Python torch.nn.CosineEmbeddingLoss() Examples

def eval(data_iter, model):
    loss_tot, y_list, y_pred_list = 0, [], []
    model.eval()
    for x1, x2, y in data_iter:
        # if args.cuda:
        #     x1, x2, y = Variable(x1).cuda(), Variable(x2).cuda(), Variable(y).cuda()
        # else:
        #     x1, x2, y = Variable(x1), Variable(x2), Variable(y)
        out1, out2, sim = model(x1, x2)
        # loss = F.cross_entropy(output, y, size_average=False)
        criterion = nn.CosineEmbeddingLoss()
        loss = criterion(out1, out2, (2*y-1).float())
        loss_tot += loss.item()  # 0-dim scaler
        y_pred = sim.data >= 0.5
        y_pred_list.append(y_pred)
        y_list.append(y)
    y_pred = torch.cat(y_pred_list, 0)
    y = torch.cat(y_list, 0)
    acc, p, r, f1 = metrics(y, y_pred)
    size = len(data_iter.dataset)
    loss_avg = loss_tot / float(size)
    model.train()
    return loss_avg, acc, p, r, f1 

def nomal_loss(pred, targetN,mask1):
    valid_mask = (mask1 > 0.0).detach()
    pred_n = pred.permute(0,2,3,1)
    pred_n = pred_n[valid_mask]
    target_n = targetN[valid_mask]

    pred_n = pred_n.contiguous().view(-1,3)
    pred_n = F.normalize(pred_n)
    target_n = target_n.contiguous().view(-1, 3)

    loss_function = nn.CosineEmbeddingLoss()
    loss = loss_function(pred_n, target_n, Variable(torch.Tensor(pred_n.size(0)).cuda().fill_(1.0)))
    return loss 

def nomal_loss(pred, targetN,params,depthI,depthJ):
    depthI = depthI.permute(0, 2, 3, 1)
    depthJ = depthJ.permute(0, 2, 3, 1)

    predN_1 = torch.zeros_like(targetN)
    predN_2 = torch.zeros_like(targetN)

    f = params[:, :, :, 0]
    cx = params[:, :, :, 1]
    cy = params[:, :, :, 2]

    z1 = depthJ - pred
    z1 = torch.squeeze(z1)
    depthJ = torch.squeeze(depthJ)
    predN_1[:, :, :, 0] = ((MatJ - cx) * z1 + depthJ) * 1.0 / f
    predN_1[:, :, :, 1] = (MatI - cy) * z1 * 1.0 / f
    predN_1[:, :, :, 2] = z1

    z2 = depthI - pred
    z2 = torch.squeeze(z2)
    depthI = torch.squeeze(depthI)
    predN_2[:, :, :, 0] = (MatJ - cx) * z2  * 1.0 / f
    predN_2[:, :, :, 1] = ((MatI - cy) * z2 + depthI) * 1.0 / f
    predN_2[:, :, :, 2] = z2

    predN = torch.cross(predN_1, predN_2)
    pred_n = F.normalize(predN)
    pred_n = pred_n.contiguous().view(-1, 3)
    target_n = targetN.contiguous().view(-1, 3)

    loss_function = nn.CosineEmbeddingLoss()
    loss = loss_function(pred_n, target_n, Variable(torch.Tensor(pred_n.size(0)).cuda().fill_(1.0)))
    return loss 

def nomal_loss(pred, targetN,params,depthI,depthJ):
    depthI = depthI.permute(0, 2, 3, 1)
    depthJ = depthJ.permute(0, 2, 3, 1)

    predN_1 = torch.zeros_like(targetN)
    predN_2 = torch.zeros_like(targetN)

    f = params[:, :, :, 0]
    cx = params[:, :, :, 1]
    cy = params[:, :, :, 2]

    z1 = depthJ - pred
    z1 = torch.squeeze(z1)
    depthJ = torch.squeeze(depthJ)
    predN_1[:, :, :, 0] = ((MatJ - cx) * z1 + depthJ) * 1.0 / f
    predN_1[:, :, :, 1] = (MatI - cy) * z1 * 1.0 / f
    predN_1[:, :, :, 2] = z1

    z2 = depthI - pred
    z2 = torch.squeeze(z2)
    depthI = torch.squeeze(depthI)
    predN_2[:, :, :, 0] = (MatJ - cx) * z2  * 1.0 / f
    predN_2[:, :, :, 1] = ((MatI - cy) * z2 + depthI) * 1.0 / f
    predN_2[:, :, :, 2] = z2

    predN = torch.cross(predN_1, predN_2)
    pred_n = F.normalize(predN)
    pred_n = pred_n.contiguous().view(-1, 3)
    target_n = targetN.contiguous().view(-1, 3)

    loss_function = nn.CosineEmbeddingLoss()
    loss = loss_function(pred_n, target_n, Variable(torch.Tensor(pred_n.size(0)).cuda().fill_(1.0)))
    return loss 

def __init__(self, args, src_dict, tgt_dict, src_embedding, tgt_embedding, device):
        super(E2E, self).__init__(args)

        self.args = args
        self.src_dict = src_dict
        self.tgt_dict = tgt_dict

