOCHES):
# Tick / tock
epoch_start_time = time()
# Batch average meterics
batch_time = utils.network_utils.AverageMeter()
data_time = utils.network_utils.AverageMeter()
encoder_losses = utils.network_utils.AverageMeter()
refiner_losses = utils.network_utils.AverageMeter()
# # switch models to training mode
encoder.train()
decoder.train()
merger.train()
refiner.train()
batch_end_time = time()
n_batches = len(train_data_loader)
# print("****debug: length of train data loder",n_batches)
for batch_idx, (rendering_images, ground_truth_volumes) in enumerate(
train_data_loader()):
# # debug
# if batch_idx>1:
# break
# Measure data time
data_time.update(time() - batch_end_time)
# print("****debug: batch_idx",batch_idx)
# print(rendering_images.shape)
# print(ground_truth_volumes.shape)
# Get data from data loader
rendering_images = utils.network_utils.var_or_cuda(
rendering_images)
ground_truth_volumes = utils.network_utils.var_or_cuda(
ground_truth_volumes)
# Train the encoder, decoder, and merger
image_features = encoder(rendering_images)
raw_features, generated_volumes = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volumes = merger(raw_features, generated_volumes)
# else:
# mergered_volumes = paddle.mean(generated_volumes, aixs=1)
encoder_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volumes = refiner(generated_volumes)
refiner_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
# else:
# refiner_loss = encoder_loss
# Gradient decent
encoder_solver.clear_grad()
decoder_solver.clear_grad()
merger_solver.clear_grad()
refiner_solver.clear_grad()
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
encoder_loss.backward(retain_graph=True)
refiner_loss.backward()
# else:
# encoder_loss.backward()
encoder_solver.step()
decoder_solver.step()
merger_solver.step()
refiner_solver.step()
# Append loss to average metrics
encoder_losses.update(encoder_loss.numpy())
refiner_losses.update(refiner_loss.numpy())
# Append loss to TensorBoard
n_itr = epoch_idx * n_batches + batch_idx
train_writer.add_scalar(tag='EncoderDecoder/BatchLoss',
step=n_itr,
value=encoder_loss.numpy())
train_writer.add_scalar('Refiner/BatchLoss',
value=refiner_loss.numpy(),
step=n_itr)
# Tick / tock
batch_time.update(time() - batch_end_time)
batch_end_time = time()
if (batch_idx % int(cfg.CONST.INFO_BATCH)) == 0:
print(
'[INFO] %s [Epoch %d/%d][Batch %d/%d] BatchTime = %.3f (s) DataTime = %.3f (s) EDLoss = %.4f RLoss = %.4f'
% (dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES,
batch_idx + 1, n_batches, batch_time.val, data_time.val,
encoder_loss.numpy(), refiner_loss.numpy()))
# Append epoch loss to TensorBoard
train_writer.add_scalar(tag='EncoderDecoder/EpochLoss',
step=epoch_idx + 1,
value=encoder_losses.avg)
train_writer.add_scalar('Refiner/EpochLoss',
value=refiner_losses.avg,
step=epoch_idx + 1)
# update scheduler each step
encoder_lr_scheduler.step()
decoder_lr_scheduler.step()
merger_lr_scheduler.step()
refiner_lr_scheduler.step()
# Tick / tock
epoch_end_time = time()
print(
'[INFO] %s Epoch [%d/%d] EpochTime = %.3f (s) EDLoss = %.4f RLoss = %.4f'
% (dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES, epoch_end_time -
epoch_start_time, encoder_losses.avg, refiner_losses.avg))
# Update Rendering Views
if cfg.TRAIN.UPDATE_N_VIEWS_RENDERING:
n_views_rendering = random.randint(1, cfg.CONST.N_VIEWS_RENDERING)
train_data_loader.dataset.set_n_views_rendering(n_views_rendering)
