app.py

import math
import os
import sys
import tempfile

import matplotlib.pyplot as plt
import numpy as np
import streamlit as st
from PIL import Image
from tensorflow.keras import layers, models

# get default figure dimensions for matplotlib
fig_size = plt.rcParams['figure.figsize']

# set default color map to grayscale
plt.rcParams['image.cmap'] = 'gray'

# title for the app
st.title("Filters and Feature Maps Visualization")

# file uploader widget to upload h5 files
model = st.file_uploader(label="Upload model", type=["h5"])

if model:
    # start a temporary directory
    with tempfile.TemporaryDirectory() as tempdir:
        with open(os.path.join(tempdir, "temp.h5"), mode='wb') as f:
            # write the model to a temporary file in a temporary folder
            f.write(model.getvalue())

        # load the model into a `model` variable
        model = models.load_model(os.path.join(tempdir, "temp.h5"))

    # dropdown menu to visualize filters or maps
    viz_option = st.selectbox("What would you like to visualize?",
                              options=["Filters", "Feature Maps"])

    # get indices of conv layers in model
    conv_indices = [i for i in range(len(model.layers)) if isinstance(
        model.layers[i], layers.Conv2D)]

    if viz_option.lower() == "filters":
        # filter visualization

        # dropdown menu to select a convolutional layer
        layer_index = st.selectbox(
            "Select a layer to see its filters", options=conv_indices)

        weights = model.layers[layer_index].get_weights()[0]
        num_filters = weights.shape[-1]
        num_channels = weights.shape[-2]

        st.write(
            f"This layer has {num_filters} filters and {num_channels} channels per filter.")

        channel_index = st.selectbox(
            "Which channel would you like to view?", options=range(1, num_channels + 1))

        # make subplots
        nrows = math.ceil(math.sqrt(num_filters))
        ncols = math.ceil(math.sqrt(num_filters))
        fig, ax = plt.subplots(nrows, ncols, figsize=(
            fig_size[0] * ncols, fig_size[1] * nrows))

        # remove excess subplots
        for i in range(num_filters - (nrows * ncols), 0):
            ax.flatten()[i].remove()

        # plot filters
        for i in range(num_filters):
            ax.flatten()[i].imshow(weights[:, :, channel_index - 1, i])
            ax.flatten()[i].set(xticklabels=[],
                                yticklabels=[], title=f"Filter {i + 1}")

        fig.tight_layout()
        # show plot
        st.pyplot(fig)
    else:
        # feature map visualization

        # upload image
        img = st.file_uploader(label="Upload image", type=['jpg', 'png'])

        if img:
            # read image
            img = np.asarray(Image.open(img))
            st.image(img)

            # adjust shape of image
            img = np.expand_dims(np.expand_dims(img, axis=-1), axis=0)

            # choose layer
            layer_index = st.selectbox(
                "Feature Map at which layer?", options=conv_indices)

            # create temp model
            temp_model = models.Model(
                inputs=model.inputs, outputs=model.layers[layer_index].output)
            output = np.squeeze(temp_model.predict(img))

            num_channels = output.shape[-1]

            nrows = math.ceil(math.sqrt(num_channels))
            ncols = math.ceil(math.sqrt(num_channels))
            fig, ax = plt.subplots(nrows, ncols, figsize=(
                fig_size[0] * ncols, fig_size[1] * nrows))

            # remove excess subplots
            for i in range(num_channels - (nrows * ncols), 0):
                ax.flatten()[i].remove()

            # plot filters
            for i in range(num_channels):
                ax.flatten()[i].imshow(output[:, :, i])
                ax.flatten()[i].set(xticklabels=[],
                                    yticklabels=[], title=f"Channel {i + 1}")

            fig.tight_layout()
            # show plot
            st.pyplot(fig)