utils.py

""" Utilities """
import os
import logging
import shutil
import torch
import torchvision.datasets as dset
import numpy as np
import preproc


def get_data(dataset, data_path, cutout_length, validation):
    """ Get torchvision dataset """
    dataset = dataset.lower()

    if dataset == 'cifar10':
        dset_cls = dset.CIFAR10
        n_classes = 10
    elif dataset == 'mnist':
        dset_cls = dset.MNIST
        n_classes = 10
    elif dataset == 'fashionmnist':
        dset_cls = dset.FashionMNIST
        n_classes = 10
    else:
        raise ValueError(dataset)

    trn_transform, val_transform = preproc.data_transforms(dataset, cutout_length)
    trn_data = dset_cls(root=data_path, train=True, download=True, transform=trn_transform)

    # assuming shape is NHW or NHWC
    shape = trn_data.data.shape   # zouxing: torchvision >= 1.2 train_data -> data
    input_channels = 3 if len(shape) == 4 else 1
    assert shape[1] == shape[2], "not expected shape = {}".format(shape)
    input_size = shape[1]

    ret = [input_size, input_channels, n_classes, trn_data]
    if validation: # append validation data
        ret.append(dset_cls(root=data_path, train=False, download=True, transform=val_transform))

    return ret


def get_logger(file_path):
    """ Make python logger """
    # [!] Since tensorboardX use default logger (e.g. logging.info()), we should use custom logger
    logger = logging.getLogger('darts')
    log_format = '%(asctime)s | %(message)s'
    formatter = logging.Formatter(log_format, datefmt='%m/%d %I:%M:%S %p')
    file_handler = logging.FileHandler(file_path)
    file_handler.setFormatter(formatter)
    stream_handler = logging.StreamHandler()
    stream_handler.setFormatter(formatter)

    logger.addHandler(file_handler)
    logger.addHandler(stream_handler)
    logger.setLevel(logging.INFO)

    return logger


def param_size(model):
    """ Compute parameter size in MB """
    n_params = sum(
        np.prod(v.size()) for k, v in model.named_parameters() if not k.startswith('aux_head'))
    return n_params / 1024. / 1024.


class AverageMeter():
    """ Computes and stores the average and current value """
    def __init__(self):
        self.reset()

    def reset(self):
        """ Reset all statistics """
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        """ Update statistics """
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

 # output: (bs, num_class) 是64行10列,  target: (bs, 1)， topk=(1,5)
def accuracy(output, target, topk=(1,)):
    """ Computes the precision@k for the specified values of k """
    maxk = max(topk)  # 5
    batch_size = target.size(0)

    # maxk=5，表示dim=1按行取值
    # output的值是精度，选top5是选这一行精度最大的五个对应的列，也就是属于哪一类
    # pred是(bs,5) 值为类别号，0，1，...,9  
    _, pred = output.topk(maxk, 1, True, True)
    pred = pred.t()     # 转置，pred:(5, bs)
    # one-hot case
    if target.ndimension() > 1:
        target = target.max(1)[1]

    # pred和target对应位置值相等返回1，不等返回0
    # target原来是64行1列，值为类别；target.view(1, -1)把target拉成一行，expand_as(pred)又把target变成5行64列 
    correct = pred.eq(target.view(1, -1).expand_as(pred))

    res = []
    for k in topk:  # k=1 和 k=5
        correct_k = correct[:k].contiguous().view(-1).float().sum(0)  # zouxing: view -> .contiguous().view
        res.append(correct_k.mul_(1.0 / batch_size))
    # res里是两个值，一个是top1的概率，一个是top5的概率
    return res


def save_checkpoint(state, ckpt_dir, is_best=False):
    filename = os.path.join(ckpt_dir, 'checkpoint.pth.tar')
    torch.save(state, filename)
    if is_best:
        best_filename = os.path.join(ckpt_dir, 'best.pth.tar')
        shutil.copyfile(filename, best_filename)