def train(model,
criterion,
optimizer,
epochs,
device,
train_loader,
valid_loader,
print_every=60,
step_track_every=30):
total_start = time()
model.epochs = epochs
save_checkpoint(model, optimizer)
train_losses, valid_losses = [], []
steps = 0
running_loss = 0
valid_loader_size = len(valid_loader)
print("==========================")
print("Starting training of NN...")
for epoch in range(epochs):
print("==========================")
print(f"Starting epoch #{epoch}...")
start_epoch = start_step = time()
for inputs, labels in train_loader:
steps += 1
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
logps = model.forward(inputs)
loss = criterion(logps, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
if steps % step_track_every == 0:
print(
f"Time in step #{steps}: {(time() - start_step):.3f} seconds"
)
start_step = time()
if steps % print_every == 0:
start_test = time()
valid_loss, accuracy = get_valid_loss_and_accuracy(
model, valid_loader, device, criterion)
cur_train_loss = running_loss / valid_loader_size
running_loss = 0
train_losses.append(cur_train_loss)
valid_losses.append(valid_loss)
print(f"Epoch {epoch}/{epochs}..."
f"Training loss: {cur_train_loss:.4f}..."
f"Validation loss: {valid_loss:.4f}..."
f"Test accuracy: {accuracy:.4f}\n")
print(
f"Time taken to test losses in epoch #{epoch}: {(time() - start_test):.3f} seconds"
)
print(f"Time per epoch: {(time() - start_epoch):.3f} seconds")
print(f"Total training time: {(time() - total_start):.3f} seconds")
save_checkpoint(model, optimizer)
def do_run():
config = get_input_config()
#setup some globals
global STANZA
STANZA = config.get("name")
http_proxy = config.get("http_proxy")
https_proxy = config.get("https_proxy")
proxies = {}
if not http_proxy is None:
proxies["http"] = http_proxy
if not https_proxy is None:
proxies["https"] = https_proxy
request_timeout = int(config.get("request_timeout", 30))
try:
req_args = {"verify": True, "timeout": float(request_timeout)}
if proxies:
req_args["proxies"] = proxies
req = requests.get(
url=
"https://publicdashacc.blob.core.windows.net/publicdata?restype=container&comp=list&prefix=data",
params=req_args)
xmldom = etree.fromstring(req.content)
blobs = xmldom.xpath('/EnumerationResults/Blobs/Blob')
for blob in blobs:
blob_etag = blob.xpath('Properties/Etag')[0].text
blob_name = blob.xpath('Name')[0].text
logging.info("Found file=%s etag=%s" % (blob_name, blob_etag))
blob_url = "https://publicdashacc.blob.core.windows.net/publicdata/%s" % (
blob_name)
if not load_checkpoint(config, blob_etag):
print("Processing file={}".format(blob_url))
data_req = requests.get(url=blob_url, params=req_args)
data_json = data_req.json()
iterate_json_data("overview", data_json, blob_name)
iterate_json_data("countries", data_json, blob_name)
iterate_json_data("regions", data_json, blob_name)
iterate_json_data("utlas", data_json, blob_name)
logging.info("Marking file={} etag={} as processed".format(
blob_name, blob_etag))
save_checkpoint(config, blob_etag)
except RuntimeError, e:
logging.error("Looks like an error: %s" % str(e))
sys.exit(2)
def main():
in_args = train_input_args()
train_dir = in_args.data_dir + '/train'
valid_dir = in_args.data_dir + '/valid'
train_transforms = transforms.Compose([
transforms.RandomRotation(30),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
valid_transforms = transforms.Compose([
transforms.Resize(255),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
train_data = datasets.ImageFolder(train_dir, transform=train_transforms)
valid_data = datasets.ImageFolder(valid_dir, transform=valid_transforms)
trainloader = torch.utils.data.DataLoader(train_data,
batch_size=64,
shuffle=True)
validloader = torch.utils.data.DataLoader(valid_data, batch_size=64)
dataloader = [trainloader, validloader]
if in_args.arch == 'densenet121':
model = models.densenet121(pretrained=True)
classifier = nn.Sequential(
OrderedDict([("fc1", nn.Linear(1024, in_args.hidden_units)),
("ReLU1", nn.ReLU()), ("dropout", nn.Dropout(0.5)),
("fc2", nn.Linear(in_args.hidden_units, 102)),
("output", nn.LogSoftmax(dim=1))]))
elif in_args.arch == 'vgg16':
model = models.vgg16(pretrained=True)
classifier = nn.Sequential(
OrderedDict([("fc1", nn.Linear(25088, in_args.hidden_units)),
("ReLU1", nn.ReLU()), ("dropout", nn.Dropout(0.5)),
("fc2", nn.Linear(in_args.hidden_units, 102)),
("output", nn.LogSoftmax(dim=1))]))
# Freeze parameters so we don't backprop through them
for param in model.parameters():
param.requires_grad = False
model.classifier = classifier
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.classifier.parameters(),
lr=in_args.learning_rate)
device = torch.device('cuda' if in_args.gpu == 'gpu' else 'cpu')
train(model, in_args, dataloader, optimizer, criterion, device)
model.class_to_idx = train_data.class_to_idx
save_checkpoint(in_args.save_dir, optimizer, in_args, classifier, model)
def main(train_set, learning_rate, n_epochs, batch_size, num_workers, hidden_size, model_file,
cuda, checkpoint_interval, seed, n_disc):
# make data between -1 and 1
data_transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))])
train_dataset = datasets.ImageFolder(root=os.path.join(os.getcwd(), train_set),
transform=data_transform)
train_dataloader = DataLoader(train_dataset, batch_size=batch_size,
shuffle=True, num_workers=num_workers,
drop_last=True)
# initialize model
if model_file:
try:
total_examples, fixed_noise, gen_losses, disc_losses, gen_loss_per_epoch, \
disc_loss_per_epoch, prev_epoch, gen, disc = load_model_wgan(model_file, hidden_size) # TODO: upsampling method?
