Convert a preprocessing function into a convolutional layer (in Keras). It takes an input from -127 to 127 and runs it through the preprocess_input function and then returns a convolutional layer with the appropriate weights and biases to reproduce (as closely as possible) the preprocessing step. This allows models to be packaged without worryin…

prep_convert.py

from keras.applications.mobilenet_v2 import MobileNetV2, preprocess_input
from keras import layers, models
from sklearn.linear_model import LinearRegression
import numpy as np

def prep_to_conv(
    in_prep_func: Callable[[np.array], np.array],
    *,
    min_val: float=-127,
    max_val: float=127,
    channels: int=3,
    verbose: bool=False
) -> layers.Layer:
    """Function to turn a preprocessing step into a convolutional layer"""
    test_channel = np.linspace(min_val, max_val, 9).reshape((1, 3, 3)).astype('float32')
    test_input = np.stack([test_channel]*3, axis=-1)
    test_output = in_prep_func(test_input.copy())
    W = np.zeros((1, 1, channels, channels), dtype='float32')
    b = np.zeros((channels,), dtype='float32')
    for i in range(channels):
        x = test_input[:, :, :, i].ravel().reshape((-1, 1))
        y = test_output[:, :, :, i].ravel().reshape((-1, 1))
        lin_reg = LinearRegression()
        lin_reg.fit(x, y)
        W[0, 0, i, i] = lin_reg.coef_[0]
        b[i] = lin_reg.intercept_
    conv1 = layers.Conv2D(channels, 
                          kernel_size=(1,1), 
                          activation='linear', 
                          use_bias=True, 
                          weights=[W,b],
                          input_shape=(None, None, channels))
    conv1.trainable=False
    if verbose:
        import matplotlib.pyplot as plt
        # check the inputs and outputs
        s = models.Sequential()
        s.add(conv1)
        pred_output = s.predict(test_input)
        fig, m_axs = plt.subplots(1, channels, figsize=(4*channels, 4))
        for i, c_ax in enumerate(m_axs):
            x = test_input[:, :, :, i].ravel().reshape((-1, 1))
            y = test_output[:, :, :, i].ravel().reshape((-1, 1))
            z = pred_output[:, :, :, i].ravel().reshape((-1, 1))
            c_ax.plot(x, y, 's', label='Given')
            c_ax.plot(x, z, '-', label='Predicted')
            c_ax.legend()
    return conv1

prep_to_conv_nl.png