def train(model: Hidden, device: torch.device,
hidden_config: HiDDenConfiguration, train_options: TrainingOptions,
this_run_folder: str, tb_logger):
"""
Trains the HiDDeN model
:param model: The model
:param device: torch.device object, usually this is GPU (if avaliable), otherwise CPU.
:param hidden_config: The network configuration
:param train_options: The training settings
:param this_run_folder: The parent folder for the current training run to store training artifacts/results/logs.
:param tb_logger: TensorBoardLogger object which is a thin wrapper for TensorboardX logger.
Pass None to disable TensorboardX logging
:return:
"""
train_data, val_data = utils.get_data_loaders(hidden_config, train_options)
images_to_save = 8
saved_images_size = (512, 512)
best_epoch = train_options.best_epoch
best_cond = train_options.best_cond
for epoch in range(train_options.start_epoch,
train_options.number_of_epochs + 1):
logging.info(
f'\nStarting epoch {epoch}/{train_options.number_of_epochs} [{best_epoch}]'
)
training_losses = defaultdict(functions.AverageMeter)
epoch_start = time.time()
for image, _ in tqdm(train_data, ncols=80):
image = image.to(device) #.squeeze(0)
message = torch.Tensor(
np.random.choice(
[0, 1],
(image.shape[0], hidden_config.message_length))).to(device)
losses, _ = model.train_on_batch([image, message])
for name, loss in losses.items():
training_losses[name].update(loss)
train_duration = time.time() - epoch_start
logging.info('Epoch {} training duration {:.2f} sec'.format(
epoch, train_duration))
logging.info('-' * 40)
utils.write_losses(os.path.join(this_run_folder, 'train.csv'),
training_losses, epoch, train_duration)
if tb_logger is not None:
tb_logger.save_losses('train_loss', training_losses, epoch)
tb_logger.save_grads(epoch)
tb_logger.save_tensors(epoch)
tb_logger.writer.flush()
validation_losses = defaultdict(functions.AverageMeter)
logging.info('Running validation for epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
val_image_patches = ()
val_encoded_patches = ()
val_noised_patches = ()
for image, _ in tqdm(val_data, ncols=80):
image = image.to(device) #.squeeze(0)
message = torch.Tensor(
np.random.choice(
[0, 1],
(image.shape[0], hidden_config.message_length))).to(device)
losses, (encoded_images, noised_images,
decoded_messages) = model.validate_on_batch(
[image, message])
for name, loss in losses.items():
validation_losses[name].update(loss)
if hidden_config.enable_fp16:
image = image.float()
encoded_images = encoded_images.float()
pick = np.random.randint(0, image.shape[0])
val_image_patches += (F.interpolate(
image[pick:pick + 1, :, :, :].cpu(),
size=(hidden_config.W, hidden_config.H)), )
val_encoded_patches += (F.interpolate(
encoded_images[pick:pick + 1, :, :, :].cpu(),
size=(hidden_config.W, hidden_config.H)), )
val_noised_patches += (F.interpolate(
noised_images[pick:pick + 1, :, :, :].cpu(),
size=(hidden_config.W, hidden_config.H)), )
if tb_logger is not None:
tb_logger.save_losses('val_loss', validation_losses, epoch)
tb_logger.writer.flush()
val_image_patches = torch.stack(val_image_patches).squeeze(1)
val_encoded_patches = torch.stack(val_encoded_patches).squeeze(1)
val_noised_patches = torch.stack(val_noised_patches).squeeze(1)
utils.save_images(val_image_patches[:images_to_save, :, :, :],
val_encoded_patches[:images_to_save, :, :, :],
val_noised_patches[:images_to_save, :, :, :],
epoch,
os.path.join(this_run_folder, 'images'),
resize_to=saved_images_size)
curr_cond = validation_losses['encoder_mse'].avg + validation_losses[
'bitwise-error'].avg
if best_cond is None or curr_cond < best_cond:
best_cond = curr_cond
best_epoch = epoch
utils.log_progress(validation_losses)
logging.info('-' * 40)
utils.save_checkpoint(model, train_options.experiment_name, epoch,
best_epoch, best_cond,
os.path.join(this_run_folder, 'checkpoints'))
logging.info(
f'Current best epoch = {best_epoch}, loss = {best_cond:.6f}')
utils.write_losses(os.path.join(this_run_folder, 'validation.csv'),
validation_losses, epoch,
time.time() - epoch_start)
def train(model: Hidden, device: torch.device,
hidden_config: HiDDenConfiguration, train_options: TrainingOptions,
this_run_folder: str, tb_logger):
"""
Trains the HiDDeN model
:param model: The model
:param device: torch.device object, usually this is GPU (if avaliable), otherwise CPU.
