def main():
args = get_args()
logger = get_logger(args.logdir)
logger.info(vars(args))
with tf.device("/gpu:0"):
if args.densenet:
model = DenseNet()
else:
model = CnnLstm()
tfrecord = TFRecord()
tfrecord.make_iterator(args.tfr_fname)
total = sum(1 for _ in tf.python_io.tf_record_iterator(args.tfr_fname)) // tfrecord.batch_size
saver = tf.train.Saver(tf.global_variables())
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
for epoch in range(args.num_epochs):
logger.info(f"Epoch {epoch+1}")
sess.run(tfrecord.init_op)
loss_list = []
progress_bar = tqdm(total=total, desc="[TRAIN] Loss: 0", unit="batch", leave=False)
while True:
try:
step = sess.run(model.global_step)
spec, label = tfrecord.load(sess, training=True)
_, loss, merged = model.train(sess, spec, label)
progress_bar.update(1)
progress_bar.set_description(f"[TRAIN] Batch Loss: {loss:.4f}")
loss_list.append(loss)
writer.add_summary(summary=merged, global_step=step)
except tf.errors.OutOfRangeError:
break
progress_bar.close()
mean_loss = np.mean(loss_list)
logger.info(f" - [TRAIN] Mean Loss: {mean_loss:.4f}")
saver.save(sess, args.logdir + "/" + args.save_fname + ".ckpt", global_step=sess.run(model.global_step))
class Experiment(object):
def __init__(self,
directory,
epochs=1,
cuda=False,
save=False,
log_interval=30,
load=None,
split=(0.6, 0.2, 0.2),
cache=False,
minibatch_size=10,
pretrained=False):
self.dataset = Dataset(directory,
split=split,
cache=cache,
minibatch_size=minibatch_size)
self.epochs = epochs
self.cuda = cuda
self.save = save
self.log_interval = log_interval
self.model = DenseNet(pretrained)
self.optimizer = torch.optim.Adam(self.model.parameters(), lr=0.01)
if load is not None:
state = torch.load(load)
self.model.load_state_dict(state['model'])
self.optimizer.load_state_dict(state['optim'])
if cuda:
self.model = self.model.cuda()
def train(self):
print('Training %s epochs.' % self.epochs)
loss_fun = nn.CrossEntropyLoss()
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer,
'min',
verbose=True,
patience=3)
last_print = time.time()
for epoch in range(self.epochs):
tprint('Starting epoch: %s' % epoch)
self.model.train()
self.optimizer.zero_grad()
for minibatch, targets in self.dataset.train:
minibatch = Variable(torch.stack(minibatch))
targets = Variable(torch.LongTensor(targets))
if self.cuda:
minibatch = minibatch.cuda()
targets = targets.cuda()
out = self.model.forward(minibatch)
loss = loss_fun(out, targets)
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
if time.time() - last_print > self.log_interval:
last_print = time.time()
numer, denom = self.dataset.train.progress()
tprint('Training: %s, %s/%s' % (epoch, numer, denom))
tprint('Training complete. Beginning validation.')
self.dataset.train.reload()
self.model.eval()
last_print = time.time()
for minibatch, targets in self.dataset.validate:
minibatch = Variable(torch.stack(minibatch), volatile=True)
targets = Variable(torch.LongTensor(targets), volatile=True)
if self.cuda:
minibatch = minibatch.cuda()
targets = targets.cuda()
out = self.model.forward(minibatch)
validation_loss = loss_fun(out, targets)
if time.time() - last_print > self.log_interval:
last_print = time.time()
numer, denom = self.dataset.validate.progress()
tprint('Validating: %s, %s/%s' % (epoch, numer, denom))
self.dataset.validate.reload()
scheduler.step(validation_loss.data[0])
if self.save:
torch.save(
{
'model': self.model.state_dict(),
'optim': self.optimizer.state_dict(),
}, 'signet.%s.pth' % int(time.time()))
def test(self):
tprint('Beginning testing.')
confusion_matrix = np.zeros((7, 7)).astype(np.int)
last_print = time.time()
for minibatch, targets in self.dataset.test:
minibatch = Variable(torch.stack(minibatch), volatile=True)
targets = Variable(torch.LongTensor(targets), volatile=True)
if self.cuda:
minibatch = minibatch.cuda()
targets = targets.cuda()
out = self.model.forward(minibatch)
_, predicted = torch.max(out.data, 1)
predicted = predicted.cpu().numpy()
targets = targets.data.cpu().numpy()
confusion_matrix += sklearn.metrics.confusion_matrix(
predicted, targets, labels=[0, 1, 2, 3, 4, 5,
6]).astype(np.int)
if time.time() - last_print > self.log_interval:
last_print = time.time()
numer, denom = self.dataset.test.progress()
tprint('Testing: %s/%s' % (numer, denom))
tprint('Testing complete.')
