如何在 Tensorflow Data API 中批量处理不同形状的张量？

Question

在学习了Research paper 之后，我正在尝试从头开始实现一个全卷积神经网络。由于 FCN 适用于任何大小的输入，因此我很难为它实现 DataLoader。我正在尝试使用 batch() 函数对输入进行批处理，但由于可变长度张量而出现错误。有没有可行的方法来批量处理任意形状的张量？

我的代码：

import tensorflow as tf
import matplotlib.pyplot as plt
from PIL import Image
import numpy as np
import os
from albumentations import ElasticTransform,VerticalFlip,RandomSizedCrop,Compose,RandomRotate90

# Train and Val contains the respective paths for [TrainImage,Labels] and [ValImage,Labels] respectively.

num_classes = 21
Train = tf.data.Dataset.from_tensor_slices(Train)
Val = tf.data.Dataset.from_tensor_slices(Val)


def Create_Mask(Img):
  Seg_Labels = np.zeros((Img.shape[0],Img.shape[1],num_classes),dtype=np.float32)
  for class_ in range(num_classes):
    Seg_Labels[:,:,class_] = (Img == class_)
  return tf.cast(Seg_Labels,dtype=tf.float32)


def Create_PreProcess_Mask_Img(Instance):
  Img = np.asarray(Image.open(Instance[0].numpy()))
  Mask = np.asarray(Image.open(Instance[1].numpy()))    
  # Since Mask is in 'P' mode it will automatically convert to labels using Color Palette 

  Normalization = tf.keras.layers.experimental.preprocessing.Rescaling(1./255)

  if tf.random.uniform(()) > 0.5:  # Applying data Augmentation
    Num = tf.random.uniform((),minval=0,maxval=3,dtype=tf.int32)
    if Num == 0:
      aug = ElasticTransform(p=1)
      Augmented = aug(image = Img,mask = Mask)

    elif Num == 1:
      Height,Width = Img.shape[0],Img.shape[1]
      aug = RandomSizedCrop(min_max_height=(50,101),height=Height,width=Width,p = 1)
      Augmented = aug(image = Img,mask = Mask)

    elif Num == 2:
      aug = Compose([VerticalFlip(p=0.5),RandomRotate90(p=0.5)])
      Augmented = aug(image = Img,mask = Mask)

    Img = Augmented["image"]
    Mask = Augmented["mask"]
  
  return Normalization(Img),Create_Mask(Mask)


def Preprocess(Instance):
  Img,Mask = tf.py_function(Create_PreProcess_Mask_Img,[Instance],[tf.float32,tf.float32])
  return tf.ensure_shape(Img,[None,None,3]),tf.ensure_shape(Mask,[None,None,num_classes])  
  #tf.ensure_shape returns the matrix if shape matches else error

def DataLoader(dataset,BATCH_SIZE = 32,BUFFER_SIZE = 256):
  data = dataset.map(Preprocess,num_parallel_calls = tf.data.AUTOTUNE)
  data = data.cache().shuffle(BUFFER_SIZE).batch(BATCH_SIZE).repeat(1)
  data = data.prefetch(buffer_size = tf.data.AUTOTUNE)
  return data

Train = DataLoader(Train)
Val = DataLoader(Val)

for X,Y in Train.take(1):
  print(X.shape)
  print(Y.shape)

堆栈跟踪：

---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/eager/context.py in execution_mode(mode)
   2112       ctx.executor = executor_new
-> 2113       yield
   2114     finally:

9 frames
InvalidArgumentError: Cannot batch tensors with different shapes in component 0. First element had shape [500,375,3] and element 1 had shape [500,383,3]. [Op:IteratorGetNext]

During handling of the above exception, another exception occurred:

InvalidArgumentError                      Traceback (most recent call last)
/usr/local/lib/python3.7/dist-packages/tensorflow/python/eager/executor.py in wait(self)
     67   def wait(self):
     68     """Waits for ops dispatched in this executor to finish."""
---> 69     pywrap_tfe.TFE_ExecutorWaitForAllPendingNodes(self._handle)
     70 
     71   def clear_error(self):

InvalidArgumentError: Cannot batch tensors with different shapes in component 0. First element had shape [500,375,3] and element 1 had shape [500,383,3].

Answer 1

您不能拥有具有不同图像大小的批次，但如果您想解决该问题，您可以：

• 调整所有图像的大小以形成一个共同的形状（如果图像的比例接近，则影响最小）look at this link。
• 将批次的大小减少到一张图像（这种方法会使训练非常不稳定）。

你可以去this post看看