Conv2D layers in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output shape of Conv2D layer

Given the following Conv2D layer in TensorFlow, what is the output shape when the input shape is (32, 32, 3)?

TensorFlow

import tensorflow as tf
layer = tf.keras.layers.Conv2D(filters=16, kernel_size=3, strides=1, padding='valid')
input_shape = (1, 32, 32, 3)
output = layer(tf.random.uniform(input_shape))
print(output.shape)

A (1, 30, 30, 16)

B (1, 32, 32, 16)

C (1, 31, 31, 16)

D (1, 29, 29, 16)

💡 Hint

Remember that 'valid' padding means no padding is added, so output size shrinks by kernel_size - 1.

With 'valid' padding and kernel size 3, the output height and width are input_size - kernel_size + 1 = 32 - 3 + 1 = 30. The number of filters is 16, so output shape is (1, 30, 30, 16).

❓ Model Choice

intermediate

Choosing padding for same output size

You want a Conv2D layer that keeps the input height and width unchanged after convolution. Which padding option should you use?

A 'same'

B 'valid'

C 'causal'

D 'full'

💡 Hint

Padding adds pixels around the input to keep output size same as input.

'same' padding adds zeros around the input so the output height and width remain the same as input when stride is 1.

❓ Hyperparameter

advanced

Effect of stride on output size

If you apply a Conv2D layer with kernel_size=5, padding='valid', and stride=2 on an input of shape (64, 64, 3), what will be the output height and width?

A 31 x 31

B 30 x 30

C 32 x 32

D 29 x 29

💡 Hint

Output size formula: floor((input_size - kernel_size)/stride) + 1

Output size = floor((64 - 5)/2) + 1 = floor(59/2) + 1 = 29 + 1 = 30. The output height and width are 30 x 30.

❓ Metrics

advanced

Calculating number of parameters in Conv2D

How many trainable parameters does a Conv2D layer have with 10 filters, kernel size 3x3, input channels 3, and use_bias=True?

A 300

B 290

C 270

D 280

💡 Hint

Parameters = (kernel_height * kernel_width * input_channels * filters) + filters (if bias used).

Parameters = (3*3*3*10) + 10 = 270 + 10 = 280.

🔧 Debug

expert

Identifying error in Conv2D input shape

What error will this code raise? import tensorflow as tf layer = tf.keras.layers.Conv2D(8, 3) input_data = tf.random.uniform((32, 3, 32, 3)) output = layer(input_data)

A No error, runs successfully

B ValueError: Input rank must be 4, got 4 with wrong dimension order

C TypeError: Conv2D expects input shape (batch, height, width, channels)

D ValueError: Input tensor has wrong shape, expected channels last

💡 Hint

TensorFlow Conv2D expects input shape with channels last by default.

The input shape (32, 3, 32, 3) is invalid for the default 'channels_last' data format because the height and width dimensions are swapped. TensorFlow expects input shape (batch, height, width, channels). This will raise a TypeError indicating the expected input shape.