Python torch.cuda方法代碼示例

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def forward(self, input):
        laplacian = input.exp() + self.eps
        output = input.clone()
        for b in range(input.size(0)):
            lap = laplacian[b].masked_fill(
                Variable(torch.eye(input.size(1)).cuda().ne(0)), 0)
            lap = -lap + torch.diag(lap.sum(0))
            # store roots on diagonal
            lap[0] = input[b].diag().exp()
            inv_laplacian = lap.inverse()

            factor = inv_laplacian.diag().unsqueeze(1)\
                                         .expand_as(input[b]).transpose(0, 1)
            term1 = input[b].exp().mul(factor).clone()
            term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone()
            term1[:, 0] = 0
            term2[0] = 0
            output[b] = term1 - term2
            roots_output = input[b].diag().exp().mul(
                inv_laplacian.transpose(0, 1)[0])
            output[b] = output[b] + torch.diag(roots_output)
        return output 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def run_weighted_rmse_net_helper(X_train, Y_train, X_test, Y_test, params, weights, i):
    X_train_ = Variable(torch.Tensor(X_train[:,:-1])).cuda()
    Y_train_ = Variable(torch.Tensor(Y_train)).cuda()
    X_test_ = Variable(torch.Tensor(X_test[:,:-1])).cuda()
    Y_test_ = Variable(torch.Tensor(Y_test)).cuda()

    model = model_classes.Net(X_train[:,:-1], Y_train, [200, 200])
    model.cuda()
    opt = optim.Adam(model.parameters(), lr=1e-3)
    solver = model_classes.SolveScheduling(params)
    for j in range(100):

        model.train()
        batch_train_weightrmse(100, i*100 + j, X_train_.data, Y_train_.data, model, opt, weights.data)

    # Rebalance weights
    model.eval()
    mu_pred_train, sig_pred_train = model(X_train_)
    Y_sched_train = solver(mu_pred_train.double(), sig_pred_train.double())
    weights2 = task_loss_no_mean(
        Y_sched_train.float(), Y_train_, params).cuda()
    model.set_sig(X_train_, Y_train_)

    return model, weights2 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def forward(self, log_prices):
        prices = torch.exp(log_prices)
        
        nBatch = prices.size(0)
        T = self.T

        Q = self.Q.unsqueeze(0).expand(nBatch, self.Q.size(0), self.Q.size(1))
        c = torch.cat(
            [prices, -prices, 
            -Variable(self.lam * self.B * torch.ones(T)).unsqueeze(0).expand(nBatch,T).cuda()], 
            1)
        A = self.A.unsqueeze(0).expand(nBatch, self.A.size(0), self.A.size(1))
        b = self.b.unsqueeze(0).expand(nBatch, self.b.size(0))
        Ae = self.Ae.unsqueeze(0).expand(nBatch, self.Ae.size(0), self.Ae.size(1))
        be = self.be.unsqueeze(0).expand(nBatch, self.be.size(0))
                
        out = QPFunction(verbose=True)\
            (Q.double(), c.double(), A.double(), b.double(), Ae.double(), be.double())
        
        return out 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def __init__(self, X, Y, hidden_layer_sizes):
        super(Net, self).__init__()

        # Initialize linear layer with least squares solution
        X_ = np.hstack([X, np.ones((X.shape[0],1))])
        Theta = np.linalg.solve(X_.T.dot(X_), X_.T.dot(Y))
        
        self.lin = nn.Linear(X.shape[1], Y.shape[1])
        W,b = self.lin.parameters()
        W.data = torch.Tensor(Theta[:-1,:].T)
        b.data = torch.Tensor(Theta[-1,:])
        
