CODEWITHSUNDEEP
pythontensorflowmachine-learningkerasconv-neural-network
watchtower
watchtower

Reputation: 4298

tf.keras.layers.Conv2D usage

I am new to CNN and tensorflow. It has been about 2 days since working through Geron's Hands-on TF book. I'd appreciate if someone can help me.

Goal: Understand how and why keras official documentation (https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv2D) uses Conv2D (m,n...) notation.

Problem: I wrote two sets of code. I understood the first one, which uses explicit set of declaration for number of filters and kernel_size.

TF and Python Version:

sys.version
Out: '3.7.9 (default, Aug 31 2020, 17:10:11) [MSC v.1916 64 bit (AMD64)]'
tf.__version__
Out: '2.3.0'

Code 1:

import tensorflow as tf

input_shape = (4, 30, 60, 3) #Sample 30x60 images with RGB channel. `batch_size` = 4

a1=tf.keras.layers.Conv2D(filters=10,kernel_size=(3,3), input_shape=input_shape[1:])
a1(tf.random.normal(input_shape)).shape
a1.filters
a1.kernel_size

model = tf.keras.Sequential()
model.add(a1)
model.output_shape
model.summary()

Output:

Out[99]: TensorShape([4, 28, 58, 10]) #we are not using padding. So, the shape of tensor is 4 batch x 28x58 x 10 filters
Out[99]: 10 # number of filters
Out[99]: (3, 3) #kernel size 
Out[99]: (None, 28, 58, 10) #this is the feature map for one image
Model: "sequential_24"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_34 (Conv2D)           (None, 28, 58, 10)        280       
=================================================================
Total params: 280
Trainable params: 280
Non-trainable params: 0
_________________________________________________________________

I am good with above output. I've added my thoughts above.

Code 2: Now, I modified the code above to not explicitly mention kernel_size and filters as per official documentation above.

a2=tf.keras.layers.Conv2D(10,3,3, input_shape=input_shape[1:]) #here's the change.
a2(tf.random.normal(input_shape)).shape
a2.filters
a2.kernel_size

model = tf.keras.Sequential()
model.add(a2)
model.output_shape
model.summary()

Output:

Out[100]: TensorShape([4, 10, 20, 10])
Out[100]: 10
Out[100]: (3, 3)
Out[100]: (None, 10, 20, 10)
Model: "sequential_25"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_35 (Conv2D)           (None, 10, 20, 10)        280       
=================================================================
Total params: 280
Trainable params: 280
Non-trainable params: 0
_________________________________________________________________

As we can see, the only difference is that Code 1 uses Conv2D(filters=10,kernel_size=(3,3),... while Code 2 uses Conv2D(10,3,3,.... Moreover filters and kernel_size are also the same. However, the output_shape are completely different.

Why is this? Can someone please explain? I couldn't find anything on keras official documentation (https://www.tensorflow.org/api_docs/python/tf/keras/layers/Conv2D).

Upvotes: 0

Views: 1151

Answers (1)

Davide Anghileri
Davide Anghileri

Reputation: 931

When using tf.keras.layers.Conv2D() you should pass the second parameter (kernel_size) as a tuple (3, 3) otherwise your are assigning the second parameter, kernel_size=3 and then the third parameter which is stride=3. By using a stride of 3 you see an input_shape which is 1/3 of the original inputh shape, rounded to the nearest integer.

So, change your code to:

a2=tf.keras.layers.Conv2D(10, (3,3), input_shape=input_shape[1:])

Upvotes: 2

Related Questions