if episode_length % 1000 == 0:
if tag_txt:
path = checkpoint_path + '/' + str(episode_length) + '_txt.model'
torch.save(model_txt.state_dict(), path)
if tag_img:
path = checkpoint_path + '/' + str(episode_length) + '_img.model'
torch.save(model_img.state_dict(), path)
if tag_txt:
model_txt.train()
if tag_img:
model_img.train()
if dataset == 'mir':
ori_img_fc7, pos_img_fc7, neg_img_fc7, ori_img_fc6, pos_img_fc6, neg_img_fc6, ori_img_p5, pos_img_p5, neg_img_p5, ori_txt, pos_txt, neg_txt, lab = traintest.get_batch(
batch_size)
if tag_img:
ori_img_fc7 = Variable(ori_img_fc7).cuda()
pos_img_fc7 = Variable(pos_img_fc7).cuda()
neg_img_fc7 = Variable(neg_img_fc7).cuda()
ori_img_fc6 = Variable(ori_img_fc6).cuda()
pos_img_fc6 = Variable(pos_img_fc6).cuda()
neg_img_fc6 = Variable(neg_img_fc6).cuda()
ori_img_p5 = Variable(ori_img_p5).cuda()
pos_img_p5 = Variable(pos_img_p5).cuda()
neg_img_p5 = Variable(neg_img_p5).cuda()
hash_o_img = Variable(torch.zeros(batch_size, 1).cuda())
if update_lr_flag:
if opt.learning_rate_decay_start >= 0 and epoch > opt.learning_rate_decay_start:
frac = (epoch - opt.learning_rate_decay_start
) // opt.learning_rate_decay_every
decay_factor = opt.learning_rate_decay_rate**frac
current_lr = opt.learning_rate * decay_factor
utils.set_lr(optimizer, current_lr)
if opt.scheduled_sampling_start >= 0 and epoch > opt.scheduled_sampling_start:
frac = (epoch - opt.scheduled_sampling_start
) // opt.scheduled_sampling_increase_every
model.ss_prob = min(opt.scheduled_sampling_increase_prob**frac,
opt.scheduled_sampling_max_prob)
update_lr_flag = False
# 加载训练数据
data = loader.get_batch('train')
torch.cuda.synchronize()
buf = [
data['fc_feats'], data['att_feats'], data['captions'], data['masks']
]
buf = [
Variable(torch.from_numpy(d), requires_grad=False).cuda() for d in buf
]
fc_feats, att_feats, captions, masks = buf
optimizer.zero_grad()
output = model(fc_feats, att_feats, captions)
loss = criterion(output, captions[:, 1:], masks[:, 1:])
loss.backward()