print('model loaded successfully!')
except:
print('could not load model! creating new model...')
model_file = None
if not model_file:
print('creating new model...')
gen = Generator(hidden_dim=hidden_size, leaky=0.2)
disc = Discriminator(leaky=0.2)
gen.weight_init(mean=0, std=0.02)
disc.weight_init(mean=0, std=0.02)
total_examples = 0
disc_losses = []
gen_losses = []
disc_loss_per_epoch = []
gen_loss_per_epoch = []
prev_epoch = 0
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if cuda:
fixed_noise = Variable(torch.randn(9, hidden_size).cuda())
else:
fixed_noise = Variable(torch.rand(9, hidden_size))
if cuda:
gen.cuda()
disc.cuda()
# Adam optimizer
gen_optimizer = optim.RMSprop(gen.parameters(), lr=learning_rate, eps=1e-8)
disc_optimizer = optim.RMSprop(disc.parameters(), lr=learning_rate, eps=1e-8)
# results save folder
gen_images_dir = 'results/wgan_generated_images'
train_summaries_dir = 'results/wgan_training_summaries'
checkpoint_dir = 'results/wgan_checkpoints'
if not os.path.isdir('results'):
os.mkdir('results')
if not os.path.isdir(gen_images_dir):
os.mkdir(gen_images_dir)
if not os.path.isdir(train_summaries_dir):
os.mkdir(train_summaries_dir)
if not os.path.isdir(checkpoint_dir):
os.mkdir(checkpoint_dir)
np.random.seed(seed) # reset training seed to ensure that batches remain the same between runs!
try:
for epoch in range(prev_epoch, n_epochs):
disc_losses_epoch = []
gen_losses_epoch = []
for idx, (true_batch, _) in enumerate(train_dataloader):
disc.zero_grad()
# Sample minibatch of examples from data generating distribution
if cuda:
true_batch = Variable(true_batch.cuda())
else:
true_batch = Variable(true_batch)
# discriminator on true data
true_disc_output = disc.forward(true_batch)
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if cuda:
z = Variable(torch.randn(batch_size, hidden_size).cuda())
else:
z = Variable(torch.randn(batch_size, hidden_size))
# discriminator on fake data
fake_batch = gen.forward(z.view(-1, hidden_size, 1, 1))
fake_disc_output = disc.forward(
fake_batch.detach()) # detach so gradients not computed for generator
# Optimize with new loss function
disc_loss = wgan_Dloss(true_disc_output, fake_disc_output)
disc_loss.backward()
disc_optimizer.step()
# Weight clipping as done by WGAN
for p in disc.parameters():
p.data.clamp_(-0.01, 0.01)
# Store losses
disc_losses_epoch.append(disc_loss.data[0])
# Train generator after the discriminator has been trained n_disc times
if (idx+1) % n_disc == 0:
gen.zero_grad()
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if cuda:
z = Variable(torch.randn(batch_size, hidden_size).cuda())
else:
z = Variable(torch.rand(batch_size, hidden_size))
# train generator
fake_batch = gen.forward(z.view(-1, hidden_size, 1, 1))
fake_disc_output = disc.forward(fake_batch)
# Optimize generator
gen_loss = wgan_Gloss(fake_disc_output)
gen_loss.backward()
gen_optimizer.step()
# Store losses
gen_losses_epoch.append(gen_loss.data[0])
if (total_examples != 0) and (total_examples % n_disc*4 == 0):
print('epoch {}: step {}/{} disc loss: {:.4f}, gen loss: {:.4f}'
.format(epoch + 1, idx + 1, len(train_dataloader), disc_loss.data[0], gen_loss.data[0]))
# Checkpoint model
total_examples += batch_size
if (checkpoint_interval != 0) and (total_examples % checkpoint_interval == 0):
disc_losses.extend(disc_losses_epoch)
gen_losses.extend(gen_losses_epoch)
save_checkpoint(total_examples=total_examples, fixed_noise=fixed_noise, disc=disc, gen=gen,
gen_losses=gen_losses, disc_losses=disc_losses,
disc_loss_per_epoch=disc_loss_per_epoch,
gen_loss_per_epoch=gen_loss_per_epoch, epoch=epoch, directory=checkpoint_dir)
print("Checkpoint saved!")
# sample images for inspection
save_image_sample(batch=gen.forward(fixed_noise.view(-1, hidden_size, 1, 1)),
cuda=cuda, total_examples=total_examples, directory=gen_images_dir)
print("Saved images!")