print(confusion_matrix)
print(tabulate.tabulate(stats(confusion_matrix)))
class Trainer(object):
"""
The Trainer class encapsulates all the logic necessary for
training the DenseNet model. It use SGD to update the weights
of the model given hyperparameters constraints provided by the
user in the config file.
"""
def __init__(self, config, data_loader):
"""
Construct a new Trainer instance.
Params
------
- config: object containing command line arguments.
- data_loader: data iterator
"""
self.config = config
if config.is_train:
self.train_loader = data_loader[0]
self.valid_loader = data_loader[1]
else:
self.test_loader = data_loader
# network params
self.num_blocks = config.num_blocks
self.num_layers_total = config.num_layers_total
self.growth_rate = config.growth_rate
self.bottleneck = config.bottleneck
self.theta = config.compression
# training params
self.epochs = config.epochs
self.start_epoch = 0
self.best_valid_acc = 0.
self.init_lr = config.init_lr
self.lr = self.init_lr
self.is_decay = True
self.momentum = config.momentum
self.weight_decay = config.weight_decay
self.dropout_rate = config.dropout_rate
if config.lr_sched == '':
self.is_decay = False
else:
self.lr_decay = [float(x) for x in config.lr_sched.split(',')]
# other params
self.ckpt_dir = config.ckpt_dir
self.logs_dir = config.logs_dir
self.num_gpu = config.num_gpu
self.use_tensorboard = config.use_tensorboard
self.resume = config.resume
self.print_freq = config.print_freq
self.dataset = config.dataset
if self.dataset == 'cifar10':
self.num_classes = 10
elif self.dataset == 'cifar100':
self.num_classes = 100
else:
self.num_classes = 1000
# build densenet model
self.model = DenseNet(self.num_blocks, self.num_layers_total,
self.growth_rate, self.num_classes, self.bottleneck,
self.dropout_rate, self.theta)
print('[*] Number of model parameters: {:,}'.format(
sum([p.data.nelement() for p in self.model.parameters()])))
# define loss and optimizer
self.criterion = nn.CrossEntropyLoss()
self.optimizer = torch.optim.SGD(self.model.parameters(), lr=self.init_lr,
momentum=self.momentum, weight_decay=self.weight_decay)
if self.num_gpu > 0:
self.model.cuda()
self.criterion.cuda()
# finally configure tensorboard logging
if self.use_tensorboard:
tensorboard_dir = self.logs_dir + self.get_model_name()
print('[*] Saving tensorboard logs to {}'.format(tensorboard_dir))
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
configure(tensorboard_dir)
def train(self):
"""
Train the model on the training set.
A checkpoint of the model is saved after each epoch
and if the validation accuracy is improved upon,
a separate ckpt is created for use on the test set.
"""
# switch to train mode for dropout
self.model.train()
# load the most recent checkpoint
if self.resume:
self.load_checkpoint(best=False)
for epoch in trange(self.start_epoch, self.epochs):
# decay learning rate
if self.decay:
self.anneal_learning_rate(epoch)
# train for 1 epoch
self.train_one_epoch(epoch)
# evaluate on validation set
valid_acc = self.validate(epoch)
is_best = valid_acc > self.best_valid_acc
self.best_valid_acc = max(valid_acc, self.best_valid_acc)
self.save_checkpoint({
'epoch': epoch + 1,
'state_dict': self.model.state_dict(),
'best_valid_acc': self.best_valid_acc}, is_best)
def test(self):
"""
Test the model on the held-out test data.
This function should only be called at the very
end once the model has finished training.
