def eval(model, dev_loader, encoder, gpu_mode, write_to_file=False):
model.eval()
print('evaluating model...')
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3, write_to_file)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
with torch.no_grad():
mx = len(dev_loader)
for i, (img_id, img, verb, labels) in enumerate(dev_loader):
print("{}/{} batches\r".format(i + 1, mx)),
#print(img_id[0], encoder.verb2_role_dict[encoder.verb_list[verb[0]]])
if gpu_mode >= 0:
img = torch.autograd.Variable(img.cuda())
labels = torch.autograd.Variable(labels.cuda())
verb = torch.autograd.Variable(verb.cuda())
else:
img = torch.autograd.Variable(img)
labels = torch.autograd.Variable(labels)
verb = torch.autograd.Variable(verb)
verb_predict, role_predict = model(img)
top1.add_point_eval5_log_sorted(img_id, verb_predict, verb,
role_predict, labels)
top5.add_point_eval5_log_sorted(img_id, verb_predict, verb,
role_predict, labels)
del verb_predict, img, verb
return top1, top5, 0
def eval(model, dev_loader, encoder, gpu_mode, write_to_file=False):
model.eval()
print('evaluating model...')
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3, write_to_file)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
with torch.no_grad():
for i, (img_id, img, verb, labels) in enumerate(dev_loader):
#print(img_id[0], encoder.verb2_role_dict[encoder.verb_list[verb[0]]])
if gpu_mode >= 0:
img = torch.autograd.Variable(img.cuda())
verb = torch.autograd.Variable(verb.cuda())
labels = torch.autograd.Variable(labels.cuda())
labels = torch.autograd.Variable(labels.cuda())
else:
img = torch.autograd.Variable(img)
verb = torch.autograd.Variable(verb)
labels = torch.autograd.Variable(labels)
role_predict = model(img, verb)
top1.add_point_noun(verb, role_predict, labels)
top5.add_point_noun(verb, role_predict, labels)
del role_predict, img, verb, labels
return top1, top5, 0
def train(model,
train_loader,
dev_loader,
optimizer,
scheduler,
max_epoch,
model_dir,
encoder,
gpu_mode,
clip_norm,
model_name,
model_saving_name,
eval_frequency=4000):
model.train()
train_loss = 0
total_steps = 0
print_freq = 400
dev_score_list = []
if gpu_mode >= 0:
ngpus = 2
device_array = [i for i in range(0, ngpus)]
pmodel = torch.nn.DataParallel(model, device_ids=device_array)
else:
pmodel = model
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
for epoch in range(max_epoch):
for i, (_, img, verb, labels) in enumerate(train_loader):
total_steps += 1
if gpu_mode >= 0:
img = torch.autograd.Variable(img.cuda())
verb = torch.autograd.Variable(verb.cuda())
labels = torch.autograd.Variable(labels.cuda())
else:
img = torch.autograd.Variable(img)
verb = torch.autograd.Variable(verb)
labels = torch.autograd.Variable(labels)
role_predict = pmodel(img, verb)
loss = model.calculate_loss(verb, role_predict, labels)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), clip_norm)
optimizer.step()
optimizer.zero_grad()
train_loss += loss.item()
top1.add_point_noun(verb, role_predict, labels)
top5.add_point_noun(verb, role_predict, labels)
if total_steps % print_freq == 0:
top1_a = top1.get_average_results_nouns()
top5_a = top5.get_average_results_nouns()
print("{},{},{}, {} , {}, loss = {:.2f}, avg loss = {:.2f}".
format(total_steps - 1, epoch, i,
utils.format_dict(top1_a, "{:.2f}", "1-"),
utils.format_dict(top5_a, "{:.2f}", "5-"),
loss.item(), train_loss /
((total_steps - 1) % eval_frequency)))
if total_steps % eval_frequency == 0:
top1, top5, val_loss = eval(model, dev_loader, encoder,
gpu_mode)
model.train()
top1_avg = top1.get_average_results_nouns()
top5_avg = top5.get_average_results_nouns()
avg_score = top1_avg["verb"] + top1_avg["value"] + top1_avg["value-all"] + top5_avg["verb"] + \
top5_avg["value"] + top5_avg["value-all"] + top5_avg["value*"] + top5_avg["value-all*"]
avg_score /= 8
print('Dev {} average :{:.2f} {} {}'.format(
total_steps - 1, avg_score * 100,
utils.format_dict(top1_avg, '{:.2f}', '1-'),
utils.format_dict(top5_avg, '{:.2f}', '5-')))
dev_score_list.append(avg_score)
max_score = max(dev_score_list)
if max_score == dev_score_list[-1]:
torch.save(
model.state_dict(), model_dir +
"/{}_{}.model".format(model_name, model_saving_name))
print('New best model saved! {0}'.format(max_score))
print('current train loss', train_loss)
train_loss = 0
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
del role_predict, loss, img, verb, labels
print('Epoch ', epoch, ' completed!')