# save learning curves for inspection
save_learning_curve(gen_losses=gen_losses, disc_losses=disc_losses,
total_examples=total_examples, directory=train_summaries_dir)
print("Saved learning curves!")
disc_loss_per_epoch.append(np.average(disc_losses_epoch))
gen_loss_per_epoch.append(np.average(gen_losses_epoch))
# Save epoch learning curve
save_learning_curve_epoch(gen_losses=gen_loss_per_epoch, disc_losses=disc_loss_per_epoch,
total_epochs=epoch + 1, directory=train_summaries_dir)
print("Saved learning curves!")
print('epoch {}/{} disc loss: {:.4f}, gen loss: {:.4f}'
.format(epoch + 1, n_epochs, np.array(disc_losses_epoch).mean(), np.array(gen_losses_epoch).mean()))
disc_losses.extend(disc_losses_epoch)
gen_losses.extend(gen_losses_epoch)
except KeyboardInterrupt:
print("Saving before quit...")
save_checkpoint(total_examples=total_examples, fixed_noise=fixed_noise, disc=disc, gen=gen,
disc_loss_per_epoch=disc_loss_per_epoch,
gen_loss_per_epoch=gen_loss_per_epoch,
gen_losses=gen_losses, disc_losses=disc_losses, epoch=epoch, directory=checkpoint_dir)
print("Checkpoint saved!")
# sample images for inspection
save_image_sample(batch=gen.forward(fixed_noise.view(-1, hidden_size, 1, 1)),
cuda=cuda, total_examples=total_examples, directory=gen_images_dir)
print("Saved images!")
# save learning curves for inspection
save_learning_curve(gen_losses=gen_losses, disc_losses=disc_losses,
total_examples=total_examples, directory=train_summaries_dir)
print("Saved learning curves!")
def train_and_eval(model,
train_loader,
valid_loader,
learning_rate,
epochs,
model_outdir,
wts,
task,
metrics_every_iter,
restore_chkpt=None,
run_suffix=None):
"""
Contains the powerhouse of the network, ie. the training and validation iterations called through run_model().
All parameters from the command line/json are parsed and then passed into run_model().
Performs checkpointing each epoch, saved as 'last.pth.tar' and the best model thus far (based on validation AUC), saved as 'best.pth.tar'
:param model: (nn.Module)
:param train_loader: (torch DataLoader)
:param valid_loader: (torch DataLoader)
:param learning_rate: (float) - the learning rate, defaults to 1e-05
:param epochs: (int) - the number of epochs
:param wts: (tensor) - class weights
:param model_outdir: (str) - the output directory for checkpointing, checkpoints will be saved as output_dir/task/view/*.tar
:param restore_chkpt: (str) - the directory to reload the checkpoint, if specified
:param run_suffix: (str) - suffix to be appended to the event file
:return:
"""
# output/task/my_run
# goes back 3 levels up, putting the name in the same level as the output. two levels up would put them into output/
recover_root_dir = os.path.dirname(
os.path.dirname(os.path.dirname(
model_outdir))) # removes the task and view from the directory
log_dir = os.path.join(recover_root_dir, "logs")
run_name = re.split(r'/|\\', model_outdir)[-1]
# task = re.split(r'/|\\', model_outdir)[-2]
# have log folder naming structure same as models
log_fn = os.path.join(log_dir, task, run_name)
dtnow = datetime.now()
# dtnow.strftime("%Y%m%d_%H%M%S")
log_fn = os.path.join(log_fn, dtnow.strftime("%Y_%m_%d-%H_%M_%S"))
# make directory if it doesn't exist.
if not os.path.exists(log_fn):
os.makedirs(log_fn)
print('{} does not exist, creating..!'.format(log_fn))
else:
print('{} already exists!'.format(log_fn))
# each tensorboard event file should ideally be saved to a unique folder, else the resulting graph will look like
# it's time traveling because of overlapping logs
# if run_suffix:
# writer = tf.summary.create_file_writer(log_fn, filename_suffix=run_suffix)
# else:
# writer = tf.summary.create_file_writer(log_fn)
# use cpu or cuda depending on availability
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
current_best_val_loss = float('Inf')
# this needs to be outside of the loop else it'll keep resetting, right? same with the model
optimizer = torch.optim.Adam(model.parameters(),
learning_rate,
weight_decay=0.01)
# taken directly from MRNet code
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
patience=5, # how many epochs to wait for before acting
factor=0.3, # factor to reduce LR by, LR = factor * LR
threshold=1e-4) # threshold to measure new optimum
# weight loss by training class positive weights, if use_wts is False then no weights are applied
# criterion_d = {'bladder': torch.nn.BCEWithLogitsLoss(), 'view': torch.nn.CrossEntropyLoss(), 'granular':torch.nn.CrossEntropyLoss()}
if wts is None:
criterion = torch.nn.CrossEntropyLoss()
else:
wts = wts.to(device)
criterion = torch.nn.CrossEntropyLoss(weight=wts)