:param hidden_config: The network configuration
:param train_options: The training settings
:param this_run_folder: The parent folder for the current training run to store training artifacts/results/logs.
:param tb_logger: TensorBoardLogger object which is a thin wrapper for TensorboardX logger.
Pass None to disable TensorboardX logging
:return:
"""
train_data, val_data = utils.get_data_loaders(hidden_config, train_options)
file_count = len(train_data.dataset)
if file_count % train_options.batch_size == 0:
steps_in_epoch = file_count // train_options.batch_size
else:
steps_in_epoch = file_count // train_options.batch_size + 1
print_each = 10
images_to_save = 8
saved_images_size = (512, 512)
for epoch in range(train_options.start_epoch,
train_options.number_of_epochs + 1):
logging.info('\nStarting epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
logging.info('Batch size = {}\nSteps in epoch = {}'.format(
train_options.batch_size, steps_in_epoch))
training_losses = defaultdict(AverageMeter)
epoch_start = time.time()
step = 1
#train
for image, _ in train_data:
image = image.to(device)
"""
message = torch.Tensor(np.random.choice([0, 1], (image.shape[0], hidden_config.message_length))).to(device)
losses, _ = model.train_on_batch([image, message])
print(losses)
"""
#crop imgs
imgs = cropImg(32, image)
#iterate img
bitwise_arr = []
main_losses = None
encoded_imgs = []
for img in imgs:
img = img.to(device)
message = torch.Tensor(
np.random.choice(
[0, 1], (img.shape[0],
hidden_config.message_length))).to(device)
losses, (encoded_images, noised_images,
decoded_messages) = model.train_on_batch(
[img, message])
encoded_imgs.append(
encoded_images[0][0].cpu().detach().numpy())
main_losses = losses
for name, loss in losses.items():
if (name == 'bitwise-error '):
bitwise_arr.append(loss)
Total = 0
Vcount = 0
V_average = 0
H_average = 0
for i in range(0, len(encoded_imgs) - 1):
if ((i + 1) % 4 != 0):
img = encoded_imgs[i]
img_next = encoded_imgs[i + 1]
average_img = 0
average_img_next = 0
for j in range(0, 32):
for k in range(0, 10):
average_img = average_img + img[j][31 - k]
average_img_next = average_img_next + img_next[j][k]
average_blocking = np.abs(average_img -
average_img_next) / 320
V_average = V_average + average_blocking
for j in range(0, 32):
distinct = np.abs(img[j][31] - img_next[j][0])
Total = Total + 1
if (distinct > 0.5):
Vcount = Vcount + 1
V_average = V_average / 12
Hcount = 0
for i in range(0, len(encoded_imgs) - 4):
img = encoded_imgs[i]
img_next = encoded_imgs[i + 4]
average_img = 0
average_img_next = 0
for j in range(0, 32):
for k in range(0, 10):
average_img = average_img + img[31 - k][j]
average_img_next = average_img_next + img_next[k][j]
average_blocking = np.abs(average_img - average_img_next) / 320
H_average = H_average + average_blocking
for j in range(0, 32):
distinct = np.abs(img[31][j] - img_next[0][j])
Total = Total + 1
if (distinct > 0.5):
Hcount = Hcount + 1
H_average = H_average / 12
bitwise_arr = np.array(bitwise_arr)
bitwise_avg = np.average(bitwise_arr)
#blocking_loss = (Vcount+Hcount)/Total
blocking_loss = (H_average + V_average) / 2
for name, loss in main_losses.items():
if (name == 'bitwise-error '):
training_losses[name].update(bitwise_avg)
else:
if (name == 'blocking_effect'):
training_losses[name].update(blocking_loss)
else:
training_losses[name].update(loss)
if step % print_each == 0 or step == steps_in_epoch:
logging.info('Epoch: {}/{} Step: {}/{}'.format(
epoch, train_options.number_of_epochs, step,
steps_in_epoch))
utils.log_progress(training_losses)
logging.info('-' * 40)
step += 1
train_duration = time.time() - epoch_start
logging.info('Epoch {} training duration {:.2f} sec'.format(
epoch, train_duration))
logging.info('-' * 40)
utils.write_losses(os.path.join(this_run_folder, 'train.csv'),
