所有输入数组 (x) 应具有相同数量的样本。得到数组形状：[(52, 224, 224, 3), (64, 35)]

Question

我收到了这个错误，很奇怪第一个时期已经圆满完成，而第二个时期的尺寸不匹配。 All input arrays (x) should have the same number of samples. Got array shapes: [(52, 224, 224, 3), (64, 35)] see error screenshot

我第一次使用 ImageDataGenerator() 和方法 flow_from_directory() 从数据集中获取所有图像。这是我认为应该是错误的代码的一部分：

# generator function - two inputs (images and histogram vector) and one label (OCEAN labels)
def generator(dataset_path, OCEAN_histogram, batch_size):

    gen = ImageDataGenerator()
    gen_faces = gen.flow_from_directory(dataset_path,
                                          target_size = (224, 224),
                                          class_mode = None,
                                          batch_size = batch_size,
                                          shuffle=True)    


    batch_OCEAN = np.zeros((batch_size, 5))
    batch_histogram = np.zeros((batch_size, n_features))

    # loop where we continually feed the nn in batches
    while True:

    # it takes the next batch of images
    batch_faces = gen_faces.next()


    # list with all images names in the current batch
    all_faces = [f.rsplit('.', 1)[0].rsplit('/', 1)[1] for f in gen_faces.filenames]

    for i in range(batch_size):
        batch_OCEAN[i] = OCEAN_histogram[all_faces[i]][0]
        batch_histogram[i] = OCEAN_histogram[all_faces[i]][1]

    # shapes : batch_faces -> (64, 224, 224, 3)  batch_histogram -> (64, 35) batch_OCEAN -> (64, 5)
    yield [ batch_faces, batch_histogram ], batch_OCEAN


train_generator = generator(train_path, train_OCEAN_histogram, batch_size)
validation_generator = generator(validation_path, validation_OCEAN_histogram, batch_size)

# Train model on dataset
print("[INFO] training model...")
custom_vgg_model.fit_generator(train_generator, epochs = 50, steps_per_epoch = train_size//batch_size,
                               validation_data = validation_generator,
                               validation_steps = validation_size//batch_size, verbose = 1)

这个错误怎么可能发生？有什么想法吗？

提前致谢！

Answer 1

原因是因为您的总训练样本为 55,796，如果您的批量大小为 64，您将有 871 个步骤，余数为 52。在 871 步之后，您的生成器将在下一次迭代中返回一个形状为 (52, 224, 224, 3) 的张量。

我建议在每个 epoch 之后，对数据集进行洗牌，并且只在生成器中保留 55,796//64 样本。