时间戳

首页 > 解决方案 > TensorFlow 单层网络 Conv2d；无效参数：转置需要大小为 3 的向量。但 input(1) 是大小为 4 的向量

python - TensorFlow 单层网络 Conv2d；无效参数：转置需要大小为 3 的向量。但 input(1) 是大小为 4 的向量

问题描述

我无法弄清楚尺寸不匹配的位置。似乎所有图层都是正确的形状。Conv2d 需要一个大小为 4 的向量。

#Here is the network structure 

# inside parse_Depth_input I print the image shape (not sure why it's backwards but I did try reversing just in case)

d_image  (228, 304, 1)

   Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 304, 228, 3)       30        
_________________________________________________________________
reshape (Reshape)            (None, 69312, 3)          0         
=================================================================
Total params: 30
Trainable params: 30
Non-trainable params: 0

(0) 无效参数：transpose 需要一个大小为 3 的向量。但 input(1) 是一个大小为 4 的向量 [[{{node gradient_tape/simple_cnn/sequential/conv2d/Conv2D/Conv2DBackpropFilter-0-TransposeNHWCToNCHW-LayoutOptimizer}}] ] [[cond/then/_25/cond/cond/then/_124/cond/cond/remove_squeezable_dimensions/Equal/_110]] （1）无效参数：转置需要一个大小为3的向量。但输入（1）是一个向量大小为 4 [[{{node gradient_tape/simple_cnn/sequential/conv2d/Conv2D/Conv2DBackpropFilter-0-TransposeNHWCToNCHW-LayoutOptimizer}}]]

import tensorflow as tf
import numpy as np
from sklearn.model_selection import train_test_split
from PIL import Image
from tensorflow.keras import datasets, layers, models, preprocessing
import os
from natsort import natsorted
from tensorflow.keras.models import Model

BATCH_SIZE = 32
EPOCHS = 15
LEARNING_RATE = 1e-4

#jpegs with values from 0 to 255
img_dir = ".../normalized_imgs"
# .npy files of size (69312,3)
pts_dir = ".../normalized_pts"
img_files = [os.path.join(img_dir, f)
               for f in natsorted(os.listdir(img_dir))]

pts_files = [os.path.join(pts_dir, f)
            for f in natsorted(os.listdir(pts_dir))]

img = Image.open(img_files[0])
pts = np.load(pts_files[0])

def parse_img_input(img_file, pts_file):
        def _parse_input(img_file, pts_file):
                # get image
                d_filepath = img_file.numpy().decode()
                d_image_decoded = tf.image.decode_jpeg(tf.io.read_file(d_filepath), channels=1)
                d_image = tf.cast(d_image_decoded, tf.float32) / 255.0
    
                # get numpy data
                pts_filepath = pts_file.numpy().decode()
                pts = np.load(pts_filepath, allow_pickle= True)

                print("d_image ",d_image.shape )
                return d_image, pts
        return tf.py_function(_parse_input,
                              inp=[img_file, pts_file],
                              Tout=[tf.float32, tf.float32])

class SimpleCNN(Model):
        def __init__(self):
                super(SimpleCNN, self).__init__()
                input_shape = (img.size[0], img.size[1], 1)
                self.model = model = models.Sequential()
                model.add(tf.keras.Input(shape= input_shape))
                model.add(layers.Conv2D(3, (3,3), padding='same'))
                model.add(layers.Reshape((pts.shape[0], pts.shape[1])))


# split input data into train, test sets
X_train_file, X_test_file, y_train_file, y_test_file = train_test_split(img_files, pts_files, 
                                                    test_size=0.2, 
                                                    random_state=0)

model = SimpleCNN()

dataset_train = tf.data.Dataset.from_tensor_slices((X_train_file, y_train_file))
dataset_train = dataset_train.map(parse_img_input)

dataset_test = tf.data.Dataset.from_tensor_slices((X_test_file, y_test_file))
dataset_test = dataset_test.map(parse_img_input)

model.compile(optimizer=tf.keras.optimizers.Adam(LEARNING_RATE), loss= tf.losses.MeanSquaredError(), metrics= [tf.keras.metrics.get('accuracy')])
model.fit(dataset_train, epochs=EPOCHS, shuffle=True, validation_data= dataset_test)

奇怪的是，如果我尝试将形状重塑为与输入相同的形状，尽管它正确显示了输入大小，但仍会出现错误

            input_shape = (img.size[0], img.size[1], 1)
            model.add(tf.keras.Input(shape= input_shape))
            model.add(layers.Reshape((input_shape[0], input_shape[1], 1)))

reshape 的输入是一个有 69312 个值的张量，但请求的形状有 15803136

标签： pythontensorflowconv-neural-network

解决方案

我的问题是我没有事先对数据集进行批处理。我添加了这个

def prepare_dataset(ds, shuffle = False):
  if shuffle:
    ds = ds.shuffle(buffer_size=500)
  ds = ds.batch(BATCH_SIZE)
  ds = ds.prefetch(buffer_size=AUTOTUNE)
  return ds

dataset_train = prepare_dataset(dataset_train, shuffle=True)

推荐阅读