"""
# switch to test mode for dropout
self.model.eval()
accs = AverageMeter()
batch_time = AverageMeter()
# load the best checkpoint
self.load_checkpoint(best=True)
tic = time.time()
for i, (image, target) in enumerate(self.test_loader):
if self.num_gpu > 0:
image = image.cuda()
target = target.cuda(async=True)
input_var = torch.autograd.Variable(image)
target_var = torch.autograd.Variable(target)
# forward pass
output = self.model(input_var)
# compute loss & accuracy
acc = self.accuracy(output.data, target)
accs.update(acc, image.size()[0])
# measure elapsed time
toc = time.time()
batch_time.update(toc-tic)
# print to screen
if i % self.print_freq == 0:
print('Test: [{0}/{1}]\t'
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Test Acc: {acc.val:.3f} ({acc.avg:.3f})'.format(
i, len(self.test_loader), batch_time=batch_time,
acc=accs))
print('[*] Test Acc: {acc.avg:.3f}'.format(acc=accs))
def train_one_epoch(self, epoch):
"""
Train the model for 1 epoch of the training set.
An epoch corresponds to one full pass through the entire
training set in successive mini-batches.
This is used by train() and should not be called manually.
"""
batch_time = AverageMeter()
losses = AverageMeter()
accs = AverageMeter()
tic = time.time()
for i, (image, target) in enumerate(self.train_loader):
if self.num_gpu > 0:
image = image.cuda()
target = target.cuda(async=True)
input_var = torch.autograd.Variable(image)
target_var = torch.autograd.Variable(target)
# forward pass
output = self.model(input_var)
# compute loss & accuracy
loss = self.criterion(output, target_var)
acc = self.accuracy(output.data, target)
losses.update(loss.data[0], image.size()[0])
accs.update(acc, image.size()[0])
# compute gradients and update SGD
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# measure elapsed time
toc = time.time()
batch_time.update(toc-tic)
# print to screen
if i % self.print_freq == 0:
print('Epoch: [{0}][{1}/{2}]\t'
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Train Loss: {loss.val:.4f} ({loss.avg:.4f})\t'
'Train Acc: {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch, i, len(self.train_loader), batch_time=batch_time,
loss=losses, acc=accs))
# log to tensorboard
if self.use_tensorboard:
log_value('train_loss', losses.avg, epoch)
log_value('train_acc', accs.avg, epoch)
def validate(self, epoch):
"""
Evaluate the model on the validation set.
"""
batch_time = AverageMeter()
losses = AverageMeter()
accs = AverageMeter()
tic = time.time()
for i, (image, target) in enumerate(self.valid_loader):
if self.num_gpu > 0:
image = image.cuda()
target = target.cuda(async=True)
input_var = torch.autograd.Variable(image)
target_var = torch.autograd.Variable(target)
# forward pass
output = self.model(input_var)
# compute loss & accuracy
loss = self.criterion(output, target_var)
acc = self.accuracy(output.data, target)
losses.update(loss.data[0], image.size()[0])
accs.update(acc, image.size()[0])
# measure elapsed time
toc = time.time()
batch_time.update(toc-tic)
# print to screen
if i % self.print_freq == 0:
print('Valid: [{0}/{1}]\t'
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Valid Loss: {loss.val:.4f} ({loss.avg:.4f})\t'
'Valid Acc: {acc.val:.3f} ({acc.avg:.3f})'.format(
i, len(self.valid_loader), batch_time=batch_time,
loss=losses, acc=accs))
print('[*] Valid Acc: {acc.avg:.3f}'.format(acc=accs))
# log to tensorboard
if self.use_tensorboard:
log_value('val_loss', losses.avg, epoch)
log_value('val_acc', accs.avg, epoch)
return accs.avg
def save_checkpoint(self, state, is_best):
"""
Save a copy of the model so that it can be loaded at a future
date. This function is used when the model is being evaluated
on the test data.
Furthermore, the model with the highest accuracy is saved as
with a special name.
"""
print("[*] Saving model to {}".format(self.ckpt_dir))
filename = self.get_model_name() + '_ckpt.pth.tar'
ckpt_path = os.path.join(self.ckpt_dir, filename)
torch.save(state, ckpt_path)
if is_best:
filename = self.get_model_name() + '_model_best.pth.tar'
shutil.copyfile(ckpt_path,
os.path.join(self.ckpt_dir, filename))
print("[*] ==== Best Valid Acc Achieved ====")
def load_checkpoint(self, best=False):
"""
Load the best copy of a model. This is useful for 2 cases:
- Resuming training with the most recent model checkpoint.
- Loading the best validation model to evaluate on the test data.
Params
------
- best: if set to True, loads the best model. Use this if you want
to evaluate your model on the test data. Else, set to False in
which case the most recent version of the checkpoint is used.