print('[INFO] %s Epoch [%d/%d] Update #RenderingViews to %d' %
(dt.now(), epoch_idx + 2, cfg.TRAIN.NUM_EPOCHES,
n_views_rendering))
# Validate the training models
iou = test_net(cfg, epoch_idx + 1, output_dir, val_data_loader,
val_writer, encoder, decoder, merger, refiner)
# Save weights to file
if (epoch_idx + 1) % cfg.TRAIN.SAVE_FREQ == 0:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
utils.network_utils.save_checkpoints(
cfg, os.path.join(ckpt_dir,
'ckpt-epoch-%04d' % (epoch_idx + 1)),
epoch_idx + 1, encoder, encoder_solver, decoder,
decoder_solver, merger, merger_solver, refiner, refiner_solver,
best_iou, best_epoch)
if iou > best_iou:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
best_iou = iou
best_epoch = epoch_idx + 1
utils.network_utils.save_checkpoints(
cfg, os.path.join(ckpt_dir, 'best-ckpt'), epoch_idx + 1,
encoder, encoder_solver, decoder, decoder_solver, merger,
merger_solver, refiner, refiner_solver, best_iou, best_epoch)
class VaeGanModule(pl.LightningModule):
def __init__(self, hparams):
super().__init__()
self.hparams = hparams
# Encoder
self.encoder = Encoder(ngf=self.hparams.ngf, z_dim=self.hparams.z_dim)
self.encoder.apply(weights_init)
device = "cuda" if isinstance(self.hparams.gpus, int) else "cpu"
# Decoder
self.decoder = Decoder(ngf=self.hparams.ngf, z_dim=self.hparams.z_dim)
self.decoder.apply(weights_init)
# Discriminator
self.discriminator = Discriminator()
self.discriminator.apply(weights_init)
# Losses
self.criterionFeat = torch.nn.L1Loss()
self.criterionGAN = GANLoss(gan_mode="lsgan")
if self.hparams.use_vgg:
self.criterion_perceptual_style = [Perceptual_Loss(device)]
@staticmethod
def reparameterize(mu, logvar, mode='train'):
if mode == 'train':
std = torch.exp(0.5 * logvar)
eps = torch.randn_like(std)
return mu + eps * std
else:
return mu
def discriminate(self, fake_image, real_image):
input_concat_fake = torch.cat(
(fake_image.detach(), real_image),
dim=1) # non sono sicuro che .detach() sia necessario in lightning
input_concat_real = torch.cat((real_image, real_image), dim=1)
return (self.discriminator(input_concat_fake),
self.discriminator(input_concat_real))
def training_step(self, batch, batch_idx, optimizer_idx):
x, _ = batch
# train VAE
if optimizer_idx == 0:
# encode
mu, log_var = self.encoder(x)
z_repar = VaeGanModule.reparameterize(mu, log_var)
# decode
fake_image = self.decoder(z_repar)
# reconstruction
reconstruction_loss = self.criterionFeat(fake_image, x)
kld_loss = -0.5 * torch.mean(1 + log_var - mu.pow(2) -
log_var.exp())
# Discriminate
input_concat_fake = torch.cat((fake_image, x), dim=1)
pred_fake = self.discriminator(input_concat_fake)
# Losses
loss_G_GAN = self.criterionGAN(pred_fake, True)
if self.hparams.use_vgg:
loss_G_perceptual = \
self.criterion_perceptual_style[0](fake_image, x)
else:
loss_G_perceptual = 0.0
g_loss = (reconstruction_loss *
20) + kld_loss + loss_G_GAN + loss_G_perceptual
# Results are collected in a TrainResult object
result = pl.TrainResult(g_loss)
result.log("rec_loss", reconstruction_loss * 10, prog_bar=True)
result.log("loss_G_GAN", loss_G_GAN, prog_bar=True)
result.log("kld_loss", kld_loss, prog_bar=True)
result.log("loss_G_perceptual", loss_G_perceptual, prog_bar=True)
# train Discriminator
if optimizer_idx == 1:
# Measure discriminator's ability to classify real from generated samples
# Encode
mu, log_var = self.encoder(x)
z_repar = VaeGanModule.reparameterize(mu, log_var)
# Decode
fake_image = self.decoder(z_repar)