training_losses, epoch, train_duration)
if tb_logger is not None:
tb_logger.save_losses(training_losses, epoch)
tb_logger.save_grads(epoch)
tb_logger.save_tensors(epoch)
first_iteration = True
validation_losses = defaultdict(AverageMeter)
logging.info('Running validation for epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
#val
for image, _ in val_data:
image = image.to(device)
#crop imgs
imgs = cropImg(32, image)
#iterate img
bitwise_arr = []
main_losses = None
encoded_imgs = []
blocking_imgs = []
for img in imgs:
img = img.to(device)
message = torch.Tensor(
np.random.choice(
[0, 1], (img.shape[0],
hidden_config.message_length))).to(device)
losses, (encoded_images, noised_images,
decoded_messages) = model.validate_on_batch(
[img, message])
encoded_imgs.append(encoded_images)
blocking_imgs.append(
encoded_images[0][0].cpu().detach().numpy())
main_losses = losses
for name, loss in losses.items():
if (name == 'bitwise-error '):
bitwise_arr.append(loss)
Total = 0
Vcount = 0
V_average = 0
H_average = 0
for i in range(0, len(blocking_imgs) - 1):
if ((i + 1) % 4 != 0):
img = blocking_imgs[i]
img_next = blocking_imgs[i + 1]
average_img = 0
average_img_next = 0
for j in range(0, 32):
for k in range(0, 10):
average_img = average_img + img[j][31 - k]
average_img_next = average_img_next + img_next[j][k]
average_blocking = np.abs(average_img -
average_img_next) / 320
V_average = V_average + average_blocking
for j in range(0, 32):
distinct = np.abs(img[j][31] - img_next[j][0])
Total = Total + 1
if (distinct > 0.5):
Vcount = Vcount + 1
V_average = V_average / 12
Hcount = 0
for i in range(0, len(blocking_imgs) - 4):
img = blocking_imgs[i]
img_next = blocking_imgs[i + 4]
for j in range(0, 32):
for k in range(0, 10):
average_img = average_img + img[31 - k][j]
average_img_next = average_img_next + img_next[k][j]
average_blocking = np.abs(average_img - average_img_next) / 320
H_average = H_average + average_blocking
for j in range(0, 32):
distinct = np.abs(img[31][j] - img_next[0][j])
Total = Total + 1
if (distinct > 0.5):
Hcount = Hcount + 1
H_average = H_average / 12
bitwise_arr = np.array(bitwise_arr)
bitwise_avg = np.average(bitwise_arr)
#blocking_loss = (Vcount+Hcount)/Total
blocking_loss = (H_average + V_average) / 2
for name, loss in main_losses.items():
if (name == 'bitwise-error '):
validation_losses[name].update(bitwise_avg)
else:
if (name == 'blocking_effect'):
validation_losses[name].update(blocking_loss)
else:
validation_losses[name].update(loss)
#concat image
encoded_images = concatImgs(encoded_imgs)
if first_iteration:
if hidden_config.enable_fp16:
image = image.float()
encoded_images = encoded_images.float()
utils.save_images(
image.cpu()[:images_to_save, :, :, :],
encoded_images[:images_to_save, :, :, :].cpu(),
epoch,
os.path.join(this_run_folder, 'images'),
resize_to=saved_images_size)
first_iteration = False
utils.log_progress(validation_losses)
logging.info('-' * 40)
utils.save_checkpoint(model, train_options.experiment_name, epoch,
os.path.join(this_run_folder, 'checkpoints'))
utils.write_losses(os.path.join(this_run_folder, 'validation.csv'),
validation_losses, epoch,
time.time() - epoch_start)
def train(model: Hidden, device: torch.device,
hidden_config: HiDDenConfiguration, train_options: TrainingOptions,
this_run_folder: str, tb_logger, vocab):
"""
Trains the HiDDeN model
:param model: The model
:param device: torch.device object, usually this is GPU (if avaliable), otherwise CPU.
:param hidden_config: The network configuration
:param train_options: The training settings
:param this_run_folder: The parent folder for the current training run to store training artifacts/results/logs.
:param tb_logger: TensorBoardLogger object which is a thin wrapper for TensorboardX logger.
