生成器只做 12 次迭代 - 无论批量大小如何

Question

我有以下数据生成器。 它工作并返回预期数据。 除了无论我设置的 epoch 或 batchsize 等于什么，它只进行 12 次迭代，然后给出错误(见下文(

我尝试更改纪元数和批量大小。

# initialize the number of epochs to train for and batch size
NUM_EPOCHS = 10 #100
BS = 32 #64 #32
NUM_TRAIN_IMAGES = len(train_uxo_scrap)
NUM_TEST_IMAGES = len(test_uxo_scrap)

def datagenerator(imgfns, imglabels, batchsize, mode="train", class_mode='binary'):
    cnt=0
    while True:
        images = []
        labels = []
        #cnt=0
        while len(images) < batchsize and cnt < len(imgfns):
            images.append(imgfns[cnt])
            labels.append(imglabels[cnt])
            cnt=cnt+1
        print(images)
        print(labels)
        print('********** cnt = ', cnt)
        yield images, labels

train_gen = datagenerator(train_uxo_scrap, train_uxo_scrap_labels, batchsize=BS, class_mode='binary')
valid_gen = datagenerator(test_uxo_scrap, test_uxo_scrap_labels, batchsize=BS, class_mode='binary')

# train the network
H = model.fit_generator(
    train_gen,
    steps_per_epoch=NUM_TRAIN_IMAGES // BS,
    validation_data=valid_gen,
    validation_steps=NUM_TEST_IMAGES // BS,
    epochs=NUM_EPOCHS)

我希望代码在每次迭代中经历 10 个纪元和 32 个样本。 我每次迭代得到 32 个样本，但我在第 1 个纪元中只得到 12 次迭代，然后出现以下错误。 无论设置了什么批次大小或纪元，都会发生这种情况。

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-83-26f81894773d> in <module>()
      5     validation_data=valid_gen,
      6     validation_steps=NUM_TEST_IMAGES // BS,
----> 7     epochs=NUM_EPOCHS)
~AppDataLocalContinuumanaconda3envsdltf1libsite-packagestensorflowpythonkerasenginetraining.py in fit_generator(self, generator, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch)
   1424         use_multiprocessing=use_multiprocessing,
   1425         shuffle=shuffle,
-> 1426         initial_epoch=initial_epoch)
   1427 
   1428   def evaluate_generator(self,
~AppDataLocalContinuumanaconda3envsdltf1libsite-packagestensorflowpythonkerasenginetraining_generator.py in model_iteration(model, data, steps_per_epoch, epochs, verbose, callbacks, validation_data, validation_steps, class_weight, max_queue_size, workers, use_multiprocessing, shuffle, initial_epoch, mode, batch_size, **kwargs)
    182       # `batch_size` used for validation data if validation
    183       # data is NumPy/EagerTensors.
--> 184       batch_size = int(nest.flatten(batch_data)[0].shape[0])
    185 
    186       # Callbacks batch begin.
IndexError: tuple index out of range

以下是打印输出的示例：

['C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\uxo_48-81\JBCC_Norm_Formatted_48-81_#615.npy', ..., 'C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\scrap_48-81\JBCC_Norm_Formatted_48-81_#224.npy']
[1, 0, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 0, 0, 0]
********** cnt =  352
['C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\uxo_48-81\JBCC_Norm_Formatted_48-81_#532.npy', 'C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\uxo_48-81\JBCC_Norm_Formatted_48-81_#953.npy', 
...
, 'C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\scrap_48-81\JBCC_Norm_Formatted_48-81_#1081.npy', 'C:\Users\jfhauris\Documents\xtemp\ML GEO\MLGeoCode\FormattedDataStore\scrap_48-81\JBCC_Norm_Formatted_48-81_#1050.npy']
[1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 0, 1, 0, 0]
********** cnt =  384

Answer 1

看看这是否有效：

def datagenerator(imgfns, imglabels, batchsize, mode="train", class_mode='binary'):
    while True:
        start = 0
        end = batchsize
        while start  < len(imgfns): 
            x = imgfns[start:end]
            y = imglabels[start:end]
            yield x, y
            start += batchsize
            end += batchsize

假设imgfns, imglabels是 Numpy 数组。