Conv2d tensorflow specification

Question

I'm training a binary classifier for grayscale images 129x107 pixels. Here's the model setup, with Conv2d commented out.

  model = tf.keras.Sequential([
    keras.Input(name="MBInput", shape=[129, 107, 1]),
    keras.layers.Conv2D(3, kernel_size=3, input_shape=(129,107,1)), # Without this line it all works
    keras.layers.Flatten(),
    keras.layers.Dense(100, activation='relu'),
    keras.layers.Dense(1, activation='sigmoid')
    ])
  model.compile(optimizer='adam', loss='binary_crossentropy')

When I include the convolution step tensorflow complains

Input 0 of layer sequential is incompatible with the layer: : expected min_ndim=4, found ndim=3. Full shape received: (None, 129, 107)

I'm probably missing something obvious. What's the right code for 3x3 convolutional filter in this context?

Answer 1

If you specify input_shape in Conv2D layer then there is no need of explicitly add Input layer to the model and don't mix keras and tf.keras

Working code as shown below

import tensorflow as tf

model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(3, kernel_size=3, input_shape=(129,107,1))) 
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(100, activation='relu'))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

model.compile(optimizer='adam', loss='binary_crossentropy')
model.summary()

Output:

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 127, 105, 3)       30        
_________________________________________________________________
flatten (Flatten)            (None, 40005)             0         
_________________________________________________________________
dense (Dense)                (None, 100)               4000600   
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 101       
=================================================================
Total params: 4,000,731
Trainable params: 4,000,731
Non-trainable params: 0
_________________________________________________________________

If you would like to use Input layer, then you can remove input_shape from Conv2D layer

import tensorflow as tf

model = tf.keras.Sequential()
model.add(tf.keras.Input(name="MBInput", shape=[129, 107, 1]))
model.add(tf.keras.layers.Conv2D(3, kernel_size=3)) 
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(100, activation='relu'))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))

model.compile(optimizer='adam', loss='binary_crossentropy')
model.summary()

Output:

Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_1 (Conv2D)            (None, 127, 105, 3)       30        
_________________________________________________________________
flatten_1 (Flatten)          (None, 40005)             0         
_________________________________________________________________
dense_2 (Dense)              (None, 100)               4000600   
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 101       
=================================================================
Total params: 4,000,731
Trainable params: 4,000,731
Non-trainable params: 0
_________________________________________________________________

Here both models are same.