# how well can it label as real?
pred_fake, pred_real = self.discriminate(fake_image, x)
# Fake loss
d_loss_fake = self.criterionGAN(pred_fake, False)
# Real Loss
d_loss_real = self.criterionGAN(pred_real, True)
# Total loss is average of prediction of fakes and reals
loss_D = (d_loss_fake + d_loss_real) / 2
# Results are collected in a TrainResult object
result = pl.TrainResult(loss_D)
result.log("loss_D_real", d_loss_real, prog_bar=True)
result.log("loss_D_fake", d_loss_fake, prog_bar=True)
return result
def training_epoch_end(self, training_step_outputs):
z_appr = torch.normal(mean=0,
std=1,
size=(16, self.hparams.z_dim),
device=training_step_outputs[0].minimize.device)
# Generate images from latent vector
sample_imgs = self.decoder(z_appr)
grid = torchvision.utils.make_grid(sample_imgs,
normalize=True,
range=(-1, 1))
# where to save the image
path = os.path.join(self.hparams.generated_images_folder,
f"generated_images_{self.current_epoch}.png")
torchvision.utils.save_image(sample_imgs,
path,
normalize=True,
range=(-1, 1))
# Log images in tensorboard
self.logger.experiment.add_image(f'generated_images', grid,
self.current_epoch)
# Epoch level metrics
epoch_loss = torch.mean(
torch.stack([x['minimize'] for x in training_step_outputs]))
results = pl.TrainResult()
results.log("epoch_loss", epoch_loss, prog_bar=False)
return results
def validation_step(self, batch, batch_idx):
x, _ = batch
# Encode
mu, log_var = self.encoder(x)
z_repar = VaeGanModule.reparameterize(mu, log_var)
# Decode
recons = self.decoder(z_repar)
reconstruction_loss = nn.functional.mse_loss(recons, x)
# Results are collected in a EvalResult object
result = pl.EvalResult(checkpoint_on=reconstruction_loss)
return result
testing_step = validation_step
def configure_optimizers(self):
params_vae = concat_generators(self.encoder.parameters(),
self.decoder.parameters())
opt_vae = torch.optim.Adam(params_vae,
lr=self.hparams.learning_rate_vae)
parameters_discriminator = self.discriminator.parameters()
opt_d = torch.optim.Adam(parameters_discriminator,
lr=self.hparams.learning_rate_d)
return [opt_vae, opt_d]
@staticmethod
def add_argparse_args(parser):
parser.add_argument('--generated_images_folder',
required=False,
default="./output",
type=str)
parser.add_argument('--ngf', type=int, default=128)
parser.add_argument('--z_dim', type=int, default=128)
parser.add_argument('--learning_rate_vae',
default=1e-03,
required=False,
type=float)
parser.add_argument('--learning_rate_d',
default=1e-03,
required=False,
type=float)
parser.add_argument("--use_vgg", action="store_true", default=False)
return parser
class Trainer:
def __init__(self, config):
self.config = config
self.config.data.n_datasets = len(config.data.datasets)
print("No of datasets used:", self.config.data.n_datasets)
torch.manual_seed(config.env.seed)
torch.cuda.manual_seed(config.env.seed)
self.expPath = self.config.env.expPath
self.logger = Logger("Training", "logs/training.log")
self.data = [
DatasetSet(data_path, config.data.seq_len, config.data)
for data_path in config.data.datasets
]
self.losses_recon = [
LossMeter(f'recon {i}') for i in range(self.config.data.n_datasets)
]
self.loss_d_right = LossMeter('d')
self.loss_total = [
