Python model.cuda方法代码示例

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def get_loader(self):
        paths = [os.path.join(self.cache_dir, phase + '.pkl') for phase in self.config.get('train', 'phase').split()]
        dataset = utils.data.Dataset(
            utils.data.load_pickles(paths),
            transform=transform.augmentation.get_transform(self.config, self.config.get('transform', 'augmentation').split()),
            one_hot=None if self.config.getboolean('train', 'cross_entropy') else len(self.category),
            shuffle=self.config.getboolean('data', 'shuffle'),
            dir=os.path.join(self.model_dir, 'exception'),
        )
        logging.info('num_examples=%d' % len(dataset))
        try:
            workers = self.config.getint('data', 'workers')
            if torch.cuda.is_available():
                workers = workers * torch.cuda.device_count()
        except configparser.NoOptionError:
            workers = multiprocessing.cpu_count()
        collate_fn = utils.data.Collate(
            transform.parse_transform(self.config, self.config.get('transform', 'resize_train')),
            utils.train.load_sizes(self.config),
            maintain=self.config.getint('data', 'maintain'),
            transform_image=transform.get_transform(self.config, self.config.get('transform', 'image_train').split()),
            transform_tensor=transform.get_transform(self.config, self.config.get('transform', 'tensor').split()),
            dir=os.path.join(self.model_dir, 'exception'),
        )
        return torch.utils.data.DataLoader(dataset, batch_size=self.args.batch_size * torch.cuda.device_count() if torch.cuda.is_available() else self.args.batch_size, shuffle=True, num_workers=workers, collate_fn=collate_fn, pin_memory=torch.cuda.is_available()) 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def batchify(data, bsz):
    # Work out how cleanly we can divide the dataset into bsz parts.
    if isinstance(data, tuple):
        nbatch = data[0].size(0) // bsz
        # Trim off any extra elements that wouldn't cleanly fit (remainders).
        tag_data = data[1].narrow(0, 0, nbatch * bsz)
        data = data[0].narrow(0, 0, nbatch * bsz)
        # Evenly divide the data across the bsz batches.
        tag_data = tag_data.view(bsz, -1).t().contiguous()
    else:
        nbatch = data.size(0) // bsz
        # Trim off any extra elements that wouldn't cleanly fit (remainders).
        data = data.narrow(0, 0, nbatch * bsz)
    
    # Evenly divide the data across the bsz batches.
    data = data.view(bsz, -1).t().contiguous()
    # Turning the data over to CUDA at this point may lead to more OOM errors
    #if args.cuda:
     #    data = data.cuda()
    if isinstance(data,tuple):
        return data, tag_data
    return data 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def evaluate(lm_data_source, ccg_data_source):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss = 0
    ntokens = len(corpus.dictionary)
    if (not args.single) and (torch.cuda.device_count() > 1):
        #"module" is necessary when using DataParallel
        hidden = model.module.init_hidden(eval_batch_size)
    else:
        hidden = model.init_hidden(eval_batch_size)
    for i in range(0, lm_data_source.size(0) + ccg_data_source.size(0) - 1, args.bptt):
        # TAG
        if i > lm_data_source.size(0):
            data, targets = get_batch(ccg_data_source, i - lm_data_source.size(0), evaluation=True)
        # LM
        else:
            data, targets = get_batch(lm_data_source, i, evaluation=True)
        output, hidden = model(data, hidden)
        output_flat = output.view(-1, ntokens)
        curr_loss = len(data) * criterion(output_flat, targets).data
        total_loss += curr_loss
        hidden = repackage_hidden(hidden)
    if len(ccg_data_source) == 0:
        return total_loss / len(lm_data_source)
    return total_loss[0] / (len(lm_data_source)+len(ccg_data_source)) 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def ensure_model(model):
    if torch.cuda.is_available():
        model.cuda()
        if torch.cuda.device_count() > 1:
            logging.info('%d GPUs are used' % torch.cuda.device_count())
            model = nn.DataParallel(model).cuda()
    return model 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def eval(self, **kwargs):
        logging.info('evaluating')
        if torch.cuda.is_available():
            self.inference.cpu()
        try:
            e = _eval.Eval(self.args, self.config)
            cls_ap = e()
            self.backup_best(cls_ap, e.path)
        except:
            traceback.print_exc()
        if torch.cuda.is_available():
            self.inference.cuda() 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def batchify(data, bsz):
    # Work out how cleanly we can divide the dataset into bsz parts.
    nbatch = data.size(0) // bsz
    # Trim off any extra elements that wouldn't cleanly fit (remainders).
    data = data.narrow(0, 0, nbatch * bsz)
    # Evenly divide the data across the bsz batches.
    data = data.view(bsz, -1).t().contiguous()
    if args.cuda:
        data = data.cuda()
    return data 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def evaluate(split):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss, nbatches = 0, 0
    ntokens = len(corpus.dictionary.idx2word)
    hidden = model.init_hidden(args.eval_batch_size)
    for source, target in corpus.iter(split, args.eval_batch_size, args.bptt, use_cuda=args.cuda):
        model.softmax.set_target(target.data.view(-1))
        output, hidden = model(source, hidden)
        total_loss += criterion(output, target.view(-1)).data.sum()
        hidden = repackage_hidden(hidden)
        nbatches += 1
    return total_loss / nbatches 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def test_get_batch(source, evaluation=False):
    if isinstance(source, tuple):
        seq_len = len(source[0]) - 1
        data = Variable(source[0][:seq_len], volatile=evaluation)
        target = Variable(source[1][:seq_len], volatile=evaluation)
        