Pass None to disable TensorboardX logging
:return:
"""
train_data, val_data = utils.get_data_loaders(hidden_config, train_options,
vocab)
file_count = len(train_data.dataset)
if file_count % train_options.batch_size == 0:
steps_in_epoch = file_count // train_options.batch_size
else:
steps_in_epoch = file_count // train_options.batch_size + 1
print_each = 10
images_to_save = 8
saved_images_size = (512, 512)
for epoch in range(train_options.start_epoch,
train_options.number_of_epochs + 1):
logging.info('\nStarting epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
logging.info('Batch size = {}\nSteps in epoch = {}'.format(
train_options.batch_size, steps_in_epoch))
training_losses = defaultdict(AverageMeter)
epoch_start = time.time()
step = 1
for image, ekeys, dkeys, caption, length in train_data:
image, caption, ekeys, dkeys = image.to(device), caption.to(
device), ekeys.to(device), dkeys.to(device)
losses, _ = model.train_on_batch(
[image, ekeys, dkeys, caption, length])
for name, loss in losses.items():
training_losses[name].update(loss)
if step % print_each == 0 or step == steps_in_epoch:
logging.info('Epoch: {}/{} Step: {}/{}'.format(
epoch, train_options.number_of_epochs, step,
steps_in_epoch))
utils.log_progress(training_losses)
logging.info('-' * 40)
step += 1
train_duration = time.time() - epoch_start
logging.info('Epoch {} training duration {:.2f} sec'.format(
epoch, train_duration))
logging.info('-' * 40)
utils.write_losses(os.path.join(this_run_folder, 'train.csv'),
training_losses, epoch, train_duration)
if tb_logger is not None:
tb_logger.save_losses(training_losses, epoch)
tb_logger.save_grads(epoch)
tb_logger.save_tensors(epoch)
first_iteration = True
validation_losses = defaultdict(AverageMeter)
logging.info('Running validation for epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
for image, ekeys, dkeys, caption, length in val_data:
image, caption, ekeys, dkeys = image.to(device), caption.to(
device), ekeys.to(device), dkeys.to(device)
losses, (encoded_images, noised_images, decoded_messages, predicted_sents) = \
model.validate_on_batch([image, ekeys, dkeys, caption, length])
#print(predicted)
#exit()
predicted_sents = predicted_sents.cpu().numpy()
for i in range(train_options.batch_size):
try:
#print(''.join([vocab.idx2word[int(w)] + ' ' for w in predicted.cpu().numpy()[i::train_options.batch_size]][1:length[i]-1]))
print("".join([
vocab.idx2word[int(idx)] + ' '
for idx in predicted_sents[i]
]))
break
except IndexError:
print(f'{i}th batch does not have enough length.')
for name, loss in losses.items():
validation_losses[name].update(loss)
if first_iteration:
if hidden_config.enable_fp16:
image = image.float()
encoded_images = encoded_images.float()
utils.save_images(
image.cpu()[:images_to_save, :, :, :],
encoded_images[:images_to_save, :, :, :].cpu(),
epoch,
os.path.join(this_run_folder, 'images'),
resize_to=saved_images_size)
first_iteration = False
utils.log_progress(validation_losses)
logging.info('-' * 40)
utils.save_checkpoint(model, train_options.experiment_name, epoch,
os.path.join(this_run_folder, 'checkpoints'))
utils.write_losses(os.path.join(this_run_folder, 'validation.csv'),
validation_losses, epoch,
time.time() - epoch_start)
def train(model: Hidden, device: torch.device,
hidden_config: HiDDenConfiguration, train_options: TrainingOptions,
this_run_folder: str, tb_logger):
"""
Trains the HiDDeN model
:param model: The model
:param device: torch.device object, usually this is GPU (if avaliable), otherwise CPU.
:param hidden_config: The network configuration
:param train_options: The training settings
:param this_run_folder: The parent folder for the current training run to store training artifacts/results/logs.
:param tb_logger: TensorBoardLogger object which is a thin wrapper for TensorboardX logger.
