def train(model, criterion, optimizer, loader, epoch, args):
model.train()
loss = 0.0
acc = 0.0
num_samples = 0
bar = progressbar.ProgressBar(max_value=len(loader))
for idx, data_batch in enumerate(loader):
subj_batch_var = data_batch["subject"]["embedding"].cuda()
obj_batch_var = data_batch["object"]["embedding"].cuda()
predicate = data_batch["predicate"].cuda()
label_batch_var = torch.squeeze(data_batch["label"]).cuda()
if args.model == "drnet":
img = data_batch["bbox_img"].cuda()
mask_batch_var = data_batch["bbox_mask"].cuda()
output = model(subj_batch_var, obj_batch_var, img, mask_batch_var,
predicate)
elif args.model == "vtranse":
img = data_batch["full_img"].cuda()
ts_batch_var = data_batch["subject"]["t"].cuda()
to_batch_var = data_batch["object"]["t"].cuda()
bboxs_batch_var = data_batch["subject"]["bbox"].cuda()
bboxo_batch_var = data_batch["object"]["bbox"].cuda()
output = model(
subj_batch_var,
obj_batch_var,
img,
ts_batch_var,
to_batch_var,
bboxs_batch_var,
bboxo_batch_var,
predicate,
)
elif args.model == "vipcnn" or args.model == "pprfcn":
img = data_batch["full_img"].cuda()
bbox_s = data_batch["subject"]["bbox"].cuda()
bbox_o = data_batch["object"]["bbox"].cuda()
output = model(img, bbox_s, bbox_o, predicate)
loss_batch_var = criterion(output, label_batch_var)
loss_batch = loss_batch_var.item()
loss += len(data_batch["label"]) * loss_batch
acc += num_true_positives(output, label_batch_var)
num_samples += len(data_batch["label"])
optimizer.zero_grad()
loss_batch_var.backward()
optimizer.step()
bar.update(idx)
loss /= num_samples
acc /= num_samples / 100.0
return loss, acc
def train(model, criterion, optimizer, loader, epoch, args):
model.train()
loss = 0.
acc = 0.
num_samples = 0
bar = progressbar.ProgressBar(max_value=len(loader))
for idx, data_batch in enumerate(loader):
subj_batch_var = data_batch['subject']['embedding'].cuda()
obj_batch_var = data_batch['object']['embedding'].cuda()
predicate = data_batch['predicate'].cuda()
label_batch_var = torch.squeeze(data_batch['label']).cuda()
if args.model == 'drnet':
img = data_batch['bbox_img'].cuda()
mask_batch_var = data_batch['bbox_mask'].cuda()
output = model(subj_batch_var, obj_batch_var, img, mask_batch_var,
predicate)
elif args.model == 'vtranse':
img = data_batch['full_img'].cuda()
ts_batch_var = data_batch['subject']['t'].cuda()
to_batch_var = data_batch['object']['t'].cuda()
bboxs_batch_var = data_batch['subject']['bbox'].cuda()
bboxo_batch_var = data_batch['object']['bbox'].cuda()
output = model(subj_batch_var, obj_batch_var, img, ts_batch_var,
to_batch_var, bboxs_batch_var, bboxo_batch_var,
predicate)
elif args.model == 'vipcnn' or args.model == 'pprfcn':
img = data_batch['full_img'].cuda()
bbox_s = data_batch['subject']['bbox'].cuda()
bbox_o = data_batch['object']['bbox'].cuda()
output = model(img, bbox_s, bbox_o, predicate)
loss_batch_var = criterion(output, label_batch_var)
loss_batch = loss_batch_var.item()
loss += (len(data_batch['label']) * loss_batch)
acc += num_true_positives(output, label_batch_var)
num_samples += len(data_batch['label'])
optimizer.zero_grad()
loss_batch_var.backward()
optimizer.step()
bar.update(idx)
loss /= num_samples
acc /= (num_samples / 100.)
return loss, acc
def train(model, criterion, optimizer, loader, epoch, args):
"train for one epoch"
model.train()
loss = 0.0
acc = 0.0
num_samples = 0
bar = progressbar.ProgressBar(max_value=len(loader))
for idx, data_batch in enumerate(loader):
subj_batch_var = data_batch["subject"]["bbox"]
obj_batch_var = data_batch["object"]["bbox"]
label_batch_var = torch.squeeze(data_batch["label"]).view(-1, 1)
predi_batch_var = data_batch["predicate"]
if torch.cuda.is_available():
subj_batch_var = subj_batch_var.cuda()
obj_batch_var = obj_batch_var.cuda()
label_batch_var = label_batch_var.cuda()
predi_batch_var = predi_batch_var.cuda()
output_var = model(subj_batch_var, obj_batch_var, predi_batch_var)
loss_batch_var = criterion(output_var, label_batch_var)
loss_batch = loss_batch_var.item()
loss += len(data_batch["label"]) * loss_batch
acc += num_true_positives(output_var, label_batch_var)
num_samples += len(data_batch["label"])
optimizer.zero_grad()
loss_batch_var.backward()
optimizer.step()
bar.update(idx)
loss /= num_samples
acc /= num_samples / 100.0
return loss, acc
def train(model, criterion, optimizer, loader, epoch, args):
'train for one epoch'
model.train()
loss = 0.
acc = 0.
num_samples = 0
bar = progressbar.ProgressBar(max_value=len(loader))
for idx, data_batch in enumerate(loader):
subj_batch_var = data_batch['subject']['embedding']
obj_batch_var = data_batch['object']['embedding']
label_batch_var = torch.squeeze(data_batch['label'])
predi_batch_var = data_batch['predicate']
if torch.cuda.is_available():
subj_batch_var = subj_batch_var.cuda()
obj_batch_var = obj_batch_var.cuda()
label_batch_var = label_batch_var.cuda()
predi_batch_var = predi_batch_var.cuda()
output_var = model(subj_batch_var, obj_batch_var, predi_batch_var)
loss_batch_var = criterion(output_var, label_batch_var)
loss_batch = loss_batch_var.item()
loss += (len(data_batch['label']) * loss_batch)
acc += num_true_positives(output_var, label_batch_var)
num_samples += len(data_batch['label'])
optimizer.zero_grad()
loss_batch_var.backward()
optimizer.step()
bar.update(idx)
loss /= num_samples
acc /= (num_samples / 100.)
return loss, acc