    else:
        seq_len = len(source) - 1
        data = Variable(source[:seq_len], volatile=evaluation)
        target = Variable(source[1:1+seq_len].view(-1))
    # This is where data should be CUDA-fied to lessen OOM errors
    if args.cuda:
        return data.cuda(), target.cuda()
    else:
        return data, target 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def get_batch(source, i, evaluation=False):
    if isinstance(source, tuple):
        seq_len = min(args.bptt, len(source[0]) - 1 - i)
        data = Variable(source[0][i:i+seq_len], volatile=evaluation)
        target = Variable(source[1][i:i+seq_len].view(-1))
    else:
        seq_len = min(args.bptt, len(source) - 1 - i)
        data = Variable(source[i:i+seq_len], volatile=evaluation)
        target = Variable(source[i+1:i+1+seq_len].view(-1))
    #This is where data should be CUDA-fied to lessen OOM errors
    if args.cuda:
        return data.cuda(), target.cuda()
    else:
        return data, target 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def train():
    global lr, best_val_loss
    # Turn on training mode which enables dropout.
    model.train()
    total_loss, nbatches = 0, 0
    start_time = time.time()
    ntokens = len(corpus.dictionary.idx2word)
    hidden = model.init_hidden(args.batch_size)
    for b, batch in enumerate(corpus.iter('train', args.batch_size, args.bptt, use_cuda=args.cuda)):
        model.train()
        source, target = batch
        # Starting each batch, we detach the hidden state from how it was previously produced.
        # If we didn't, the model would try backpropagating all the way to start of the dataset.
        hidden = repackage_hidden(hidden)
        model.zero_grad()
        model.softmax.set_target(target.data.view(-1))
        output, hidden = model(source, hidden)
        loss = criterion(output, target.view(-1))
        loss.backward()


        # `clip_grad_norm` helps prevent the exploding gradient problem in RNNs.
        torch.nn.utils.clip_grad_norm(model.parameters(), args.clip)
        for p in model.parameters():
            if p.grad is not None:
                p.data.add_(-lr, p.grad.data)

        total_loss += loss.data.cpu()

        if b % args.log_interval == 0 and b > 0:
            cur_loss = total_loss[0] / args.log_interval
            elapsed = time.time() - start_time
            val_loss = evaluate('valid')
            print('| epoch {:3d} | batch {:5d} | lr {:02.5f} | ms/batch {:5.2f} | '
                    'loss {:5.2f} | ppl {:8.2f} | valid loss {:5.2f} | valid ppl {:8.2f}'.format(
                epoch, b, lr,
                elapsed * 1000 / args.log_interval, cur_loss, math.exp(cur_loss),
                val_loss, math.exp(val_loss)))

            # Save the model if the validation loss is the best we've seen so far.
            if not best_val_loss or val_loss < best_val_loss:
                with open(args.save, 'wb') as f:
                    torch.save(model, f)
                best_val_loss = val_loss
            else:
                # Anneal the learning rate if no improvement has been seen in the validation dataset.
                lr *= args.ar

            total_loss = 0
            start_time = time.time()



# At any point you can hit Ctrl + C to break out of training early. 

# 需要导入模块: import model [as 别名]
# 或者: from model import cuda [as 别名]
def test_evaluate(test_lm_sentences, test_ccg_sentences, lm_data_source, ccg_data_source):
    # Turn on evaluation mode which disables dropout.
    model.eval()
    total_loss = 0.
    ntokens = len(corpus.dictionary)
    if args.words:
        print('word sentid sentpos wlen surp entropy')#,end='')
        if args.guess:
            for i in range(args.guessn):
                print(' guess'+str(i))#,end='')
                if args.guessscores:
                    print(' gscore'+str(i))#,end='')
        sys.stdout.write('\n')
    bar = Bar('Processing', max=len(lm_data_source)+len(ccg_data_source))
    for i in range(len(lm_data_source)+len(ccg_data_source)):
        if i >= len(lm_data_source):
            sent_ids = ccg_data_source[i-len(lm_data_source)]
            sent = test_ccg_sentences[i-len(lm_data_source)]
        else:
            sent_ids = lm_data_source[i]
            sent = test_lm_sentences[i]
        if args.cuda:
            sent_ids = sent_ids.cuda()
        if (not args.single) and (torch.cuda.device_count() > 1):
            # "module" is necessary when using DataParallel
            hidden = model.module.init_hidden(1) # number of parallel sentences being processed
        else:
            hidden = model.init_hidden(1) # number of parallel sentences being processed
        data, targets = test_get_batch(sent_ids, evaluation=True)
        data=data.unsqueeze(1) # only needed if there is just a single sentence being processed
        output, hidden = model(data, hidden)
        output_flat = output.view(-1, ntokens)
        curr_loss = criterion(output_flat, targets).item()
        #curr_loss = len(data) * criterion(output_flat, targets).data # needed if there is more than a single sentence being processed
        total_loss += curr_loss
        if args.words:
            # output word-level complexity metrics
            if i >= len(lm_data_source):
                get_complexity_apply(output_flat,targets,i-len(lm_data_source),tags=True)
            else:
                get_complexity_apply(output_flat,targets,i)
        else:
            # output sentence-level loss
            print(str(sent)+":"+str(curr_loss[0]))
        hidden = repackage_hidden(hidden)
        bar.next()
    bar.finish()
    return total_loss / (len(lm_data_source)+len(ccg_data_source))