Pass None to disable TensorboardX logging
:return:
"""
train_data, val_data = utils.get_data_loaders(hidden_config, train_options)
file_count = len(train_data.dataset)
if file_count % train_options.batch_size == 0:
steps_in_epoch = file_count // train_options.batch_size
else:
steps_in_epoch = file_count // train_options.batch_size + 1
steps_in_epoch = 313
print_each = 10
images_to_save = 8
saved_images_size = (512, 512)
for epoch in range(train_options.start_epoch,
train_options.number_of_epochs + 1):
logging.info('\nStarting epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
logging.info('Batch size = {}\nSteps in epoch = {}'.format(
train_options.batch_size, steps_in_epoch))
training_losses = defaultdict(AverageMeter)
epoch_start = time.time()
step = 1
for image, _ in train_data:
image = image.to(device)
message = torch.Tensor(
np.random.choice(
[0, 1],
(image.shape[0], hidden_config.message_length))).to(device)
losses, _ = model.train_on_batch([image, message])
for name, loss in losses.items():
training_losses[name].update(loss)
if step % print_each == 0 or step == steps_in_epoch:
#import pdb; pdb.set_trace()
logging.info('Epoch: {}/{} Step: {}/{}'.format(
epoch, train_options.number_of_epochs, step,
steps_in_epoch))
utils.log_progress(training_losses)
logging.info('-' * 40)
step += 1
if step == steps_in_epoch:
break
train_duration = time.time() - epoch_start
logging.info('Epoch {} training duration {:.2f} sec'.format(
epoch, train_duration))
logging.info('-' * 40)
utils.write_losses(os.path.join(this_run_folder, 'train.csv'),
training_losses, epoch, train_duration)
if tb_logger is not None:
tb_logger.save_losses(training_losses, epoch)
tb_logger.save_grads(epoch)
tb_logger.save_tensors(epoch)
first_iteration = True
validation_losses = defaultdict(AverageMeter)
logging.info('Running validation for epoch {}/{}'.format(
epoch, train_options.number_of_epochs))
step = 1
for image, _ in val_data:
image = image.to(device)
message = torch.Tensor(
np.random.choice(
[0, 1],
(image.shape[0], hidden_config.message_length))).to(device)
losses, (encoded_images, noised_images,
decoded_messages) = model.validate_on_batch(
[image, message])
for name, loss in losses.items():
validation_losses[name].update(loss)
if first_iteration:
if hidden_config.enable_fp16:
image = image.float()
encoded_images = encoded_images.float()
utils.save_images(
image.cpu()[:images_to_save, :, :, :],
encoded_images[:images_to_save, :, :, :].cpu(),
epoch,
os.path.join(this_run_folder, 'images'),
resize_to=saved_images_size)
first_iteration = False
step += 1
if step == steps_in_epoch // 10:
break
utils.log_progress(validation_losses)
logging.info('-' * 40)
utils.save_checkpoint(model, train_options.experiment_name, epoch,
os.path.join(this_run_folder, 'checkpoints'))
utils.write_losses(os.path.join(this_run_folder, 'validation.csv'),
validation_losses, epoch,
time.time() - epoch_start)
def main():
# device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
device = torch.device('cpu')
parser = argparse.ArgumentParser(description='Training of HiDDeN nets')
# parser.add_argument('--size', '-s', default=128, type=int, help='The size of the images (images are square so this is height and width).')
parser.add_argument('--data-dir',
'-d',
required=True,
type=str,
help='The directory where the data is stored.')
parser.add_argument(
'--runs_root',
'-r',
default=os.path.join('.', 'experiments'),
type=str,
help='The root folder where data about experiments are stored.')
args = parser.parse_args()
print_each = 25
completed_runs = [
o for o in os.listdir(args.runs_root)
if os.path.isdir(os.path.join(args.runs_root, o))
and o != 'no-noise-defaults'
]
print(completed_runs)
write_csv_header = True
for run_name in completed_runs:
current_run = os.path.join(args.runs_root, run_name)
print(f'Run folder: {current_run}')
options_file = os.path.join(current_run, 'options-and-config.pickle')
train_options, hidden_config, noise_config = utils.load_options(
options_file)
train_options.train_folder = os.path.join(args.data_dir, 'val')
train_options.validation_folder = os.path.join(args.data_dir, 'val')
train_options.batch_size = 4
checkpoint = utils.load_last_checkpoint(
os.path.join(current_run, 'checkpoints'))
noiser = Noiser(noise_config, device)
model = Hidden(hidden_config, device, noiser, tb_logger=None)
utils.model_from_checkpoint(model, checkpoint)
print('Model loaded successfully. Starting validation run...')
_, val_data = utils.get_data_loaders(hidden_config, train_options)
file_count = len(val_data.dataset)
if file_count % train_options.batch_size == 0:
steps_in_epoch = file_count // train_options.batch_size
else:
steps_in_epoch = file_count // train_options.batch_size + 1
losses_accu = {}
step = 0
for image, _ in val_data:
step += 1
image = image.to(device)
message = torch.Tensor(
np.random.choice(
[0, 1],
(image.shape[0], hidden_config.message_length))).to(device)
losses, (encoded_images, noised_images,
decoded_messages) = model.validate_on_batch(
[image, message])
if not losses_accu: # dict is empty, initialize
for name in losses:
losses_accu[name] = []
for name, loss in losses.items():
losses_accu[name].append(loss)
if step % print_each == 0:
print(f'Step {step}/{steps_in_epoch}')
utils.print_progress(losses_accu)
print('-' * 40)
utils.print_progress(losses_accu)
write_validation_loss(os.path.join(args.runs_root,
'validation_run.csv'),
losses_accu,
run_name,
checkpoint['epoch'],
write_header=write_csv_header)
write_csv_header = False