        # Set up non-linear network of 
        # Linear -> BatchNorm -> ReLU -> Dropout layers
        layer_sizes = [X.shape[1]] + hidden_layer_sizes
        layers = reduce(operator.add, 
            [[nn.Linear(a,b), nn.BatchNorm1d(b), nn.ReLU(), nn.Dropout(p=0.2)] 
                for a,b in zip(layer_sizes[0:-1], layer_sizes[1:])])
        layers += [nn.Linear(layer_sizes[-1], Y.shape[1])]
        self.net = nn.Sequential(*layers)
        self.sig = Parameter(torch.ones(1, Y.shape[1]).cuda()) 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def __init__(self, params, eps=1e-2):
        super(SolveNewsvendor, self).__init__()
        k = len(params['d'])
        self.Q = Variable(torch.diag(torch.Tensor(
            [params['c_quad']] + [params['b_quad']]*k + [params['h_quad']]*k)) \
                .cuda())
        self.p = Variable(torch.Tensor(
            [params['c_lin']] + [params['b_lin']]*k + [params['h_lin']]*k) \
                .cuda())
        self.G = Variable(torch.cat([
            torch.cat([-torch.ones(k,1), -torch.eye(k), torch.zeros(k,k)], 1),
            torch.cat([torch.ones(k,1), torch.zeros(k,k), -torch.eye(k)], 1),
            -torch.eye(1 + 2*k)], 0).cuda())
        self.h = Variable(torch.Tensor(
            np.concatenate([-params['d'], params['d'], np.zeros(1+ 2*k)])).cuda())
        self.one = Variable(torch.Tensor([1])).cuda()
        self.eps_eye = eps * Variable(torch.eye(1 + 2*k).cuda()).unsqueeze(0) 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def forward(self, y):
        nBatch, k = y.size()

        eps2 = 1e-8
        Q_scale = torch.cat([torch.diag(torch.cat(
            [self.one, y[i]+eps2, y[i]+eps2])).unsqueeze(0) for i in range(nBatch)], 0)
        Q = self.Q.unsqueeze(0).expand_as(Q_scale).mul(Q_scale)
        p_scale = torch.cat([Variable(torch.ones(nBatch,1).cuda()), y, y], 1)
        p = self.p.unsqueeze(0).expand_as(p_scale).mul(p_scale)
        G = self.G.unsqueeze(0).expand(nBatch, self.G.size(0), self.G.size(1))
        h = self.h.unsqueeze(0).expand(nBatch, self.h.size(0))
        e = Variable(torch.Tensor().cuda()).double()

        out = QPFunction(verbose=False)\
            (Q.double(), p.double(), G.double(), h.double(), e, e).float()

        return out[:,:1] 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def forward(self, input):
        laplacian = input.exp() + self.eps
        output = input.clone()
        for b in range(input.size(0)):
            lap = laplacian[b].masked_fill(
                torch.eye(input.size(1)).cuda().ne(0), 0)
            lap = -lap + torch.diag(lap.sum(0))
            # store roots on diagonal
            lap[0] = input[b].diag().exp()
            inv_laplacian = lap.inverse()

            factor = inv_laplacian.diag().unsqueeze(1)\
                                         .expand_as(input[b]).transpose(0, 1)
            term1 = input[b].exp().mul(factor).clone()
            term2 = input[b].exp().mul(inv_laplacian.transpose(0, 1)).clone()
            term1[:, 0] = 0
            term2[0] = 0
            output[b] = term1 - term2
            roots_output = input[b].diag().exp().mul(
                inv_laplacian.transpose(0, 1)[0])
            output[b] = output[b] + torch.diag(roots_output)
        return output 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def _process_cuda_events(self):
        """
        Service pending timers. Dequeue timers and aggregate Cuda time intervals,
        until the first "pending" timer (i.e. dependent on a not-yet-ready cuda event).
        """
        while len(self._cuda_pending_timers) > 0:
            timer = self._cuda_pending_timers[0]
            elapsed_cuda = 0.0

            for ev_start, ev_end in timer._cuda_event_intervals:
                if not ev_start.query() or not ev_end.query():
                    # Cuda events associated with this timer aren't ready yet,
                    # stop servicing the queue.
                    return
                # Use seconds (instead of ms) for consistency with "host" timers
                elapsed_cuda += ev_start.elapsed_time(ev_end) / 1000.0

            # All time intervals for this timer are now accounted for.
            # Aggregate stats and pop from pending queue.
            self._cuda_stats[timer.name].update(elapsed_cuda)
            self._cuda_pending_timers.popleft() 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def __init__(self, size, cuda=False):

        self.size = size
        self.done = False

        self.tt = torch.cuda if cuda else torch

        # The score for each translation on the beam.
        self.scores = self.tt.FloatTensor(size).zero_()
        self.all_scores = []
        self.all_length = []

        # The backpointers at each time-step.
        self.prevKs = []

        # The outputs at each time-step.
        self.nextYs = [self.tt.LongTensor(size).fill_(s2s.Constants.PAD)]
        self.nextYs[0][0] = s2s.Constants.BOS

        # The attentions (matrix) for each time.
        self.attn = []

    # Get the outputs for the current timestep. 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def recommend_auto_balance(message: str) -> str:
    """Expands a message with recommendation to :mod:`torchgpipe.balance`."""
    return f'''{message}