LossMeter(f'total {i}') for i in range(self.config.data.n_datasets)
]
self.evals_recon = [
LossMeter(f'recon {i}') for i in range(self.config.data.n_datasets)
]
self.eval_d_right = LossMeter('eval d')
self.eval_total = [
LossMeter(f'eval total {i}')
for i in range(self.config.data.n_datasets)
]
self.encoder = Encoder(config.encoder)
self.decoders = torch.nn.ModuleList([
Decoder(config.decoder) for _ in range(self.config.data.n_datasets)
])
self.classifier = DomainClassifier(
config.domain_classifier, num_classes=self.config.data.n_datasets)
states = None
if config.env.checkpoint:
checkpoint_args_path = os.path.dirname(
config.env.checkpoint) + '/args.pth'
checkpoint_args = torch.load(checkpoint_args_path)
self.start_epoch = checkpoint_args[-1] + 1
states = [
torch.load(self.config.env.checkpoint + f'_{i}.pth')
for i in range(self.config.data.n_datasets)
]
self.encoder.load_state_dict(states[0]['encoder_state'])
for i in range(self.config.data.n_datasets):
self.decoders[i].load_state_dict(states[i]['decoder_state'])
self.classifier.load_state_dict(states[0]['discriminator_state'])
self.logger.info('Loaded checkpoint parameters')
raise NotImplementedError
else:
self.start_epoch = 0
self.encoder = torch.nn.DataParallel(self.encoder).cuda()
self.classifier = torch.nn.DataParallel(self.classifier).cuda()
for i, decoder in enumerate(self.decoders):
self.decoders[i] = torch.nn.DataParallel(decoder).cuda()
self.model_optimizers = [
optim.Adam(chain(self.encoder.parameters(), decoder.parameters()),
lr=config.data.lr) for decoder in self.decoders
]
self.classifier_optimizer = optim.Adam(self.classifier.parameters(),
lr=config.data.lr)
if config.env.checkpoint and config.env.load_optimizer:
for i in range(self.config.data.n_datasets):
self.model_optimizers[i].load_state_dict(
states[i]['model_optimizer_state'])
self.classifier_optimizer.load_state_dict(
states[0]['d_optimizer_state'])
self.lr_managers = []
for i in range(self.config.data.n_datasets):
self.lr_managers.append(
torch.optim.lr_scheduler.ExponentialLR(
self.model_optimizers[i], config.data.lr_decay))
self.lr_managers[i].last_epoch = self.start_epoch
self.lr_managers[i].step()
def eval_batch(self, x, x_aug, dset_num):
x, x_aug = x.float(), x_aug.float()
z = self.encoder(x)
y = self.decoders[dset_num](x, z)
z_logits = self.classifier(z)
z_classification = torch.max(z_logits, dim=1)[1]
z_accuracy = (z_classification == dset_num).float().mean()
self.eval_d_right.add(z_accuracy.data.item())
# discriminator_right = F.cross_entropy(z_logits, dset_num).mean()
discriminator_right = F.cross_entropy(
z_logits,
torch.tensor([dset_num] * x.size(0)).long().cuda()).mean()
recon_loss = cross_entropy_loss(y, x)
self.evals_recon[dset_num].add(recon_loss.data.cpu().numpy().mean())
total_loss = discriminator_right.data.item() * self.config.domain_classifier.d_lambda + \
recon_loss.mean().data.item()
self.eval_total[dset_num].add(total_loss)
return total_loss
def train_batch(self, x, x_aug, dset_num):
x, x_aug = x.float(), x_aug.float()
# Optimize D - classifier right
z = self.encoder(x)
z_logits = self.classifier(z)
discriminator_right = F.cross_entropy(
z_logits,
torch.tensor([dset_num] * x.size(0)).long().cuda()).mean()
loss = discriminator_right * self.config.domain_classifier.d_lambda
