Python utils.Params方法代码示例

# 需要导入模块: import utils [as 别名]
# 或者: from utils import Params [as 别名]
def main():
    # Load the 'reference' parameters from parent_dir json file
    global param_template, gpu_ids, args, search_params, model_dir

    args = parser.parse_args()
    model_dir = os.path.join('experiments', args.model_name)
    json_file = os.path.join(model_dir, 'params.json')
    assert os.path.isfile(json_file), f'No json configuration file found at {args.json}'
    param_template = utils.Params(json_file)

    gpu_ids = args.gpu_ids
    logger.info(f'Running on GPU: {gpu_ids}')

    # Perform hypersearch over parameters listed below
    search_params = {
        'lstm_dropout': np.arange(0, 0.501, 0.1, dtype=np.float32).tolist(),
        'lstm_hidden_dim': np.arange(5, 60, 10, dtype=np.int).tolist()
    }

    keys = sorted(search_params.keys())
    search_range = list(product(*[[*range(len(search_params[i]))] for i in keys]))

    start_pool(search_range, len(gpu_ids)) 

# 需要导入模块: import utils [as 别名]
# 或者: from utils import Params [as 别名]
def launch_training_job(search_range):
    '''Launch training of the model with a set of hyperparameters in parent_dir/job_name
    Args:
        search_range: one combination of the params to search
    '''

    search_range = search_range[0]
    params = {k: search_params[k][search_range[idx]] for idx, k in enumerate(sorted(search_params.keys()))}
    model_param_list = '-'.join('_'.join((k, f'{v:.2f}')) for k, v in params.items())
    model_param = copy(param_template)
    for k, v in params.items():
        setattr(model_param, k, v)

    pool_id, job_idx = multiprocessing.Process()._identity
    gpu_id = gpu_ids[pool_id - 1]

    logger.info(f'Worker {pool_id} running {job_idx} using GPU {gpu_id}')

    # Create a new folder in parent_dir with unique_name 'job_name'
    model_name = os.path.join(model_dir, model_param_list)
    model_input = os.path.join(args.model_name, model_param_list)
    if not os.path.exists(model_name):
        os.makedirs(model_name)

    # Write parameters in json file
    json_path = os.path.join(model_name, 'params.json')
    model_param.save(json_path)
    logger.info(f'Params saved to: {json_path}')

    # Launch training with this config
    cmd = f'{PYTHON} train.py ' \
        f'--model-name={model_input} ' \
        f'--dataset={args.dataset} ' \
        f'--data-folder={args.data_dir} ' \
        f'--save-best '
    if args.sampling:
        cmd += ' --sampling'
    if args.relative_metrics:
        cmd += ' --relative-metrics'

    logger.info(cmd)
    check_call(cmd, shell=True, env={'CUDA_VISIBLE_DEVICES': str(gpu_id),
                                     'OMP_NUM_THREADS': '4'}) 

# 需要导入模块: import utils [as 别名]
# 或者: from utils import Params [as 别名]
def train(model: nn.Module,
          optimizer: optim,
          loss_fn,
          train_loader: DataLoader,
          test_loader: DataLoader,
          params: utils.Params,
          epoch: int) -> float:
    '''Train the model on one epoch by batches.
    Args:
        model: (torch.nn.Module) the neural network
        optimizer: (torch.optim) optimizer for parameters of model
        loss_fn: a function that takes outputs and labels per timestep, and then computes the loss for the batch
        train_loader: load train data and labels
        test_loader: load test data and labels
        params: (Params) hyperparameters
        epoch: (int) the current training epoch
    '''
    model.train()
    loss_epoch = np.zeros(len(train_loader))
    # Train_loader:
    # train_batch ([batch_size, train_window, 1+cov_dim]): z_{0:T-1} + x_{1:T}, note that z_0 = 0;
    # idx ([batch_size]): one integer denoting the time series id;
    # labels_batch ([batch_size, train_window]): z_{1:T}.
    for i, (train_batch, idx, labels_batch) in enumerate(tqdm(train_loader)):
        optimizer.zero_grad()
        batch_size = train_batch.shape[0]

        train_batch = train_batch.permute(1, 0, 2).to(torch.float32).to(params.device)  # not scaled
        labels_batch = labels_batch.permute(1, 0).to(torch.float32).to(params.device)  # not scaled
        idx = idx.unsqueeze(0).to(params.device)

        loss = torch.zeros(1, device=params.device)
        hidden = model.init_hidden(batch_size)
        cell = model.init_cell(batch_size)

        for t in range(params.train_window):
            # if z_t is missing, replace it by output mu from the last time step
            zero_index = (train_batch[t, :, 0] == 0)
            if t > 0 and torch.sum(zero_index) > 0:
                train_batch[t, zero_index, 0] = mu[zero_index]
            mu, sigma, hidden, cell = model(train_batch[t].unsqueeze_(0).clone(), idx, hidden, cell)
            loss += loss_fn(mu, sigma, labels_batch[t])

        loss.backward()
        optimizer.step()
        loss = loss.item() / params.train_window  # loss per timestep
        loss_epoch[i] = loss
        if i % 1000 == 0:
            test_metrics = evaluate(model, loss_fn, test_loader, params, epoch, sample=args.sampling)
            model.train()
            logger.info(f'train_loss: {loss}')
        if i == 0:
            logger.info(f'train_loss: {loss}')
    return loss_epoch 

# 需要导入模块: import utils [as 别名]
# 或者: from utils import Params [as 别名]
def evaluate(model, loss_fn, dataloader, metrics, params):
    """Evaluate the model on `num_steps` batches.

    Args:
        model: (torch.nn.Module) the neural network
        loss_fn: a function that takes batch_output and batch_labels and computes the loss for the batch
        dataloader: (DataLoader) a torch.utils.data.DataLoader object that fetches data
        metrics: (dict) a dictionary of functions that compute a metric using the output and labels of each batch
        params: (Params) hyperparameters
        num_steps: (int) number of batches to train on, each of size params.batch_size
    """

    # set model to evaluation mode
    model.eval()

    # summary for current eval loop
    summ = []

    # compute metrics over the dataset
    for data_batch, labels_batch in dataloader:

        # move to GPU if available
        if params.cuda:
            data_batch, labels_batch = data_batch.cuda(async=True), labels_batch.cuda(async=True)
        # fetch the next evaluation batch
        data_batch, labels_batch = Variable(data_batch), Variable(labels_batch)
        
        # compute model output
        output_batch = model(data_batch)
        loss = loss_fn(output_batch, labels_batch)

        # extract data from torch Variable, move to cpu, convert to numpy arrays
        output_batch = output_batch.data.cpu().numpy()
        labels_batch = labels_batch.data.cpu().numpy()

        # compute all metrics on this batch
        summary_batch = {metric: metrics[metric](output_batch, labels_batch)
                         for metric in metrics}
        summary_batch['loss'] = loss.data[0]
        summ.append(summary_batch)

    # compute mean of all metrics in summary
    metrics_mean = {metric:np.mean([x[metric] for x in summ]) for metric in summ[0]} 
    metrics_string = " ; ".join("{}: {:05.3f}".format(k, v) for k, v in metrics_mean.items())
    logging.info("- Eval metrics : " + metrics_string)
    return metrics_mean