        # src_flow: assume tgt embeddings are transformed from the src mog space
        self.register_buffer('src_embedding', src_embedding)
        self.register_buffer('tgt_embedding', tgt_embedding)

        if args.init_var:
            # initialize with gaussian variance
            self.register_buffer("s2t_s_var", src_dict.var)
            self.register_buffer("s2t_t_var", tgt_dict.var)
            self.register_buffer("t2s_s_var", src_dict.var)
            self.register_buffer("t2s_t_var", tgt_dict.var)
        else:
            self.s2t_s_var = args.s_var
            self.s2t_t_var = args.s2t_t_var
            self.t2s_t_var = args.t_var
            self.t2s_s_var = args.t2s_s_var

        self.register_buffer('src_freqs', torch.tensor(src_dict.freqs, dtype=torch.float))
        self.register_buffer('tgt_freqs', torch.tensor(tgt_dict.freqs, dtype=torch.float))

        # backward: t2s
        self.src_flow = MogFlow_batch(args, self.t2s_s_var)
        # backward: s2t
        self.tgt_flow = MogFlow_batch(args, self.s2t_t_var)
        
        self.s2t_valid_dico = None
        self.t2s_valid_dico = None

        self.device = device
        # use dict pairs from train data (supervise) or identical words (supervise_id) as supervisions
        self.supervise = args.supervise_id
        if self.supervise:
            self.load_training_dico()
            if args.sup_obj == 'mse':
                self.sup_loss_func = nn.MSELoss()
            elif args.sup_obj == 'cosine':
                self.sup_loss_func = CosineEmbeddingLoss()

        optim_fn, optim_params= get_optimizer(args.flow_opt_params)
        self.flow_optimizer = optim_fn(list(self.src_flow.parameters()) + list(self.tgt_flow.parameters()), **optim_params)
        self.flow_scheduler = torch.optim.lr_scheduler.ExponentialLR(self.flow_optimizer, gamma=args.lr_decay)

        self.best_valid_metric = 1e-12

        self.sup_sw = args.sup_s_weight
        self.sup_tw = args.sup_t_weight

        self.mse_loss = nn.MSELoss()
        self.cos_loss = CosineEmbeddingLoss()

        # Evaluation on trained model
        if args.load_from_pretrain_s2t != "" or args.load_from_pretrain_t2s != "":
            self.load_from_pretrain() 

def main():
   
    model = im2recipe()
    model.visionMLP = torch.nn.DataParallel(model.visionMLP)
    model.to(device)

    # define loss function (criterion) and optimizer
    # cosine similarity between embeddings -> input1, input2, target
    cosine_crit = nn.CosineEmbeddingLoss(0.1).to(device)
    if opts.semantic_reg:
        weights_class = torch.Tensor(opts.numClasses).fill_(1)
        weights_class[0] = 0 # the background class is set to 0, i.e. ignore
        # CrossEntropyLoss combines LogSoftMax and NLLLoss in one single class
        class_crit = nn.CrossEntropyLoss(weight=weights_class).to(device)
        # we will use two different criteria
        criterion = [cosine_crit, class_crit]
    else:
        criterion = cosine_crit

    print("=> loading checkpoint '{}'".format(opts.model_path))
    if device.type=='cpu':
        checkpoint = torch.load(opts.model_path, encoding='latin1', map_location='cpu')
    else:
        checkpoint = torch.load(opts.model_path, encoding='latin1')
    opts.start_epoch = checkpoint['epoch']
    model.load_state_dict(checkpoint['state_dict'])
    print("=> loaded checkpoint '{}' (epoch {})"
          .format(opts.model_path, checkpoint['epoch']))

    # data preparation, loaders
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
    
    # preparing test loader 
    test_loader = torch.utils.data.DataLoader(
        ImagerLoader(opts.img_path,
 	    transforms.Compose([
            transforms.Scale(256), # rescale the image keeping the original aspect ratio
            transforms.CenterCrop(224), # we get only the center of that rescaled
            transforms.ToTensor(),
            normalize,
        ]),data_path=opts.data_path,sem_reg=opts.semantic_reg,partition='test'),
        batch_size=opts.batch_size, shuffle=False,
        num_workers=opts.workers, pin_memory=True)
    print('Test loader prepared.')