self.loss_d_right.add(loss.data.item())
self.classifier_optimizer.zero_grad()
loss.backward()
if self.config.domain_classifier.grad_clip is not None:
clip_grad_value_(self.classifier.parameters(),
self.config.domain_classifier.grad_clip)
self.classifier_optimizer.step()
# optimize G - reconstructs well, classifier wrong
z = self.encoder(x_aug)
y = self.decoders[dset_num](x, z)
z_logits = self.classifier(z)
discriminator_wrong = -F.cross_entropy(
z_logits,
torch.tensor([dset_num] * x.size(0)).long().cuda()).mean()
if not (-100 < discriminator_right.data.item() < 100):
self.logger.debug(f'z_logits: {z_logits.detach().cpu().numpy()}')
self.logger.debug(f'dset_num: {dset_num}')
recon_loss = cross_entropy_loss(y, x)
self.losses_recon[dset_num].add(recon_loss.data.cpu().numpy().mean())
loss = (recon_loss.mean() +
self.config.domain_classifier.d_lambda * discriminator_wrong)
self.model_optimizers[dset_num].zero_grad()
loss.backward()
if self.config.domain_classifier.grad_clip is not None:
clip_grad_value_(self.encoder.parameters(),
self.config.domain_classifier.grad_clip)
clip_grad_value_(self.decoders[dset_num].parameters(),
self.config.domain_classifier.grad_clip)
self.model_optimizers[dset_num].step()
self.loss_total[dset_num].add(loss.data.item())
return loss.data.item()
def train_epoch(self, epoch):
for meter in self.losses_recon:
meter.reset()
self.loss_d_right.reset()
for i in range(len(self.loss_total)):
self.loss_total[i].reset()
self.encoder.train()
self.classifier.train()
for decoder in self.decoders:
decoder.train()
n_batches = self.config.data.epoch_len
with tqdm(total=n_batches,
desc='Train epoch %d' % epoch) as train_enum:
for batch_num in range(n_batches):
if self.config.data.short and batch_num == 3:
break
dset_num = batch_num % self.config.data.n_datasets
x, x_aug = next(self.data[dset_num].train_iter)
x = wrap_cuda(x)
x_aug = wrap_cuda(x_aug)
batch_loss = self.train_batch(x, x_aug, dset_num)
train_enum.set_description(
f'Train (loss: {batch_loss:.2f}) epoch {epoch}')
train_enum.update()
def evaluate_epoch(self, epoch):
for meter in self.evals_recon:
meter.reset()
self.eval_d_right.reset()
for i in range(len(self.eval_total)):
self.eval_total[i].reset()
self.encoder.eval()
self.classifier.eval()
for decoder in self.decoders:
decoder.eval()
n_batches = int(np.ceil(self.config.data.epoch_len / 10))
with tqdm(total=n_batches) as valid_enum, \
torch.no_grad():
for batch_num in range(n_batches):
if self.config.data.short and batch_num == 10:
break
dset_num = batch_num % self.config.data.n_datasets
x, x_aug = next(self.data[dset_num].valid_iter)
x = wrap_cuda(x)
x_aug = wrap_cuda(x_aug)
batch_loss = self.eval_batch(x, x_aug, dset_num)
valid_enum.set_description(
f'Test (loss: {batch_loss:.2f}) epoch {epoch}')
valid_enum.update()
@staticmethod
def format_losses(meters):
losses = [meter.summarize_epoch() for meter in meters]
return ', '.join('{:.4f}'.format(x) for x in losses)
def train_losses(self):
meters = [*self.losses_recon, self.loss_d_right]
return self.format_losses(meters)
def eval_losses(self):
meters = [*self.evals_recon, self.eval_d_right]
return self.format_losses(meters)
def train(self):
best_eval = [float('inf') for _ in range(self.config.data.n_datasets)]
