app.py

# -*- coding: utf-8 -*-
"""
Created on Sun Jan 28 18:48:07 2024

@author: liewchooichin
"""
import os
import pathlib
import gradio as gr
import numpy as np
import tensorflow as tf
from huggingface_hub import snapshot_download
from huggingface_hub import from_pretrained_keras

# check the tensoflow version
print(f"tensorflow version: {tf.__version__}")

# global variables
# predictions from:
pred_binary = ""  # binary labels
pred_multi = ""  # multi labels

# sample files
samples = []
labels = []
data_dir = "face_samples"
# local testing
LOCAL_TEST = False  # when in HF, set to False
HF_SPACE = True  # when in HF
# My model in the HF repo
REPO_ID_BINARY = 'liewchooichin/fake_binary'
REPO_ID_MULTILABEL = 'liewchooichin/fake_multilabel'
# tf_model = None
# keras_model = None
local_model_dir = "fake_models"
pb_name = "saved_model.pb"
keras_binary_label = os.path.join("binary_label", "all_binary_6771.keras")
keras_multilabel = os.path.join("multi_label", "multi_7036.keras")


def get_samples():
    samples_path = os.path.join(
        os.path.dirname(__file__),
        data_dir
    )
    samples_path = pathlib.Path(samples_path)
    files = list(samples_path.glob("*.jpg"))
    # hard code the examples first for test
    # first 9 are fake, the last 3 are real
    # fake faces
    for i in range(9):
        samples.append(files[i])
        # get the fake or real label
        fake = 1
        labels.append(fake)
    # real faces
    for i in range(3):
        samples.append(files[i+9])
        fake = 0
        labels.append(fake)
    # print to check the image and labels
    for i in range(12):
        print(samples[i], labels[i])


def download_keras_model():
    # set the model variables to be global
    global keras_binary_model
    global keras_multi_model

    # HF repo
    # load binary label
    if HF_SPACE:
        download_dir = snapshot_download(repo_id=REPO_ID_BINARY)
        print(f"Download dir: {download_dir}")
        keras_binary_path = os.path.join(download_dir, keras_binary_label)
        print(f"Keras binary label: {keras_binary_path}")
        # this load() does not work in HF
        #keras_binary_model = tf.keras.models.load(keras_binary_path)
        #keras_binary_model = tf.keras.saving.load_model(keras_binary_path)       
        #keras_binary_model = from_pretrained_keras("liewchooichin/fake_binary")
        keras_binary_model = tf.saved_model.load(download_dir)

    # local testing
    # check if the model exists
    # binary label
    # "C:\PY\exercises\hello_iris\fake_models\binary_label\all_binary_6771.keras"
    if LOCAL_TEST:
        model_path = os.path.join(
            os.path.dirname(__file__),
            local_model_dir,
            keras_binary_label
        )
        if not os.path.exists(model_path):
            print(f"Model not found: {model_path}")
        # load local keras model
        keras_binary_model = tf.keras.models.load_model(model_path)

    # Check with model loaded
    #print(f"\nBinary label model: {keras_binary_model.name}")

    # load multilabel
    # "C:\PY\exercises\hello_iris\fake_models\multi_label\all_multi_7036.keras"
    if LOCAL_TEST:
        model_path = os.path.join(
            os.path.dirname(__file__),
            local_model_dir,
            keras_multilabel
        )
        if not os.path.exists(model_path):
            print(f"Model not found: {model_path}")
        # load local keras model
        keras_multi_model = tf.keras.models.load_model(model_path)

    # In HF space, load model from repository
    # Load the multilabel model
    if HF_SPACE:
        # HF repo
        download_dir = snapshot_download(repo_id=REPO_ID_MULTILABEL)
        print(f"Download dir: {download_dir}")
        keras_multi_path = os.path.join(download_dir, keras_multilabel)
        print(f"Keras multi label: {keras_multi_path}")
        # load() does not work in HF
        #keras_multi_model = tf.keras.models.load(keras_multi_path)
        #keras_multi_model = tf.keras.saving.load_model(keras_multi_path)
        #keras_multi_model = from_pretrained_keras("liewchooichin/fake_multilabel")
        keras_multi_model = tf.saved_model.load(download_dir)

    # Check with model loaded
    #print(f"\nLoaded model: {keras_multi_model.name}")


def get_img_array(img_path):
    # get the dataset into array of 224x224
    img = tf.keras.utils.load_img(
        img_path,
        target_size=(224, 224)
    )
    img_array = tf.keras.utils.img_to_array(img)
    # expand the dimension for prediction
    img_array = np.expand_dims(img_array, axis=0)
    print(f"Shape of image array: {img_array.shape}")
    return img_array


def get_prediction(img_path):
    # adjust threshold for accuracy
    threshold = 0.4
    
    # check the image path
    print(f"Image path: {img_path}")
    # also display the original filename for info
    orig_filename = img_path.split("\\")[-1]
    get_img_array(img_path)
    # get the image array
    img_array = get_img_array(img_path)

    # test with local model
    # binary label
    pred_binary = keras_binary_model(img_array, training=False)
    print(f"Keras binary label: {pred_binary}")
    if pred_binary[0][0] > threshold:
        fake = "Fake"
    else:
        fake = "Real"