# # TODO: reloading checkpoint
# if restore_chkpt:
# logging.info("Restoring Checkpoint from {}".format(restore_chkpt))
# helpers.load_checkpoint(checkpoint = restore_chkpt,
# model = model,
# optimizer = optimizer,
# # scheduler = scheduler,
# epochs = epochs)
# # so epochs - loaded_epoch is where we would need to start, right?
# logging.info("Starting again at Epoch {}....".format(epochs))
# logging.info("Finished Restoring Checkpoint...")
for epoch in range(epochs):
logging.info('[Epoch {}]'.format(epoch + 1))
# main training loop
epoch_loss, epoch_preds, epoch_labels = run_model(
model=model,
loader=train_loader,
optimizer=optimizer,
criterion=criterion,
metrics_every_iter=metrics_every_iter,
train=True)
logging.info('[Epoch {}]\t\t Training Average Loss: {:.5f}'.format(
epoch + 1, epoch_loss))
# logging.info('[Epoch {}]\t\tTraining Balanced Accuracy: {:.3f}\t Training Average Loss: {:.5f}'.format(epoch + 1, epoch_auc, epoch_loss))
# main validation loop
epoch_val_loss, epoch_val_preds, epoch_val_labels = run_model(
model=model,
loader=valid_loader,
optimizer=optimizer,
criterion=criterion,
metrics_every_iter=
False, # default, just show the epoch validation metrics..
train=False)
logging.info('[Epoch {}]\t\t Validation Average Loss: {:.5f}'.format(
epoch + 1, epoch_val_loss))
# logging.info('[Epoch {}]\t\tValidation Balanced Accuracy: {:.3f}\t Validation Average Loss: {:.5f}'.format(epoch + 1, epoch_val_acc, epoch_val_loss))
scheduler.step(epoch_val_loss
) # check per epoch, how does the threshold work?!?!?
logging.info('[Epoch {}]\t\tOptimizer Learning Rate: {}'.format(
epoch + 1, {optimizer.param_groups[0]['lr']}))
# with writer.as_default():
# tf.summary.scalar('Loss/train', epoch_loss, epoch + 1)
# tf.summary.scalar('Loss/val', epoch_val_loss, epoch + 1)
# tf.summary.scalar('BACC/train', epoch_acc, epoch + 1)
# tf.summary.scalar('BACC/val', epoch_val_acc, epoch + 1)
# check whether the most recent epoch loss is better than previous best
# is_best_val_auc = epoch_val_auc >= current_best_val_auc
is_best_val_loss = epoch_val_loss < current_best_val_loss
# save state in a dictionary
state = {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
# 'validation_acc': epoch_val_acc,
'best_validation_loss': epoch_val_loss,
# 'metrics': metrics # read more into this
'scheduler_dict': scheduler.state_dict(),
'optim_dict': optimizer.state_dict()
}
# save as last epoch
helpers.save_checkpoint(state,
is_best=is_best_val_loss,
checkpoint_dir=model_outdir)
# epoch = epoch + 1)
if is_best_val_loss:
# set new best validation loss
# current_best_val_auc = epoch_val_auc
current_best_val_loss = epoch_val_loss
# logging.info('[Epoch {}]\t\t******New Best Validation:\t AUC: {:.3f}******'.format(epoch + 1, epoch_val_auc))
logging.info(
'[Epoch {}]\t\t******New Best Validation Loss: {:.3f}******'.
format(epoch + 1, epoch_val_loss))
helpers.save_checkpoint(state,
is_best=is_best_val_loss,
checkpoint_dir=model_outdir)
def main(train_set, learning_rate, n_epochs, beta_0, beta_1, batch_size,
num_workers, hidden_size, model_file, cuda, display_result_every,
checkpoint_interval, seed, label_smoothing, grad_clip, dropout,
upsampling):
# make data between -1 and 1
data_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])
train_dataset = datasets.ImageFolder(root=os.path.join(
os.getcwd(), train_set),
transform=data_transform)
train_dataloader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers,
drop_last=True)
# initialize model
if model_file:
try:
total_examples, fixed_noise, gen_losses, disc_losses, gen_loss_per_epoch, \
disc_loss_per_epoch, prev_epoch, gen, disc = load_model(model_file, hidden_size, upsampling, cuda)
print('model loaded successfully!')
except:
print('could not load model! creating new model...')
model_file = None
if not model_file:
print('creating new model...')
if upsampling == 'transpose':
from models.model import Generator, Discriminator
elif upsampling == 'nn':
from models.model_nn import Generator, Discriminator
elif upsampling == 'bilinear':
from models.model_bilinear import Generator, Discriminator
gen = Generator(hidden_dim=hidden_size, leaky=0.2, dropout=dropout)
disc = Discriminator(leaky=0.2, dropout=dropout)
gen.weight_init(mean=0, std=0.02)
disc.weight_init(mean=0, std=0.02)
total_examples = 0
disc_losses = []
gen_losses = []
disc_loss_per_epoch = []
gen_loss_per_epoch = []
prev_epoch = 0
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if cuda:
fixed_noise = Variable(torch.randn(9, hidden_size).cuda())
else:
fixed_noise = Variable(torch.rand(9, hidden_size))
if cuda:
gen.cuda()
disc.cuda()
# Binary Cross Entropy loss
BCE_loss = nn.BCELoss()
# Adam optimizer
gen_optimizer = optim.Adam(gen.parameters(),
lr=learning_rate,
betas=(beta_0, beta_1),
eps=1e-8)
disc_optimizer = optim.Adam(disc.parameters(),
lr=learning_rate,
betas=(beta_0, beta_1),
eps=1e-8)
# results save folder
gen_images_dir = 'results/generated_images'
train_summaries_dir = 'results/training_summaries'
checkpoint_dir = 'results/checkpoints'
if not os.path.isdir('results'):
os.mkdir('results')
if not os.path.isdir(gen_images_dir):
os.mkdir(gen_images_dir)
if not os.path.isdir(train_summaries_dir):
os.mkdir(train_summaries_dir)
if not os.path.isdir(checkpoint_dir):
os.mkdir(checkpoint_dir)
np.random.seed(
seed
) # reset training seed to ensure that batches remain the same between runs!
