dheemanthbhat / tensorflow-image-pipeline

TensorFlow Image Pipeline

StackOverflow Question: Why shape of 3D tensor (of an image) in filter() has None?

I am following this tutorial on tensorflow.org.

I have folder images with two folders cat and dog in it. Following above tutorial I am trying to convert .jpg and .png images to features (NumPy array) for modeling.

Problem

After processing the images to tensors I found that some images were converted to tensor with shape (28, 28, 4). So I added condition to filter out such tensors. This logic works while explicitly looping each tensor, using for loop, after converting it to numpy array, but same logic does not work when used with filter.

Please help me fix this filter() I went through filter() documentation and could not find any solution.

Source code

import tensorflow as tf
import os

print("TensorFlow version:", tf.__version__)

def process_image(file_path_tensor):
    parts = tf.strings.split(file_path_tensor, os.sep)
    label = parts[-2]

    image = tf.io.read_file(file_path_tensor)
    image = tf.image.decode_jpeg(image)
    image = tf.image.resize(image, [128, 128])
    image = tf.image.convert_image_dtype(image, tf.float32)
    image = image / 255

    return image, label


def check_shape(x, y):
    print("\nShape received in filter():", x.shape)
    d1, d2, d3 = x.shape
    return d3 == 3


images_ds = tf.data.Dataset.list_files("./images/*/*", shuffle=True)

file_path = next(iter(images_ds))
image, label = process_image(file_path)

print("Shape:", image.shape)
print("Class label:", label.numpy().decode())

# ETL pipeline.
X_y_tensors = (
    images_ds
    .map(process_image)   # Extra and Transform
    .filter(check_shape)  # Filter
    .as_numpy_iterator()  # Load
)

print("\nTechnique 1:")
print("Final X count:", len(list(X_y_tensors)))


X_y_tensors = images_ds.map(process_image)

count = 0
for x, y in X_y_tensors:
    d1, d2, d3 = x.shape
    if d3 > 3:
        continue
    count += 1

print("\nTechnique 1:")
print("Final X count:", count)

Output

TensorFlow version: 2.6.0
Shape: (128, 128, 3)
Class label: cat

Shape received in filter(): (128, 128, None)

Technique 1:
Final X count: 0

Technique 1:
Final X count: 123

As it can be seen,

1. Count is 0 when Technique 1 is used to filter tensors, since the shape of the tensor received is (128, 128, None).
2. Count is 123 (image count after filtering) when Technique 2 is used.

I do not think this is an issue since I am not using batches.

Full code with dataset