"""
print("[*] Loading model from {}".format(self.ckpt_dir))
filename = self.get_model_name() + '_ckpt.pth.tar'
if best:
filename = self.get_model_name() + '_model_best.pth.tar'
ckpt_path = os.path.join(self.ckpt_dir, filename)
ckpt = torch.load(ckpt_path)
# load variables from checkpoint
self.start_epoch = ckpt['epoch']
self.best_valid_acc = ckpt['best_valid_acc']
self.model.load_state_dict(ckpt['state_dict'])
print("[*] Loaded {} checkpoint @ epoch {} with best valid acc of {:.3f}".format(
filename, ckpt['epoch'], ckpt['best_valid_acc']))
def anneal_learning_rate(self, epoch):
"""
This function decays the learning rate at 2 instances.
- The initial learning rate is divided by 10 at
t1*epochs.
- It is further divided by 10 at t2*epochs.
t1 and t2 are floats specified by the user. The default
values used by the authors of the paper are 0.5 and 0.75.
"""
sched1 = int(self.lr_sched[0] * self.epochs)
sched2 = int(self.lr_sched[1] * self.epochs)
self.lr = self.init_lr * (0.1 ** (epoch // sched1)) \
* (0.1 ** (epoch // sched2))
# log to tensorboard
if self.use_tensorboard:
log_value('learning_rate', self.lr, epoch)
for param_group in self.optimizer.param_groups:
param_group['lr'] = self.lr
def get_model_name(self):
"""
Returns the name of the model based on the configuration
parameters.
The name will take the form DenseNet-X-Y-Z where:
- X: total number of layers specified by `config.total_num_layers`.
- Y: can be BC or an empty string specified by `config.bottleneck`.
- Z: name of the dataset specified by `config.dataset`.
For example, given 169 layers with bottleneck on CIFAR-10, this
function will output `DenseNet-BC-169-cifar10`.
"""
if self.bottleneck:
return 'DenseNet-BC-{}-{}'.format(self.num_layers_total,
self.dataset)
return 'DenseNet-{}-{}'.format(self.num_layers_total,
self.dataset)
def accuracy(self, predicted, ground_truth):
"""
Utility function for calculating the accuracy of the model.
Params
------
- predicted: (torch.FloatTensor)
- ground_truth: (torch.LongTensor)
Returns
-------
- acc: (float) % accuracy.
"""
predicted = torch.max(predicted, 1)[1]
total = len(ground_truth)
correct = (predicted == ground_truth).sum()
acc = 100 * (correct / total)
return acc
class Solver(object):
DEFAULTS = {}
def __init__(self, version, data_loader, config):
"""
Initializes a Solver object
"""
# data loader
self.__dict__.update(Solver.DEFAULTS, **config)
self.version = version
self.data_loader = data_loader
self.build_model()
# TODO: build tensorboard
# start with a pre-trained model
if self.pretrained_model:
self.load_pretrained_model()
def build_model(self):
"""
Instantiates the model, loss criterion, and optimizer
"""
# instantiate model
self.model = DenseNet(config=self.config,
channels=self.input_channels,
class_count=self.class_count,
num_features=self.num_features,
compress_factor=self.compress_factor,
expand_factor=self.expand_factor,
growth_rate=self.growth_rate)
# instantiate loss criterion
self.criterion = nn.CrossEntropyLoss()
# instantiate optimizer
self.optimizer = optim.SGD(params=self.model.parameters(),
lr=self.lr,
momentum=self.momentum,
weight_decay=self.weight_decay,
nesterov=True)
# print network
self.print_network(self.model, 'DenseNet')
# use gpu if enabled
if torch.cuda.is_available() and self.use_gpu:
self.model.cuda()
self.criterion.cuda()
def print_network(self, model, name):
"""
Prints the structure of the network and the total number of parameters
"""
num_params = 0
for p in model.parameters():
num_params += p.numel()
print(name)
print(model)
print("The number of parameters: {}".format(num_params))
def load_pretrained_model(self):
"""
loads a pre-trained model from a .pth file
"""
self.model.load_state_dict(torch.load(os.path.join(
self.model_save_path, '{}.pth'.format(self.pretrained_model))))
print('loaded trained model ver {}'.format(self.pretrained_model))
def print_loss_log(self, start_time, iters_per_epoch, e, i, loss):