try:
for epoch in range(prev_epoch, n_epochs):
disc_losses_epoch = []
gen_losses_epoch = []
for idx, (true_batch, _) in enumerate(train_dataloader):
disc.zero_grad()
# hack 6 of https://github.com/soumith/ganhacks
if label_smoothing:
true_target = torch.FloatTensor(batch_size).uniform_(
0.7, 1.2)
else:
true_target = torch.ones(batch_size)
# Sample minibatch of examples from data generating distribution
if cuda:
true_batch = Variable(true_batch.cuda())
true_target = Variable(true_target.cuda())
else:
true_batch = Variable(true_batch)
true_target = Variable(true_target)
# train discriminator on true data
true_disc_result = disc.forward(true_batch)
disc_train_loss_true = BCE_loss(true_disc_result.squeeze(),
true_target)
disc_train_loss_true.backward()
torch.nn.utils.clip_grad_norm(disc.parameters(), grad_clip)
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if label_smoothing:
fake_target = torch.FloatTensor(batch_size).uniform_(
0, 0.3)
else:
fake_target = torch.zeros(batch_size)
if cuda:
z = Variable(torch.randn(batch_size, hidden_size).cuda())
fake_target = Variable(fake_target.cuda())
else:
z = Variable(torch.randn(batch_size, hidden_size))
fake_target = Variable(fake_target)
# train discriminator on fake data
fake_batch = gen.forward(z.view(-1, hidden_size, 1, 1))
fake_disc_result = disc.forward(fake_batch.detach(
)) # detach so gradients not computed for generator
disc_train_loss_false = BCE_loss(fake_disc_result.squeeze(),
fake_target)
disc_train_loss_false.backward()
torch.nn.utils.clip_grad_norm(disc.parameters(), grad_clip)
disc_optimizer.step()
# compute performance statistics
disc_train_loss = disc_train_loss_true + disc_train_loss_false
disc_losses_epoch.append(disc_train_loss.data[0])
disc_fake_accuracy = 1 - fake_disc_result.mean().data[0]
disc_true_accuracy = true_disc_result.mean().data[0]
# Sample minibatch of m noise samples from noise prior p_g(z) and transform
if label_smoothing:
true_target = torch.FloatTensor(batch_size).uniform_(
0.7, 1.2)
else:
true_target = torch.ones(batch_size)
if cuda:
z = Variable(torch.randn(batch_size, hidden_size).cuda())
true_target = Variable(true_target.cuda())
else:
z = Variable(torch.rand(batch_size, hidden_size))
true_target = Variable(true_target)
# train generator
gen.zero_grad()
fake_batch = gen.forward(z.view(-1, hidden_size, 1, 1))
disc_result = disc.forward(fake_batch)
gen_train_loss = BCE_loss(disc_result.squeeze(), true_target)
gen_train_loss.backward()
torch.nn.utils.clip_grad_norm(gen.parameters(), grad_clip)
gen_optimizer.step()
gen_losses_epoch.append(gen_train_loss.data[0])
if (total_examples != 0) and (total_examples %
display_result_every == 0):
print(
'epoch {}: step {}/{} disc true acc: {:.4f} disc fake acc: {:.4f} '
'disc loss: {:.4f}, gen loss: {:.4f}'.format(
epoch + 1, idx + 1, len(train_dataloader),
disc_true_accuracy, disc_fake_accuracy,
disc_train_loss.data[0], gen_train_loss.data[0]))
# Checkpoint model
total_examples += batch_size
if (total_examples != 0) and (total_examples %
checkpoint_interval == 0):
disc_losses.extend(disc_losses_epoch)
gen_losses.extend(gen_losses_epoch)
save_checkpoint(total_examples=total_examples,
fixed_noise=fixed_noise,
disc=disc,
gen=gen,
gen_losses=gen_losses,
disc_losses=disc_losses,
disc_loss_per_epoch=disc_loss_per_epoch,
gen_loss_per_epoch=gen_loss_per_epoch,
epoch=epoch,
directory=checkpoint_dir)
print("Checkpoint saved!")
# sample images for inspection
save_image_sample(batch=gen.forward(
fixed_noise.view(-1, hidden_size, 1, 1)),
cuda=cuda,
total_examples=total_examples,
directory=gen_images_dir)
print("Saved images!")
# save learning curves for inspection
save_learning_curve(gen_losses=gen_losses,
disc_losses=disc_losses,
total_examples=total_examples,
directory=train_summaries_dir)
print("Saved learning curves!")
disc_loss_per_epoch.append(np.average(disc_losses_epoch))
gen_loss_per_epoch.append(np.average(gen_losses_epoch))
# Save epoch learning curve
save_learning_curve_epoch(gen_losses=gen_loss_per_epoch,
disc_losses=disc_loss_per_epoch,
total_epochs=epoch + 1,
directory=train_summaries_dir)
print("Saved learning curves!")
print('epoch {}/{} disc loss: {:.4f}, gen loss: {:.4f}'.format(
epoch + 1, n_epochs,
np.array(disc_losses_epoch).mean(),
np.array(gen_losses_epoch).mean()))
disc_losses.extend(disc_losses_epoch)
gen_losses.extend(gen_losses_epoch)
except KeyboardInterrupt:
print("Saving before quit...")
save_checkpoint(total_examples=total_examples,
fixed_noise=fixed_noise,
disc=disc,
gen=gen,
disc_loss_per_epoch=disc_loss_per_epoch,
gen_loss_per_epoch=gen_loss_per_epoch,
gen_losses=gen_losses,
disc_losses=disc_losses,
epoch=epoch,
directory=checkpoint_dir)
print("Checkpoint saved!")