scheduler.step()
def eval(model, loader, encoder, logging=False):
model.eval()
verbloss = 0
nounsloss = 0
gtloss = 0
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
with torch.no_grad():
for __, (_, img, verb, nouns) in enumerate(loader):
if torch.cuda.is_available():
img = img.cuda()
verb = verb.cuda()
nouns = nouns.cuda()
with autocast(): # automix precision stuff
pred_verb, pred_nouns, pred_gt_nouns = model(img, verb)
top1.add_point_both(pred_verb, verb, pred_nouns, nouns,
pred_gt_nouns)
top5.add_point_both(pred_verb, verb, pred_nouns, nouns,
pred_gt_nouns)
if torch.cuda.is_available():
vl = model.module.verb_loss(pred_verb, verb)
nl = model.module.nouns_loss(pred_nouns, nouns)
gtl = model.module.nouns_loss(pred_gt_nouns, nouns)
else:
vl = model.verb_loss(pred_verb, verb)
nl = model.nouns_loss(pred_nouns, nouns)
gtl = model.nouns_loss(pred_gt_nouns, nouns)
verbloss += vl.item()
nounsloss += nl.item()
gtloss += gtl.item()
verbloss /= len(loader)
nounsloss /= len(loader)
gtloss /= len(loader)
val_losses = {
'verb_loss': verbloss,
'nouns_loss': nounsloss,
'gt_loss': gtloss
}
#print scores
avg_score = 0
if logging is True:
top1_a = top1.get_average_results_both()
top5_a = top5.get_average_results_both()
avg_score = top1_a['verb'] + top1_a['value'] + top1_a['value-all'] + \
top5_a['verb'] + top5_a['value'] + top5_a['value-all'] + \
top1_a['gt-value'] + top1_a['gt-value-all']
avg_score /= 8
avg_score = avg_score * 100
print('val losses = [v: {:.2f}, n: {:.2f}, gt: {:.2f}]'.format(
val_losses['verb_loss'], val_losses['nouns_loss'],
val_losses['gt_loss']))
gt = {key: top1_a[key] for key in ['gt-value', 'gt-value-all']}
one_val = {key: top1_a[key] for key in ['verb', 'value', 'value-all']}
print('{}\n{}\n{}, mean = {:.2f}\n'.format(
utils.format_dict(one_val, '{:.2f}', '1-'),
utils.format_dict(top5_a, '{:.2f}', '5-'),
utils.format_dict(gt, '{:.2f}', ''), avg_score))
return top1, top5, val_losses, avg_score
def train(model,
train_loader,
dev_loader,
optimizer,
max_epoch,
encoder,
model_saving_name,
folder,
checkpoint=None):
model.train()
avg_scores = []
verb_losses = []
nouns_losses = []
val_avg_scores = []
val_verb_losses = []
val_nouns_losses = []
epoch = 0
# if checkpoint resume stuffs
if checkpoint is not None:
epoch = checkpoint['epoch']
avg_scores = checkpoint['avg_scores']
verb_losses = checkpoint['verb_losses']
nouns_losses = checkpoint['nouns_losses']
val_avg_scores = checkpoint['val_avg_scores']
val_verb_losses = checkpoint['val_verb_losses']
val_nouns_losses = checkpoint['val_nouns_losses']
if torch.cuda.is_available():
model.module.load_state_dict(checkpoint['model_state_dict'])
else:
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
#mix precision stuff
scaler = GradScaler()
for e in range(epoch, max_epoch):
verb_loss_accum = 0
nouns_loss_accum = 0
gt_nouns_loss_accum = 0
print('Epoch-{}, lr: {:.4f}'.format(e,
optimizer.param_groups[0]['lr']))
top1 = imsitu_scorer.imsitu_scorer(encoder, 1, 3)
top5 = imsitu_scorer.imsitu_scorer(encoder, 5, 3)
for i, (_, img, verb, nouns) in enumerate(train_loader):
if torch.cuda.is_available():
img = img.cuda()
verb = verb.cuda()
nouns = nouns.cuda()
optimizer.zero_grad()