# Begin!
for epoch in range(self.start_epoch,
self.start_epoch + self.config.env.epochs):
self.train_epoch(epoch)
self.evaluate_epoch(epoch)
self.logger.info(f'Epoch %s - Train loss: (%s), Test loss (%s)',
epoch, self.train_losses(), self.eval_losses())
for i in range(len(self.lr_managers)):
self.lr_managers[i].step()
for dataset_id in range(self.config.data.n_datasets):
val_loss = self.eval_total[dataset_id].summarize_epoch()
if val_loss < best_eval[dataset_id]:
self.save_model(f'bestmodel_{dataset_id}.pth', dataset_id)
best_eval[dataset_id] = val_loss
if not self.config.env.save_per_epoch:
self.save_model(f'lastmodel_{dataset_id}.pth', dataset_id)
else:
self.save_model(f'lastmodel_{epoch}_rank_{dataset_id}.pth',
dataset_id)
torch.save([self.config, epoch], '%s/args.pth' % self.expPath)
self.logger.debug('Ended epoch')
def save_model(self, filename, decoder_id):
save_path = self.expPath / filename
torch.save(
{
'encoder_state':
self.encoder.module.state_dict(),
'decoder_state':
self.decoders[decoder_id].module.state_dict(),
'discriminator_state':
self.classifier.module.state_dict(),
'model_optimizer_state':
self.model_optimizers[decoder_id].state_dict(),
'dataset':
decoder_id,
'd_optimizer_state':
self.classifier_optimizer.state_dict()
}, save_path)
self.logger.debug(f'Saved model to {save_path}')
class Model(pl.LightningModule):
def __init__(self, cfg_network: DictConfig, cfg_tester: DictConfig):
super().__init__()
self.cfg_network = cfg_network
self.cfg_tester = cfg_tester
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Set up networks
self.encoder = Encoder(cfg_network)
self.decoder = Decoder(cfg_network)
self.refiner = Refiner(cfg_network)
self.merger = Merger(cfg_network)
# Initialize weights of networks
self.encoder.apply(utils.network_utils.init_weights)
self.decoder.apply(utils.network_utils.init_weights)
self.refiner.apply(utils.network_utils.init_weights)
self.merger.apply(utils.network_utils.init_weights)
self.bce_loss = nn.BCELoss()
def configure_optimizers(self):
params = self.cfg_network.optimization
# Set up solver
if params.policy == 'adam':
encoder_solver = optim.Adam(filter(lambda p: p.requires_grad, self.encoder.parameters()),
lr=params.encoder_lr,
betas=params.betas)
decoder_solver = optim.Adam(self.decoder.parameters(),
lr=params.decoder_lr,
betas=params.betas)
refiner_solver = optim.Adam(self.refiner.parameters(),
lr=params.refiner_lr,
betas=params.betas)
merger_solver = optim.Adam(self.merger.parameters(),
lr=params.merger_lr,
betas=params.betas)
elif params.policy == 'sgd':
encoder_solver = optim.SGD(filter(lambda p: p.requires_grad, self.encoder.parameters()),
lr=params.encoder_lr,
momentum=params.momentum)
decoder_solver = optim.SGD(self.decoder.parameters(),
lr=params.decoder_lr,
momentum=params.momentum)
refiner_solver = optim.SGD(self.refiner.parameters(),
lr=params.refiner_lr,
momentum=params.momentum)
merger_solver = optim.SGD(self.merger.parameters(),
lr=params.merger_lr,
momentum=params.momentum)
else:
raise Exception('[FATAL] %s Unknown optimizer %s.' % (dt.now(), params.policy))
# Set up learning rate scheduler to decay learning rates dynamically
encoder_lr_scheduler = optim.lr_scheduler.MultiStepLR(encoder_solver,
milestones=params.encoder_lr_milestones,
gamma=params.gamma)
decoder_lr_scheduler = optim.lr_scheduler.MultiStepLR(decoder_solver,