# sample images for inspection
save_image_sample(batch=gen.forward(
fixed_noise.view(-1, hidden_size, 1, 1)),
cuda=cuda,
total_examples=total_examples,
directory=gen_images_dir)
print("Saved images!")
# save learning curves for inspection
save_learning_curve(gen_losses=gen_losses,
disc_losses=disc_losses,
total_examples=total_examples,
directory=train_summaries_dir)
print("Saved learning curves!")
def train_and_eval(model, train_loader, valid_loader, learning_rate, epochs,
model_outdir, #pos_wt,
metrics_every_iter, task, tensorboard = False,
restore_chkpt = None, run_suffix = None):
"""
Contains the powerhouse of the network, ie. the training and validation iterations called through run_model().
All parameters from the command line/json are parsed and then passed into run_model().
Performs checkpointing each epoch, saved as 'last.pth.tar' and the best model thus far (based on validation AUC), saved as 'best.pth.tar'
:param model: (nn.Module) -
:param train_loader: (torch DataLoader)
:param valid_loader: (torch DataLoader)
:param learning_rate: (float) - the learning rate, defaults to 1e-05
:param epochs: (int) - the number of epochs
:param model_outdir: (str) - the output directory for checkpointing, checkpoints will be saved as output_dir/task/view/*.tar
:param restore_chkpt: (str) - the directory to reload the checkpoint, if specified
:param run_suffix: (str) - suffix to be appended to the event file. removed for now.
:return:
"""
log_fn = helpers.create_tb_log_dir(model_outdir)
log_fn = log_fn.strip("/") # remove leading forward slash which messes up tf log
if tensorboard:
import tensorflow as tf
writer = tf.summary.create_file_writer(log_fn) # tf 2.0+
writer = tf.compat.v1.summary.FileWriter(log_fn) # tf v1.15
current_best_val_loss = float('Inf')
optimizer = torch.optim.Adam(model.parameters(), learning_rate, weight_decay=0.01)
# taken directly from MRNet code
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience = 5, # how many epochs to wait for before acting
factor = 0.3, # factor to reduce LR by, LR = factor * LR
threshold = 1e-4) # threshold to measure new optimum
#
losses = {'regression': torch.nn.MSELoss(),
'classification': torch.nn.BCEWithLogitsLoss(),
'multitask': torch.nn.MSELoss()}
#
criterion = losses[task]
print(criterion)
metric = {'regression':'MSE',
'classification':'AUC',
'multitask': 'MSE'} # Steve: Seems like this is mostly for logging?
# # TODO: reloading checkpoint
# if restore_chkpt:
# logging.info("Restoring Checkpoint from {}".format(restore_chkpt))
# helpers.load_checkpoint(checkpoint = restore_chkpt,
# model = model,
# optimizer = optimizer,
# # scheduler = scheduler,
# epochs = epochs)
# # print(loaded_epoch)
# # so epochs - loaded_epoch is where we would need to start, right?
# logging.info("Starting again at Epoch {}....".format(epochs))
# logging.info("Finished Restoring Checkpoint...")
for epoch in range(epochs):
logging.info('[Epoch {}]'.format(epoch + 1))
# main training loop
epoch_loss, epoch_metric, epoch_preds, epoch_labels, train_df = run_model(
model = model,
loader = train_loader,
optimizer = optimizer,
criterion = criterion,
metrics_every_iter = metrics_every_iter,
task = task,
tensorboard = tensorboard,
train = True)
logging.info('[Epoch {}]\t\tTraining {}: {:.3f}\t Training Average Loss: {:.5f}'\
.format(epoch + 1, metric[task], epoch_metric, epoch_loss))
# main validation loop
epoch_val_loss, epoch_val_metric, epoch_val_preds, epoch_val_labels, val_df = run_model(model = model,
loader = valid_loader,
optimizer = optimizer,
criterion = criterion,
task = task,
tensorboard = tensorboard,
metrics_every_iter = False, # default, just show the epoch validation metrics..
train = False)
logging.info('[Epoch {}]\t\tValidation {}: {:.3f}\t Validation Average Loss: {:.5f}'.format(epoch + 1, metric[task], epoch_val_metric, epoch_val_loss))
scheduler.step(epoch_val_loss) # check per epoch, how does the threshold work?!?!?
