使用随机洗牌的批处理作为来自图像数据张量的CNN输入



我正在尝试使用STL-10数据集训练网络。

我已经从STL-10二进制文件中提取了数据,并将它们转换为Numpy数组。然后,我使用tf.convert_to_tensor函数将它们转换为张量

现在我有一个形状的张量(5000,96,96,3(

我想从包含5000张图像的数据的张量中获得一批尺寸32,并且这些批次将在每次迭代中随机洗牌。

使用tf.train.batch给出了错误

`TypeError: `Tensor` objects are not iterable when eager execution is not enabled. To iterate over this tensor use tf.map_fn.`

我如何获得一批尺寸32的图像数据,这些数据将在每次迭代中随机洗牌?

来自 tf.train.batch的文档:

参数张量可以是张量的列表或词典。这 该函数返回的值将与张量相同。

您需要将数据转换为5000个张量的列表,每个张量(96,96,3(。

您可以直接在TensorFlow函数中使用Numpy阵列,因为TensorFlow知道如何转换它们。

# Form shuffled batch of data
def get_batch(inputs, targets, size):          
    '''
    Return a total of `size` random inputs and targets(or labels). 
    '''
    targets_shape = targets.shape
    num_data = targets_shape[0]
    # this is a list of the right number of indices in the indices range
    shuffled_indices = np.random.randint(0,num_data,size)
    #this takes the selected random elements
    inputs_shuffled = inputs[shuffled_indices,:,:,:]
    #depending on the target shape it could be  targets[idx,:,:,...]
    #this takes the corresponding targets
    targets_shuffle = targets[shuffled_indices,:]
    #return the shuffled data and targets
    return inputs_shuffled, targets_shuffle

然后您可以在培训中使用它:

#This calls the function we created    
inputs_batch, targets_batch = get_batch(inputs_all,targets_all,batch_size)
#This tells to tensorflow which input goes to which placeholder
feed_dict={inputs_placeholder: inputs_batch, 
           targets_placeholder: targets_batch}
#This runs one step of the training
sess.run(train_step,feed_dict = feed_dict)

希望我能帮助...

请检查 tf.train.batch的用法:

label_batch = tf.train.batch([label], capacity=20, batch_size=10, num_threads=2)

label的支架的消失将导致这样的错误。