AttributeError: Tensor.op is meaningless when eager execution is enabled

Question

I am trying to implement RESNET 50 from scratch. After accumulating all the layers, I call tf.keras.Model. However, it gives an error:

AttributeError: Tensor.op is meaningless when eager execution is enabled.

For testing, I am inputting a 4-D tensor. conv_diff_size and conv_same_size are two custom blocks having con2d and batch-normalization layers. I am using TensorFlow 2.0 on Google Colab.

def ResNet50(inputs, classes):
  X = tf.keras.layers.Conv2D(64, kernel_size = (7,7), strides=2, padding='valid', data_format='channels_last', input_shape = inputs.shape)(inputs)
  X = tf.keras.layers.BatchNormalization(axis=-1, momentum=0.9)(X)
  X = tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=2)(X)

  X = conv_diff_size(X, [64, 64, 256])
  X = conv_same_size(X, [64, 64, 256])
  X = conv_same_size(X, [64, 64, 256])
  
  X = conv_diff_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])
  
  X = conv_diff_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  
  X = conv_diff_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
              
  X = tf.keras.layers.AveragePooling2D(pool_size=(2, 2), name = 'avg_pool')(X)
  X = tf.keras.layers.Flatten()(X)
  X = tf.keras.layers.Dense(classes, activation='relu')(X)
  
  model = tf.keras.Model(inputs=X, outputs = X)
  return model

Answer 1

the problem in your code is you are giving X as input as well as output.

Try this

import tensorflow as tf

def ResNet50(input_shape, classes):
  inputs = tf.keras.Input(shape=input_shape)#input_shape = (224,224,3)
  X = tf.keras.layers.Conv2D(64, kernel_size = (7,7), strides=2, padding='valid', data_format='channels_last', input_shape = inputs.shape)(inputs)
  X = tf.keras.layers.BatchNormalization(axis=-1, momentum=0.9)(X)
  X = tf.keras.layers.MaxPool2D(pool_size=(3, 3), strides=2)(X)

  X = conv_diff_size(X, [64, 64, 256])
  X = conv_same_size(X, [64, 64, 256])
  X = conv_same_size(X, [64, 64, 256])

  X = conv_diff_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])
  X = conv_same_size(X, [128, 128, 512])

  X = conv_diff_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])
  X = conv_same_size(X, [256, 256, 1024])

  X = conv_diff_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
  X = conv_same_size(X, [512, 512, 2048])
            
  X = tf.keras.layers.AveragePooling2D(pool_size=(2, 2), name = 'avg_pool')(X)
  X = tf.keras.layers.Flatten()(X)
  out = tf.keras.layers.Dense(classes, activation='relu')(X)

  model = tf.keras.Model(inputs=inputs, outputs = out)
  return model

I have added a input tensor layer and assigned it to variable inputs and the final layer to variable out