logging.info('[Epoch {}]\t\tOptimizer Learning Rate: {}'.format(epoch + 1, {optimizer.param_groups[0]['lr']}))
# with writer:#.as_default():
# temp = torch.tensor([epoch + 1]) # needs to be a tesor in tf v1.5?
# writer.add_summary(tf.compat.v1.summary.scalar('Loss/train', epoch_loss), temp).eval()
# writer.add_summary(tf.compat.v1.summary.scalar('Loss/val', epoch_val_loss), temp).eval()
# writer.add_summary(tf.compat.v1.summary.scalar('{}/train'.format(metric[task]), epoch_metric), temp).eval()
# writer.add_summary(tf.compat.v1.summary.scalar('{}/val'.format(metric[task]), epoch_val_metric), temp).eval()
# writer_flush = writer.flush()
# with writer.as_default():
# tf.summary.scalar('Loss/train', epoch_loss, epoch + 1)
# tf.summary.scalar('Loss/val', epoch_val_loss, epoch + 1)
# tf.summary.scalar('{}/train'.format(metric[task]), epoch_metric, epoch + 1)
# tf.summary.scalar('{}/val'.format(metric[task]), epoch_val_metric, epoch + 1)
print('Loss/train: {} for epoch: {}'.format(str(epoch_loss), str(epoch + 1)))
print('Loss/val: {} for epoch: {}'.format(str(epoch_val_loss), str(epoch + 1)))
print('{}/train: {} for epoch: {}'.format(metric[task], str(epoch_metric), str(epoch + 1)))
print('{}/val: {} for epoch: {}'.format(metric[task], str(epoch_val_metric), str(epoch + 1)))
# check whether the most recent epoch loss is better than previous best
is_best_val_loss = epoch_val_loss < current_best_val_loss
# save state in a dictionary
state = {'epoch': epoch + 1,
'state_dict': model.state_dict(),
'validation_metric': epoch_val_metric,
'metric': metric[task],
'best_validation_loss': epoch_val_loss,
# 'metrics': metrics # read more into this
'scheduler_dict': scheduler.state_dict(),
'optim_dict': optimizer.state_dict()}
# save as last epoch
helpers.save_checkpoint(state,
is_best = is_best_val_loss,
checkpoint_dir = model_outdir)
if is_best_val_loss:
current_best_val_loss = epoch_val_loss
logging.info('[Epoch {}]\t\t******New Best Validation Loss: {:.3f}******'.format(epoch + 1, epoch_val_loss))
helpers.save_checkpoint(state,
is_best = is_best_val_loss,
checkpoint_dir = model_outdir)
#if task == 'multitask': # Steven: Seems like this should work the same if doing regression or classification. I'll try doing the same for regression by commenting out this if statement.
# train_df.to_csv(os.path.join(model_outdir, 'best_epoch_training_results.csv'))
# val_df.to_csv(os.path.join(model_outdir, 'best_epoch_validation_results.csv'))
train_df.to_csv(os.path.join(model_outdir, 'best_epoch_training_results.csv'))
val_df.to_csv(os.path.join(model_outdir, 'best_epoch_validation_results.csv'))
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# create model
# if args.pretrained:
# print("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True)
# model = autofit(model, args.arch, args.num_classes)
# else:
# print("=> creating model '{}'".format(args.arch))
# model = models.__dict__[args.arch](num_classes=args.num_classes)
model = AutoFitNet(arch=args.arch,
pretrained=args.pretrained,
num_classes=args.num_classes)
if args.distributed:
# For multiprocessing distributed, DistributedDataParallel constructor
# should always set the single device scope, otherwise,
# DistributedDataParallel will use all available devices.
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.workers = int(args.workers / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
# DataParallel will divide and allocate batch_size to all available GPUs
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
if args.gpu is not None:
# best_acc1 may be from a checkpoint from a different GPU
best_acc1 = best_acc1.to(args.gpu)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'train')
valdir = os.path.join(args.data, 'val')
testdir = os.path.join(args.data, 'test')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# train_dataset = datasets.ImageFolder(
# traindir,
# transforms.Compose([
# transforms.Resize(256),
# transforms.RandomResizedCrop(224),
# # transforms.RandomHorizontalFlip(),
# transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1),
# transforms.ToTensor(),
# normalize,
# ]))
train_dataset = CityFuncDataset(
traindir,
transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(brightness=0.1,
contrast=0.1,
saturation=0.1,
hue=0.1),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(CityFuncDataset(
valdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.test:
test_loader = torch.utils.data.DataLoader(CityFuncDataset(
testdir,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
validate(test_loader, model, criterion, args)
return
if args.evaluate:
validate(val_loader, model, criterion, args)
return
epoch_time = AverageMeter('Time', ':6.3f', 's')
end = time.time()
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
# learning rate decay
adjust_learning_rate(optimizer, epoch, args.lr)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion, args)
# remember best acc@1 and save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
if not args.multiprocessing_distributed or (
args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best)
# measure elapsed time
epoch_time.update(time.time() - end)
eta = (args.epochs - epoch - 1) * epoch_time.avg
