print("[*] Training")
best_avg_acc = 0
for epoch in range(args.epochs):
# Train
tag = '' # tag = '*' if the model was saved in the last epoch
train_total_loss = 0
running_acc_flickr, running_acc_sketches = 0, 0
for i, (first, second, first_label, second_label,
similarity) in enumerate(train_loader):
# first: flickr, second: sketches
first, second, similarity = first.to(args.device), second.to(
args.device), similarity.to(args.device)
first_label, second_label = first_label.to(
args.device), second_label.to(args.device)
pred = contrastive_net.forward((first, second, None),
include_negative=False)
pred_feats_first = pred['feats']['anchor']
pred_feats_second = pred['feats']['positive']
pred_logits_first = pred['logits']['anchor']
pred_logits_second = pred['logits']['positive']
# region: optimization step
optimizer.zero_grad()
loss_contrastive = contrastive_loss(pred_feats_first,
pred_feats_second, similarity)
loss_classifier_sketches = cross_entropy_loss(
pred_logits_first, first_label)
loss_classifier_flickr = cross_entropy_loss(
pred_logits_second, second_label)
loss = 0.6 * (0.5 *
(loss_classifier_flickr +
config.learning_rate = args.lr
print("[*] Training")
best_avg_acc = 0
for epoch in range(args.epochs):
# Train
tag = '' # tag = '*' if the model was saved in the last epoch
train_total_loss = 0
running_acc_anchor, running_acc_positive, running_acc_negative = 0, 0, 0
for i, (a, p, n, a_l, p_l, n_l) in enumerate(train_loader):
# first: flickr, second: sketches
a, p, n = a.to(args.device), p.to(args.device), n.to(args.device)
a_l, p_l, n_l = a_l.to(args.device), p_l.to(args.device), n_l.to(
args.device)
pred = siamese_net.forward((a, p, n), include_negative=True)
pred_feats_anchor = pred['feats']['anchor']
pred_feats_positive = pred['feats']['positive']
pred_feats_negative = pred['feats']['negative']
pred_logits_anchor = pred['logits']['anchor']
pred_logits_positive = pred['logits']['positive']
pred_logits_negative = pred['logits']['negative']
# region: optimization step
optimizer.zero_grad()
loss_triplet = triplet_loss(pred_feats_anchor, pred_feats_positive,
pred_feats_negative)
loss_classifier_anchor = cross_entropy_loss(
pred_logits_anchor, a_l)