milestones=params.decoder_lr_milestones,
gamma=params.gamma)
refiner_lr_scheduler = optim.lr_scheduler.MultiStepLR(refiner_solver,
milestones=params.refiner_lr_milestones,
gamma=params.gamma)
merger_lr_scheduler = optim.lr_scheduler.MultiStepLR(merger_solver,
milestones=params.merger_lr_milestones,
gamma=params.gamma)
return [encoder_solver, decoder_solver, refiner_solver, merger_solver], \
[encoder_lr_scheduler, decoder_lr_scheduler, refiner_lr_scheduler, merger_lr_scheduler]
def _fwd(self, batch):
taxonomy_names, sample_names, rendering_images, ground_truth_volumes = batch
image_features = self.encoder(rendering_images)
raw_features, generated_volumes = self.decoder(image_features)
if self.cfg_network.use_merger and self.current_epoch >= self.cfg_network.optimization.epoch_start_use_merger:
generated_volumes = self.merger(raw_features, generated_volumes)
else:
generated_volumes = torch.mean(generated_volumes, dim=1)
encoder_loss = self.bce_loss(generated_volumes, ground_truth_volumes) * 10
if self.cfg_network.use_refiner and self.current_epoch >= self.cfg_network.optimization.epoch_start_use_refiner:
generated_volumes = self.refiner(generated_volumes)
refiner_loss = self.bce_loss(generated_volumes, ground_truth_volumes) * 10
else:
refiner_loss = encoder_loss
return generated_volumes, encoder_loss, refiner_loss
def training_step(self, batch, batch_idx, optimizer_idx):
(opt_enc, opt_dec, opt_ref, opt_merg) = self.optimizers()
generated_volumes, encoder_loss, refiner_loss = self._fwd(batch)
self.log('loss/EncoderDecoder', encoder_loss,
prog_bar=True, logger=True, on_step=True, on_epoch=True)
self.log('loss/Refiner', refiner_loss,
prog_bar=True, logger=True, on_step=True, on_epoch=True)
if self.cfg_network.use_refiner and self.current_epoch >= self.cfg_network.optimization.epoch_start_use_refiner:
self.manual_backward(encoder_loss, opt_enc, retain_graph=True)
self.manual_backward(refiner_loss, opt_ref)
else:
self.manual_backward(encoder_loss, opt_enc)
for opt in self.optimizers():
opt.step()
opt.zero_grad()
def training_epoch_end(self, outputs) -> None:
# Update Rendering Views
if self.cfg_network.update_n_views_rendering:
n_views_rendering = self.trainer.datamodule.update_n_views_rendering()
print('[INFO] %s Epoch [%d/%d] Update #RenderingViews to %d' %
(dt.now(), self.current_epoch + 2, self.trainer.max_epochs, n_views_rendering))
def _eval_step(self, batch, batch_idx):
# SUPPORTS ONLY BATCH_SIZE=1
taxonomy_names, sample_names, rendering_images, ground_truth_volumes = batch
taxonomy_id = taxonomy_names[0]
sample_name = sample_names[0]
generated_volumes, encoder_loss, refiner_loss = self._fwd(batch)
self.log('val_loss/EncoderDecoder', encoder_loss, prog_bar=True,
logger=True, on_step=True, on_epoch=True)
self.log('val_loss/Refiner', refiner_loss, prog_bar=True,
logger=True, on_step=True, on_epoch=True)
# IoU per sample
sample_iou = []
for th in self.cfg_tester.voxel_thresh:
_volume = torch.ge(generated_volumes, th).float()
intersection = torch.sum(_volume.mul(ground_truth_volumes)).float()
union = torch.sum(
torch.ge(_volume.add(ground_truth_volumes), 1)).float()
sample_iou.append((intersection / union).item())
# Print sample loss and IoU
n_samples = -1
print('\n[INFO] %s Test[%d/%d] Taxonomy = %s Sample = %s EDLoss = %.4f RLoss = %.4f IoU = %s' %
(dt.now(), batch_idx + 1, n_samples, taxonomy_id, sample_name, encoder_loss.item(),