eta_str = str(datetime.timedelta(seconds=int(eta)))
print(
'Epoch: [{epoch:d}]\tTime:{time:6.3f}s\tETA:{eta:6.3f}s ({eta_str:s})'
.format(epoch=epoch, time=epoch_time.val, eta=eta,
eta_str=eta_str))
end = time.time()
def train(model,
train_data,
val_data,
epochs,
batch_size,
learning_rate,
savedir,
alpha=3,
beta=3,
vc_flag=True,
mix_flag=False):
best_check = {'epoch': 0, 'best': 0, 'val_acc': 0}
out_file_name = savedir + 'result.txt'
total_train = len(train_data)
train_loader = DataLoader(dataset=train_data, batch_size=1, shuffle=True)
val_loaders = []
for i in range(len(val_data)):
val_loader = DataLoader(dataset=val_data[i],
batch_size=1,
shuffle=True)
val_loaders.append(val_loader)
# we observed that training the backbone does not make a very big difference but not training saves a lot of memory
# if the backbone should be trained, then only with very small learning rate e.g. 1e-7
for param in model.backbone.parameters():
param.requires_grad = False
if not vc_flag:
model.conv1o1.weight.requires_grad = False
else:
model.conv1o1.weight.requires_grad = True
if not mix_flag:
model.mix_model.requires_grad = False
else:
model.mix_model.requires_grad = True
classification_loss = nn.CrossEntropyLoss()
cluster_loss = ClusterLoss()
optimizer = torch.optim.Adagrad(params=filter(
lambda param: param.requires_grad, model.parameters()),
lr=learning_rate)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=optimizer,
gamma=0.98)
print('Training')
for epoch in range(epochs):
out_file = open(out_file_name, 'a')
train_loss = 0.0
correct = 0
start = time.time()
model.train()
model.backbone.eval()
for index, data in enumerate(train_loader):
if index % 500 == 0 and index != 0:
end = time.time()
print('Epoch{}: {}/{}, Acc: {}, Loss: {} Time:{}'.format(
epoch + 1, index, total_train,
correct.cpu().item() / index,
train_loss.cpu().item() / index, (end - start)))
start = time.time()
input, _, label = data
input = input.cuda(device_ids[0])
label = label.cuda(device_ids[0])
output, vgg_feat, like = model(input)
out = output.argmax(1)
correct += torch.sum(out == label)
class_loss = classification_loss(output, label) / output.shape[0]
loss = class_loss
if alpha != 0:
clust_loss = cluster_loss(
vgg_feat, model.conv1o1.weight) / output.shape[0]
loss += alpha * clust_loss
if beta != 0:
mix_loss = like[0, label[0]]
loss += -beta * mix_loss
#with torch.autograd.set_detect_anomaly(True):
loss.backward()
# pseudo batches
if np.mod(index, batch_size) == 0: # and index!=0:
optimizer.step()
optimizer.zero_grad()
train_loss += loss.detach() * input.shape[0]
updated_clutter = update_clutter_model(model, device_ids)
model.clutter_model = updated_clutter
scheduler.step()
train_acc = correct.cpu().item() / total_train
train_loss = train_loss.cpu().item() / total_train
out_str = 'Epochs: [{}/{}], Train Acc:{}, Train Loss:{}'.format(
epoch + 1, epochs, train_acc, train_loss)
print(out_str)
out_file.write(out_str)
# Evaluate Validation images
model.eval()
with torch.no_grad():
correct = 0
val_accs = []
for i in range(len(val_loaders)):
val_loader = val_loaders[i]
correct_local = 0
total_local = 0
val_loss = 0
out_pred = torch.zeros(len(val_data[i].images))
for index, data in enumerate(val_loader):
input, _, label = data
input = input.cuda(device_ids[0])
label = label.cuda(device_ids[0])
output, _, _ = model(input)
out = output.argmax(1)
out_pred[index] = out
correct_local += torch.sum(out == label)
total_local += label.shape[0]
class_loss = classification_loss(output,
label) / output.shape[0]
loss = class_loss
val_loss += loss.detach() * input.shape[0]
correct += correct_local
val_acc = correct_local.cpu().item() / total_local
val_loss = val_loss.cpu().item() / total_local
val_accs.append(val_acc)
out_str = 'Epochs: [{}/{}], Val-Set {}, Val Acc:{} Val Loss:{}\n'.format(
epoch + 1, epochs, i, val_acc, val_loss)
print(out_str)
out_file.write(out_str)
val_acc = np.mean(val_accs)
out_file.write('Epochs: [{}/{}], Val Acc:{}\n'.format(
epoch + 1, epochs, val_acc))
if val_acc > best_check['val_acc']:
print('BEST: {}'.format(val_acc))
out_file.write('BEST: {}\n'.format(val_acc))
best_check = {
'state_dict': model.state_dict(),
'val_acc': val_acc,
'epoch': epoch
}
save_checkpoint(best_check,
savedir + 'vc' + str(epoch + 1) + '.pth', True)
print('\n')
out_file.close()
return best_check
losses['validate']['history'].append(validate_loss)
# always save latest checkpoint after an epoch, and flag if best checkpoint
if (epoch + 1) % 5 == 0 or is_best_model:
print('Saving checkpoint at epoch {}...'.format(epoch + 1))
logging.info('Saving checkpoint at epoch {}...'.format(epoch +
1))
model.cpu()
save_checkpoint(
{
'epoch':
epoch + 1,
'model':
model.state_dict(),
'losses':
losses,
'word_embeddings':
model.word_embeddings.weight.data.numpy(),
'pos_embeddings':
model.cpu().pos_embeddings.weight.data.numpy(),
'optimizer':
optimizer.state_dict(),
}, LATEST_CHECKPOINT_RELATIVE_PATH,
BEST_CHECKPOINT_RELATIVE_PATH, is_best_model)
if CUDA:
model.cuda()
if validate_loss > losses['validate']['min'][
'value'] and epoch - losses['validate']['min'][
'epoch'] > 10:
print(
'Ten epochs with no improvement have passed. Stopping training...'