"""
Prints the loss and elapsed time for each epoch
"""
total_iter = self.num_epochs * iters_per_epoch
cur_iter = e * iters_per_epoch + i
elapsed = time.time() - start_time
total_time = (total_iter - cur_iter) * elapsed / (cur_iter + 1)
epoch_time = (iters_per_epoch - i) * elapsed / (cur_iter + 1)
epoch_time = str(datetime.timedelta(seconds=epoch_time))
total_time = str(datetime.timedelta(seconds=total_time))
elapsed = str(datetime.timedelta(seconds=elapsed))
log = "Elapsed {}/{} -- {}, Epoch [{}/{}], Iter [{}/{}], " \
"loss: {:.4f}".format(elapsed,
epoch_time,
total_time,
e + 1,
self.num_epochs,
i + 1,
iters_per_epoch,
loss)
# TODO: add tensorboard
print(log)
def save_model(self, e):
"""
Saves a model per e epoch
"""
path = os.path.join(
self.model_save_path,
'{}/{}.pth'.format(self.version, e + 1)
)
torch.save(self.model.state_dict(), path)
def model_step(self, images, labels):
"""
A step for each iteration
"""
# set model in training mode
self.model.train()
# empty the gradients of the model through the optimizer
self.optimizer.zero_grad()
# forward pass
output = self.model(images)
# compute loss
loss = self.criterion(output, labels.squeeze())
# compute gradients using back propagation
loss.backward()
# update parameters
self.optimizer.step()
# return loss
return loss
def train(self):
"""
Training process
"""
self.losses = []
self.top_1_acc = []
self.top_5_acc = []
iters_per_epoch = len(self.data_loader)
# start with a trained model if exists
if self.pretrained_model:
start = int(self.pretrained_model.split('/')[-1])
else:
start = 0
# start training
start_time = time.time()
for e in range(start, self.num_epochs):
for i, (images, labels) in enumerate(tqdm(self.data_loader)):
images = to_var(images, self.use_gpu)
labels = to_var(labels, self.use_gpu)
loss = self.model_step(images, labels)
# print out loss log
if (e + 1) % self.loss_log_step == 0:
self.print_loss_log(start_time, iters_per_epoch, e, i, loss)
self.losses.append((e, loss))
# save model
if (e + 1) % self.model_save_step == 0:
self.save_model(e)
# evaluate on train dataset
if (e + 1) % self.train_eval_step == 0:
top_1_acc, top_5_acc = self.train_evaluate(e)
self.top_1_acc.append((e, top_1_acc))
self.top_5_acc.append((e, top_5_acc))
# print losses
print('\n--Losses--')
for e, loss in self.losses:
print(e, '{:.4f}'.format(loss))
# print top_1_acc
print('\n--Top 1 accuracy--')
for e, acc in self.top_1_acc:
print(e, '{:.4f}'.format(acc))
# print top_5_acc
print('\n--Top 5 accuracy--')
for e, acc in self.top_5_acc:
print(e, '{:.4f}'.format(acc))
def eval(self, data_loader):
"""
Returns the count of top 1 and top 5 predictions
"""
# set the model to eval mode
self.model.eval()
top_1_correct = 0
top_5_correct = 0
total = 0
with torch.no_grad():
for images, labels in data_loader:
images = to_var(images, self.use_gpu)
labels = to_var(labels, self.use_gpu)
output = self.model(images)
total += labels.size()[0]
# top 1
# get the max for each instance in the batch
_, top_1_output = torch.max(output.data, dim=1)
top_1_correct += torch.sum(torch.eq(labels.squeeze(),
top_1_output))
# top 5
_, top_5_output = torch.topk(output.data, k=5, dim=1)
for i, label in enumerate(labels):
if label in top_5_output[i]:
top_5_correct += 1
return top_1_correct.item(), top_5_correct, total
def train_evaluate(self, e):
"""
Evaluates the performance of the model using the train dataset
"""
top_1_correct, top_5_correct, total = self.eval(self.data_loader)
log = "Epoch [{}/{}]--top_1_acc: {:.4f}--top_5_acc: {:.4f}".format(
e + 1,
self.num_epochs,
top_1_correct / total,
top_5_correct / total
)
print(log)
return top_1_correct / total, top_5_correct / total
def test(self):
"""
Evaluates the performance of the model using the test dataset
"""
top_1_correct, top_5_correct, total = self.eval(self.data_loader)
log = "top_1_acc: {:.4f}--top_5_acc: {:.4f}".format(
top_1_correct / total,
top_5_correct / total
)
print(log)