refiner_loss.item(), ['%.4f' % si for si in sample_iou]))
return {
'taxonomy_id': taxonomy_id,
'sample_name': sample_name,
'sample_iou': sample_iou
}
def _eval_epoch_end(self, outputs):
# Load taxonomies of dataset
taxonomies = []
taxonomy_path = self.trainer.datamodule.get_test_taxonomy_file_path()
with open(taxonomy_path, encoding='utf-8') as file:
taxonomies = json.loads(file.read())
taxonomies = {t['taxonomy_id']: t for t in taxonomies}
test_iou = {}
for output in outputs:
taxonomy_id, sample_name, sample_iou = output[
'taxonomy_id'], output['sample_name'], output['sample_iou']
if taxonomy_id not in test_iou:
test_iou[taxonomy_id] = {'n_samples': 0, 'iou': []}
test_iou[taxonomy_id]['n_samples'] += 1
test_iou[taxonomy_id]['iou'].append(sample_iou)
mean_iou = []
for taxonomy_id in test_iou:
test_iou[taxonomy_id]['iou'] = torch.mean(
torch.tensor(test_iou[taxonomy_id]['iou']), dim=0)
mean_iou.append(test_iou[taxonomy_id]['iou']
* test_iou[taxonomy_id]['n_samples'])
n_samples = len(outputs)
mean_iou = torch.stack(mean_iou)
mean_iou = torch.sum(mean_iou, dim=0) / n_samples
# Print header
print('============================ TEST RESULTS ============================')
print('Taxonomy', end='\t')
print('#Sample', end='\t')
print(' Baseline', end='\t')
for th in self.cfg_tester.voxel_thresh:
print('t=%.2f' % th, end='\t')
print()
# Print body
for taxonomy_id in test_iou:
print('%s' % taxonomies[taxonomy_id]
['taxonomy_name'].ljust(8), end='\t')
print('%d' % test_iou[taxonomy_id]['n_samples'], end='\t')
if 'baseline' in taxonomies[taxonomy_id]:
n_views_rendering = self.trainer.datamodule.get_n_views_rendering()
print('%.4f' % taxonomies[taxonomy_id]['baseline']
['%d-view' % n_views_rendering], end='\t\t')
else:
print('N/a', end='\t\t')
for ti in test_iou[taxonomy_id]['iou']:
print('%.4f' % ti, end='\t')
print()
# Print mean IoU for each threshold
print('Overall ', end='\t\t\t\t')
for mi in mean_iou:
print('%.4f' % mi, end='\t')
print('\n')
max_iou = torch.max(mean_iou)
self.log('Refiner/IoU', max_iou, prog_bar=True, on_epoch=True)
def validation_step(self, batch, batch_idx):
return self._eval_step(batch, batch_idx)
def validation_epoch_end(self, outputs):
self._eval_epoch_end(outputs)
def test_step(self, batch, batch_idx):
return self._eval_step(batch, batch_idx)
def test_epoch_end(self, outputs):
self._eval_epoch_end(outputs)
def get_progress_bar_dict(self):
# don't show the loss as it's None
items = super().get_progress_bar_dict()
items.pop("loss", None)
return items
drop_last=True,
num_workers=4)
VOCAB_SIZE = train_dataset.vocab.num_words
SEQ_LEN = train_dataset.vocab.max_sentence_len
encoder = Encoder(args.ENCODER_OUTPUT_SIZE).to(device)
decoder = Decoder(embed_size=args.EMBED_SIZE,
hidden_size=args.HIDDEN_SIZE,
attention_size=args.ATTENTION_SIZE,
vocab_size=VOCAB_SIZE,
encoder_size=2048,
device=device,
seq_len=SEQ_LEN + 2).to(device)
encoder_optimizer = torch.optim.Adam(encoder.parameters(), lr=args.LR)
decoder_optimizer = torch.optim.Adam(decoder.parameters(), lr=args.LR)
criterion = nn.CrossEntropyLoss(
) # 나중에 loss 계산 할때 패딩은 모두 없앨것이므로 ignore index를 설정하지 않는다
train_losses = []
validation_losses = []
for epoch in range(args.NUM_EPOCHS):
train_loss = 0
validation_loss = 0
encoder.train()
decoder.train()
for idx, (img, caption_5,
caption_lengths_5) in enumerate(train_loader):