with autocast(): #mix precision stuff
pred_verb, pred_nouns, pred_gt_nouns = model(img, verb)
#predict and calculate lossess
if torch.cuda.is_available():
verb_loss = model.module.verb_loss(pred_verb, verb)
nouns_loss = model.module.nouns_loss(pred_nouns, nouns)
gt_nouns_loss = model.module.nouns_loss(
pred_gt_nouns, nouns)
else:
verb_loss = model.verb_loss(pred_verb, verb)
nouns_loss = model.nouns_loss(pred_nouns, nouns)
gt_nouns_loss = model.nouns_loss(pred_gt_nouns, nouns)
loss = verb_loss + nouns_loss
# backpropagate errors and stuffs
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
torch.nn.utils.clip_grad_norm_(model.parameters(), 1)
scaler.step(optimizer)
scaler.update()
top1.add_point_both(pred_verb, verb, pred_nouns, nouns,
pred_gt_nouns)
top5.add_point_both(pred_verb, verb, pred_nouns, nouns,
pred_gt_nouns)
verb_loss_accum += verb_loss.item()
nouns_loss_accum += nouns_loss.item()
gt_nouns_loss_accum += gt_nouns_loss.item()
#fin epoch
#epoch accuracy mean
top1_a = top1.get_average_results_both()
top5_a = top5.get_average_results_both()
avg_score = top1_a['verb'] + top1_a['value'] + top1_a['value-all'] + \
top5_a['verb'] + top5_a['value'] + top5_a['value-all'] + \
top1_a['gt-value'] + top1_a['gt-value-all']
avg_score /= 8
avg_score *= 100
avg_scores.append(avg_score)
#epoch lossess
verb_loss_mean = verb_loss_accum / len(train_loader)
nouns_loss_mean = nouns_loss_accum / len(train_loader)
gt_nouns_loss_mean = gt_nouns_loss_accum / len(train_loader)
verb_losses.append(verb_loss_mean)
nouns_losses.append(nouns_loss_mean)
#print stuffs
print('training losses = [v: {:.2f}, n: {:.2f}, gt: {:.2f}]'.format(
verb_loss_mean, nouns_loss_mean, gt_nouns_loss_mean))
gt = {key: top1_a[key] for key in ['gt-value', 'gt-value-all']}
one_val = {key: top1_a[key] for key in ['verb', 'value', 'value-all']}
print('{}\n{}\n{}, mean = {:.2f}\n{}'.format(
utils.format_dict(one_val, '{:.2f}', '1-'),
utils.format_dict(top5_a, '{:.2f}', '5-'),
utils.format_dict(gt, '{:.2f}', ''), avg_score, '-' * 50))
# evaluating
top1, top5, val_losses, val_avg_score = eval(model,
dev_loader,
encoder,
logging=True)
model.train()
#val mean scores
val_avg_scores.append(val_avg_score)
val_verb_losses.append(val_losses['verb_loss'])
val_nouns_losses.append(val_losses['nouns_loss'])
plt.plot(verb_losses, label='verb losses')
plt.plot(nouns_losses, label='nouns losses')
plt.plot(avg_scores, label='accuracy mean')
plt.plot(val_verb_losses, '-.', label='val verb losses')
plt.plot(val_nouns_losses, '-.', label='val nouns losses')
plt.plot(val_avg_scores, '-.', label='val accuracy mean')
plt.grid()
plt.legend()
plt.savefig(pjoin(folder, model_saving_name + '.png'))
plt.clf()
#always saving but no need if it's not the best score
checkpoint = {
'epoch': e + 1,
'avg_scores': avg_scores,
'verb_losses': verb_losses,
'nouns_losses': nouns_losses,
'val_avg_scores': val_avg_scores,
'val_verb_losses': val_verb_losses,
'val_nouns_losses': val_nouns_losses,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict()
}
if torch.cuda.is_available():
checkpoint.update({'model_state_dict': model.module.state_dict()})
torch.save(checkpoint, pjoin(folder, model_saving_name))