    # run test
    test(test_loader, model, criterion) 

def main():
   
    model = im2recipe()
    #model.visionMLP = torch.nn.DataParallel(model.visionMLP, device_ids=[0,1,2,3])
    # barelo:
    model.visionMLP = torch.nn.DataParallel(model.visionMLP, device_ids=[0])
    # model.visionMLP = torch.nn.DataParallel(model.visionMLP, device_ids=[0,1])
    if not opts.no_cuda:
        model.cuda()

    # define loss function (criterion) and optimizer
    # cosine similarity between embeddings -> input1, input2, target
    cosine_crit = nn.CosineEmbeddingLoss(0.1)
    if not opts.no_cuda:
        cosine_crit.cuda()
    # cosine_crit = nn.CosineEmbeddingLoss(0.1)
    if opts.semantic_reg:
        weights_class = torch.Tensor(opts.numClasses).fill_(1)
        weights_class[0] = 0 # the background class is set to 0, i.e. ignore
        # CrossEntropyLoss combines LogSoftMax and NLLLoss in one single class
        class_crit = nn.CrossEntropyLoss(weight=weights_class)
        if not opts.no_cuda:
            class_crit.cuda()
        # class_crit = nn.CrossEntropyLoss(weight=weights_class)
        # we will use two different criteria
        criterion = [cosine_crit, class_crit]
    else:
        criterion = cosine_crit

    print("=> loading checkpoint '{}'".format(opts.model_path))
    checkpoint = torch.load(opts.model_path)
    opts.start_epoch = checkpoint['epoch']
    model.load_state_dict(checkpoint['state_dict'])
    print("=> loaded checkpoint '{}' (epoch {})"
          .format(opts.model_path, checkpoint['epoch']))

    # data preparation, loaders
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
    
    # preparing test loader 
    test_loader = torch.utils.data.DataLoader(
        ImagerLoader(opts.img_path,
 	    transforms.Compose([
            transforms.Scale(256), # rescale the image keeping the original aspect ratio
            transforms.CenterCrop(224), # we get only the center of that rescaled
            transforms.ToTensor(),
            normalize,
        ]),data_path=opts.data_path,sem_reg=opts.semantic_reg,partition='test'),
        batch_size=opts.batch_size, shuffle=False,
        num_workers=opts.workers, pin_memory=(not opts.no_cuda))
    print 'Test loader prepared.'

    # run test
    test(test_loader, model, criterion) 

def train(x, y, wv, model, epochs, abstracts_file):
    """
    This method takes the inputs, and the labels
    and trains the LSTM network to predict the
    embeddings based on the input sequnce.

    Arguments
    ---------
    x : List of input descriptions
    y : List of target entity embeddings
    wv : Keyed Word Vectors for pre-trained embeddings
    """
    # epochs = 100
    # embed_size = 100
    # hidden_size = 50
    # seq_len = 1
    # num_layers = 2
    inputs = x
    labels = y
    itr = len(inputs)
    progress = 0.0
    criterion = nn.CosineEmbeddingLoss()
    optimizer = optim.SGD(model.parameters(), lr=0.2)
    count = 0
    itr_p = 0.0
    loss = 0
    losses = np.zeros(epochs)
    logging.info('training on {0} samples'.format(itr))
    for epoch in range(epochs):
        logging.info(
            'starting epoch {0}'.format(epoch + 1))
        flags = Variable(torch.ones(1))
        count = 0
        hidden = model.init_hidden()
        for i, l in zip(inputs, labels):
            count += 1
            progress = (count / itr) * 100
            print('INFO : PROGRESS : {0:.1f} %'.format(progress), end='\r')
            for word in i:
                output, hidden = model(
                    Variable(torch.tensor([[word]])),
                    hidden)
            optimizer.zero_grad()
            loss = criterion(output[0],
                             Variable(torch.tensor([l])),
                             flags)
            loss.backward(retain_graph=True)
            optimizer.step()
        logging.info(
            'completed epoch {0}, loss : {1}'.format(epoch + 1, loss.item()))
        losses[epoch] = loss.item()
    logging.info('saving the model to model/description_encoder')
    torch.save(model, 'model/description_encoder')
    validate(model, wv, abstracts_file)
    plt.plot(losses)
    plt.title('Model Loss')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.show()