Week 1 Lab 2 - Using tf.squeeze when transforming the image data

Course Q&A Machine Learning Engineering for Production(MLOps) Machine Learning Modeling Pipelines in Production
1

Hello there,

While reviewing the code for the transformation of the image data, the following code snippet caught my attention, i.e.

def preprocessing_fn(inputs):
    [...]
    # Convert the raw image and labels to a float array
    with tf.device("/cpu:0"):
        outputs = {
            _transformed_name(_IMAGE_KEY):
                tf.map_fn(
                    _image_parser,
                    **tf.squeeze(inputs[_IMAGE_KEY], axis=1)**,
                    dtype=tf.float32),

I’d like to know what the reason for squeezing the 1st tensor axis is for this particular type of data (i.e. binary images)?

Thanks,
Vlad

2

Image is not binary. Each pixel value is in the range [0, 255] before transformation.
Data is read in batches. We need the underlying element and hence the need for squeeze.
Consider a sample record found in the training dataset:

{'features': {'feature': {'label': {'int64List': {'value': ['2']}}, 'image': {'bytesList': {'value': ['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']}}}}}

This is how a single image is decoded if done manually:

import base64
png = base64.b64decode('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')
print(tf.image.decode_image(png, channels=1))