If your model is still under development, its optimal balance would change
frequently. In this case, we highly recommend 'torchgpipe.balance' for naive
automatic balancing:

  from torchgpipe import GPipe
  from torchgpipe.balance import balance_by_time

  partitions = torch.cuda.device_count()
  sample = torch.empty(...)
  balance = balance_by_time(partitions, model, sample)

  model = GPipe(model, balance, ...)
''' 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def make_translator(opt, report_score=True, out_file=None):
    if out_file is None:
        out_file = codecs.open(opt.output, 'w', 'utf-8')

    if opt.gpu > -1:
        torch.cuda.set_device(opt.gpu)

    dummy_parser = argparse.ArgumentParser(description='train.py')
    onmt.opts.model_opts(dummy_parser)
    dummy_opt = dummy_parser.parse_known_args([])[0]

    fields, model, model_opt = \
        onmt.ModelConstructor.load_test_model(opt, dummy_opt.__dict__)

    scorer = onmt.translate.GNMTGlobalScorer(opt.alpha,
                                             opt.beta,
                                             opt.coverage_penalty,
                                             opt.length_penalty)

    kwargs = {k: getattr(opt, k)
              for k in ["beam_size", "n_best", "max_length", "min_length",
                        "stepwise_penalty", "block_ngram_repeat",
                        "ignore_when_blocking", "dump_beam",
                        "data_type", "replace_unk", "gpu", "verbose"]}

    translator = Translator(model, fields, global_scorer=scorer,
                            out_file=out_file, report_score=report_score,
                            copy_attn=model_opt.copy_attn, **kwargs)
    return translator 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def _run_target(self, batch, data):
        data_type = data.data_type
        if data_type == 'text':
            _, src_lengths = batch.src
        else:
            src_lengths = None
        src = onmt.io.make_features(batch, 'src', data_type)
        tgt_in = onmt.io.make_features(batch, 'tgt')[:-1]

        #  (1) run the encoder on the src
        enc_states, memory_bank = self.model.encoder(src, src_lengths)
        dec_states = \
            self.model.decoder.init_decoder_state(src, memory_bank, enc_states)

        #  (2) if a target is specified, compute the 'goldScore'
        #  (i.e. log likelihood) of the target under the model
        tt = torch.cuda if self.cuda else torch
        gold_scores = tt.FloatTensor(batch.batch_size).fill_(0)
        dec_out, _, _ = self.model.decoder(
            tgt_in, memory_bank, dec_states, memory_lengths=src_lengths)

        tgt_pad = self.fields["tgt"].vocab.stoi[onmt.io.PAD_WORD]
        for dec, tgt in zip(dec_out, batch.tgt[1:].data):
            # Log prob of each word.
            out = self.model.generator.forward(dec)
            tgt = tgt.unsqueeze(1)
            scores = out.data.gather(1, tgt)
            scores.masked_fill_(tgt.eq(tgt_pad), 0)
            gold_scores += scores
        return gold_scores 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def async_copy_to(obj, dev, main_stream=None):
    if torch.is_tensor(obj):
        obj = Variable(obj)
    if isinstance(obj, Variable):
        v = obj.cuda(dev, async=True)
        if main_stream is not None:
            v.data.record_stream(main_stream)
        return v
    elif isinstance(obj, collections.Mapping):
        return {k: async_copy_to(o, dev, main_stream) for k, o in obj.items()}
    elif isinstance(obj, collections.Sequence):
        return [async_copy_to(o, dev, main_stream) for o in obj]
    else:
        return obj 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def _async_copy(inputs, device_ids):
    nr_devs = len(device_ids)
    assert type(inputs) in (tuple, list)
    assert len(inputs) == nr_devs

    outputs = []
    for i, dev in zip(inputs, device_ids):
        with cuda.device(dev):
            outputs.append(async_copy_to(i, dev))

    return tuple(outputs) 

# 需要導入模塊: import torch [as 別名]
# 或者: from torch import cuda [as 別名]
def _async_copy_stream(inputs, device_ids):
    nr_devs = len(device_ids)
    assert type(inputs) in (tuple, list)
    assert len(inputs) == nr_devs

    outputs = []
    streams = [_get_stream(d) for d in device_ids]
    for i, dev, stream in zip(inputs, device_ids, streams):
        with cuda.device(dev):
            main_stream = cuda.current_stream()
            with cuda.stream(stream):
                outputs.append(async_copy_to(i, dev, main_stream=main_stream))
            main_stream.wait_stream(stream)

    return outputs