torch自定义随机加载

版权声明：本文为博主原创文章，未经博主允许不得转载。 https://blog.csdn.net/jacke121/article/details/82770796

  import random

    data_img = []
    data_lable = []

    m_index_i = 0
    for astep, samples in enumerate(dataloader):
        start = time.time()
        images, labels = samples["image"], samples["label"]

        for i in range(images.size(0)):
            data_img.append(images[i, :, :, :])
            data_lable.append(labels[i, :, :])

            asamples = torch.Tensor(config["batch_size"] * len(config["parallels"]), 3, config["img_w"],
                                    config["img_h"])
            asamples[0] = data_img[m_index_i]
            m_index_i += 1
    m_index = [i for i in range(len(data_img))]

    # print("len(data_img)",len(data_img))
    # for j in range(200):
    #     random.shuffle(m_index)
    #     timestr = datetime.datetime.now().strftime('%Y%m%d_%H%M%S_%f')
    #     samples = torch.Tensor(config["batch_size"]* len(config["parallels"]), 3, config["img_w"],config["img_h"])
    #     for i in range(m_index_i):
    #         sample = data_img[m_index[i]]
    #         label = data_lable[m_index[i]]
    #         samples[i % config["batch_size"]* len(config["parallels"])] = sample
    #         if i % config["batch_size"]* len(config["parallels"]) == (config["batch_size"]* len(config["parallels"])-1):
    #             # samples = torch.Tensor(90, 3, 352, 352)
    #             print(i,timestr)

    best_acc = 0.2
    next_need = 0
    batch_size=config["batch_size"] * len(config["parallels"])
    samples = torch.Tensor(batch_size, 3, config["img_w"], config["img_h"])
    labels = torch.Tensor(batch_size, 10, 5)
    for epoch in range(config["epochs"]):
        recall = 0
        random.shuffle(m_index)
        timestr = datetime.datetime.now().strftime('%Y%m%d_%H%M%S_%f')
        step = 0
        for i in range(m_index_i):
            # print(i,timestr)
            sample = data_img[m_index[i]]
            label = data_lable[m_index[i]]
            samples[i % batch_size] = sample
            labels[i % batch_size] = label
            if i % batch_size == (batch_size - 1):
                # samples = torch.Tensor(90, 3, 352, 352)
                config["global_step"] += 1
                # Forward and backward
                optimizer.zero_grad()
                losses = net(samples.cuda(), labels.cuda())
                step += 1