    # multi label
    pred_multi = keras_multi_model(img_array, training=False)
    print(f"Keras multi label: {pred_multi}")
    # Cut at the sigmoid 0.5 threshold
    fake_parts = np.where(pred_multi > threshold, 1, 0)
    print(f"Multi label: {fake_parts}")
    # Format each of the fake face parts
    parts_message = dict()
    # The last one is the overall prediction
    parts_message["overall"] = "Fake" if fake_parts[0][4] == 1 else "Real"
    parts_message["left_eye"] = "Fake" if fake_parts[0][0] == 1 else "Real"
    parts_message["right_eye"] = "Fake" if fake_parts[0][1] == 1 else "Real"
    parts_message["nose"] = "Fake" if fake_parts[0][2] == 1 else "Real"
    parts_message["mouth"] = "Fake" if fake_parts[0][3] == 1 else "Real"
    # Format the display line by line
    parts_formatted = ""
    for k, v in parts_message.items():
        parts_formatted = parts_formatted + f"{k}: {v}\n"

    # Format result string
    result_binary = f"Probability: {pred_binary} \
                \nPrediction: {fake}\n"
    result_multi = f"Probability: {pred_multi} \
                   \nPrediction: {fake_parts} \
                   \n{parts_formatted}"
    # pred_multi = tf_model(img_path)
    # print(f"tf: \n{pred_multi}")

    return orig_filename, result_binary, result_multi


def clear_image():
    # Clear the previous output result
    return "", "", ""


def main():
    get_samples()
    # download_tf_model()
    download_keras_model()


with gr.Blocks() as demo:

    # call the main for preliminary work
    main()

    image_width = 256
    image_height = 256

    gr.Markdown(
        """
        # Fake or real faces detection.

        The dataset is obtained from https://www.kaggle.com/datasets/ciplab/real-and-fake-face-detection.
        Trained with EfficientNet V2 B0.

        One model is trained to do binary classification and the other \
        multilabel classification. The multilabels classification is \
        based on the last four digits provided by the filenames. \
        The last four digits are following the order of left eye, \
        right eye, nose and mouth. \
        The labels are 1 (fake) and 0 (real).
        For example: ___1010.jpg means left eye and nose are fake.

        Binary accuracy for the binary label model: 0.6771. <br>
        Binary accuracy for the multilabel model: 0.7036.

        The fake faces are also categorized into how difficult it is \
        to detect the faces as fake. The categories are easy, mid and hard.

        The top prediction and its probabilities of classes are shown.

        Try our sample faces below or upload one of your own.
        """
    )
    with gr.Row():
        with gr.Column():
            img = gr.Image(height=image_height,
                           width=image_width,
                           sources=["upload", "clipboard"],
                           interactive=True,
                           type="filepath")

        with gr.Column():
            text_1 = gr.Text(
                label="Filename",
                interactive=False, lines=1
            )
            text_2 = gr.Text(
                label="Binary label, Efficient net v2 B0",
                interactive=False, lines=2)
            text_3 = gr.Text(
                label="Multi label, Efficient net v2 B0",
                interactive=False, lines=7, 
                visible=False)
            """
            text_3 = gr.Text(label="Sashi's model",
                             interactive=False, lines=3)
            text_4 = gr.Text(label="KK's model",
                             interactive=False, lines=3)
            """
        # load the images directory
        # print(f"List of examples: {samples}")

    with gr.Row():
        gr.Markdown("""
                ## Fakes faces <br>(easy)
            """)
        examples_1 = gr.Examples(
            examples=[
                samples[0], samples[1], samples[2],
            ],
            inputs=[img],
            outputs=[text_1, text_2, text_3],
            run_on_click=True,
            fn=get_prediction
        )
    with gr.Row():
        gr.Markdown("""
                ## Fakes faces <br>(mid)
            """)
        examples_2 = gr.Examples(
            examples=[
                samples[3], samples[4], samples[5],
            ],
            inputs=[img],
            outputs=[text_1, text_2, text_3],
            run_on_click=True,
            fn=get_prediction
        )
    with gr.Row():
        gr.Markdown("""
                ## Fakes faces <br>(hard)
            """)
        examples_3 = gr.Examples(
            examples=[
                samples[6], samples[7], samples[8],
            ],
            inputs=[img],
            outputs=[text_1, text_2, text_3],
            run_on_click=True,
            fn=get_prediction
        )
    with gr.Row():
        gr.Markdown("""
            ## Real faces
        """)

        examples_4 = gr.Examples(
            examples=[
                samples[9], samples[10], samples[11]
            ],
            inputs=[img],
            outputs=[text_1, text_2, text_3],
            run_on_click=True,
            fn=get_prediction
        )

    # prediction when a file is uploaded
    img.upload(fn=get_prediction, inputs=[img],
               outputs=[text_1, text_2, text_3])
    # when an example is clicked
    # img.change(fn=get_prediction, inputs=[img],
    #           outputs=[text_1, text_2])
    # when an image is cleared
    img.clear(fn=clear_image, inputs=[],
              outputs=[text_1, text_2, text_3])

if __name__ == "__main__":
    demo.launch()