# Utils
import logging
import time
import numpy as np
from tqdm import tqdm
# from sklearn.metrics import jaccard_similarity_score
# Torch related stuff
import torch
# DeepDIVA
from util.misc import AverageMeter
from util.evaluation.metrics import accuracy
[docs]def train(train_loader, model, criterion, optimizer, writer, epoch, no_cuda=False, log_interval=25,
**kwargs):
"""
Training routine
Parameters
----------
train_loader : torch.utils.data.DataLoader
The dataloader of the train set.
model : torch.nn.module
The network model being used.
criterion : torch.nn.loss
The loss function used to compute the loss of the model.
optimizer : torch.optim
The optimizer used to perform the weight update.
writer : tensorboardX.writer.SummaryWriter
The tensorboard writer object. Used to log values on file for the tensorboard visualization.
epoch : int
Number of the epoch (for logging purposes).
no_cuda : boolean
Specifies whether the GPU should be used or not. A value of 'True' means the CPU will be used.
log_interval : int
Interval limiting the logging of mini-batches. Default value of 10.
Returns
----------
top1.avg : float
Accuracy of the model of the evaluated split
"""
multi_run = kwargs['run'] if 'run' in kwargs else None
# Instantiate the counters
batch_time = AverageMeter()
loss_meter = AverageMeter()
jss_meter = AverageMeter()
data_time = AverageMeter()
# Switch to train mode (turn on dropout & stuff)
model.train()
# Empty lists to store the predictions and target values
preds = []
targets = []
# Iterate over whole training set
end = time.time()
pbar = tqdm(enumerate(train_loader), total=len(train_loader), unit='batch', ncols=150, leave=False)
for batch_idx, (input, target) in pbar:
# Measure data loading time
data_time.update(time.time() - end)
# Moving data to GPU
if not no_cuda:
input = input.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
jss, loss, target_vals, pred_vals = train_one_mini_batch(model, criterion, optimizer, input, target,
loss_meter, jss_meter)
# Store results of each minibatch
_ = [preds.append(item) for item in pred_vals]
_ = [targets.append(item) for item in target_vals]
# Add loss and accuracy to Tensorboard
if multi_run is None:
writer.add_scalar('train/mb_loss', loss.item(), epoch * len(train_loader) + batch_idx)
# writer.add_scalar('train/mb_jaccard_similarity', jss, epoch * len(train_loader) + batch_idx)
else:
writer.add_scalar('train/mb_loss_{}'.format(multi_run), loss.item(),
epoch * len(train_loader) + batch_idx)
# writer.add_scalar('train/mb_jaccard_similarity_{}'.format(multi_run), jss,
# epoch * len(train_loader) + batch_idx)
# Measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
# Log to console
if batch_idx % log_interval == 0:
pbar.set_description('train epoch [{0}][{1}/{2}]\t'.format(epoch, batch_idx, len(train_loader)))
pbar.set_postfix(Time='{batch_time.avg:.3f}\t'.format(batch_time=batch_time),
Loss='{loss.avg:.4f}\t'.format(loss=loss_meter),
# JSS='{jss_meter.avg:.3f}\t'.format(jss_meter=jss_meter),
Data='{data_time.avg:.3f}\t'.format(data_time=data_time))
# Generate the epoch wise JSS
targets = np.array(targets).astype(np.int)
preds = np.array(preds).astype(np.int)
jss_epoch = compute_jss(targets, preds)
# try:
# np.testing.assert_approx_equal(jss_epoch, jss_meter.avg)
# except:
# logging.error('Computed JSS scores do not match')
# logging.error('JSS: {} Avg: {}'.format(jss_epoch, jss_meter.avg))
# Logging the epoch-wise accuracy
if multi_run is None:
writer.add_scalar('train/loss', loss_meter.avg, epoch)
writer.add_scalar('train/jaccard_similarity', jss_epoch, epoch)
else:
writer.add_scalar('train/loss_{}'.format(multi_run), loss_meter.avg, epoch)
writer.add_scalar('train/jaccard_similarity_{}'.format(multi_run), jss_epoch, epoch)
logging.debug('Train epoch[{}]: '
'JSS={jss_epoch:.3f}\t'
'Loss={loss.avg:.4f}\t'
'Batch time={batch_time.avg:.3f} ({data_time.avg:.3f} to load data)'
.format(epoch, batch_time=batch_time, data_time=data_time, loss=loss_meter, jss_epoch=jss_epoch))
return jss_epoch
[docs]def train_one_mini_batch(model, criterion, optimizer, input, target, loss_meter, jss_meter):
"""
This routing train the model passed as parameter for one mini-batch
Parameters
----------
model : torch.nn.module
The network model being used.
criterion : torch.nn.loss
The loss function used to compute the loss of the model.
optimizer : torch.optim
The optimizer used to perform the weight update.
input : torch.autograd.Variable
The input data for the mini-batch
target : torch.autograd.Variable
The target data (labels) for the mini-batch
loss_meter : AverageMeter
Tracker for the overall loss
jss_meter : AverageMeter
Tracker for the overall Jaccard Similarity Score
Returns
-------
loss : float
Loss for this mini-batch
"""
# Compute output
output = model(input)
# Compute and record the loss
loss = criterion(output, target)
loss_meter.update(loss.item(), len(input))
# Apply sigmoid and take everything above a threshold of 0.5
squashed_output = torch.nn.Sigmoid()(output).data.cpu().numpy()
preds = get_preds_from_minibatch(squashed_output)
target_vals = target.data.cpu().numpy().astype(np.int)
# # Compute and record the Jaccard Similarity Score
# jss = compute_jss(target_vals, preds)
# jss_meter.update(jss, len(input))
jss = None
# Reset gradient
optimizer.zero_grad()
# Compute gradients
loss.backward()
# Perform a step by updating the weights
optimizer.step()
return jss, loss, target_vals, preds
[docs]def get_preds_from_minibatch(minibatch):
preds = []
for row in minibatch:
tmp = [1 if item > 0.5 else 0 for item in row]
preds.append(tmp)
preds = np.array(preds).astype(np.int)
return preds
[docs]def compute_jss(target, preds):
score = 0
num_classes = len(target[0])
for i in range(num_classes):
score += jaccard_similarity_score(target[:, i], preds[:, i])
score = score / num_classes
return score
[docs]def jaccard_similarity_score(targets, preds):
assert len(targets) == len(preds)
assert len(targets.shape) == 1
assert len(preds.shape) == 1
locs_targets = set(np.where(targets == 1)[0])
locs_preds = set(np.where(preds == 1)[0])
try:
score = len(locs_targets.intersection(locs_preds)) / len(locs_targets.union(locs_preds))
except:
print('Exception!')
return score