app.py

import os

import mediapipe as mp
import tensorflow as tf

N_ROWS = 543
N_DIMS = 3
DIM_NAMES = ['x', 'y', 'z']
SEED = 42
NUM_CLASSES = 250
INPUT_SIZE = 32


# Tensorflow layer to process data in TFLite
# Data needs to be processed in the model itself, so we cannot use Python
class PreprocessLayer(tf.keras.layers.Layer):
    def __init__(self):
        super(PreprocessLayer, self).__init__()

    def pad_edge(self, t, repeats, side):
        if side == 'LEFT':
            return tf.concat((tf.repeat(t[:1], repeats=repeats, axis=0), t), axis=0)
        elif side == 'RIGHT':
            return tf.concat((t, tf.repeat(t[-1:], repeats=repeats, axis=0)), axis=0)

    @tf.function(
        input_signature=(tf.TensorSpec(shape=[None, N_ROWS, N_DIMS], dtype=tf.float32),),
    )
    def call(self, data0):
        # Number of Frames in Video
        N_FRAMES0 = tf.shape(data0)[0]

        # Filter Out Frames With Empty Hand Data
        frames_hands_nansum = tf.experimental.numpy.nanmean(tf.gather(data0, HAND_IDXS0, axis=1), axis=[1, 2])
        non_empty_frames_idxs = tf.where(frames_hands_nansum > 0)
        non_empty_frames_idxs = tf.squeeze(non_empty_frames_idxs, axis=1)
        data = tf.gather(data0, non_empty_frames_idxs, axis=0)

        # Cast Indices in float32 to be compatible with Tensorflow Lite
        non_empty_frames_idxs = tf.cast(non_empty_frames_idxs, tf.float32)

        # Number of Frames in Filtered Video
        N_FRAMES = tf.shape(data)[0]

        # Gather Relevant Landmark Columns
        data = tf.gather(data, LANDMARK_IDXS0, axis=1)

        # Video fits in INPUT_SIZE
        if N_FRAMES < INPUT_SIZE:
            # Pad With -1 to indicate padding
            non_empty_frames_idxs = tf.pad(non_empty_frames_idxs, [[0, INPUT_SIZE - N_FRAMES]], constant_values=-1)
            # Pad Data With Zeros
            data = tf.pad(data, [[0, INPUT_SIZE - N_FRAMES], [0, 0], [0, 0]], constant_values=0)
            # Fill NaN Values With 0
            data = tf.where(tf.math.is_nan(data), 0.0, data)
            return data, non_empty_frames_idxs
        # Video needs to be downsampled to INPUT_SIZE
        else:
            # Repeat
            if N_FRAMES < INPUT_SIZE ** 2:
                repeats = tf.math.floordiv(INPUT_SIZE * INPUT_SIZE, N_FRAMES0)
                data = tf.repeat(data, repeats=repeats, axis=0)
                non_empty_frames_idxs = tf.repeat(non_empty_frames_idxs, repeats=repeats, axis=0)

            # Pad To Multiple Of Input Size
            pool_size = tf.math.floordiv(len(data), INPUT_SIZE)
            if tf.math.mod(len(data), INPUT_SIZE) > 0:
                pool_size += 1

            if pool_size == 1:
                pad_size = (pool_size * INPUT_SIZE) - len(data)
            else:
                pad_size = (pool_size * INPUT_SIZE) % len(data)

            # Pad Start/End with Start/End value
            pad_left = tf.math.floordiv(pad_size, 2) + tf.math.floordiv(INPUT_SIZE, 2)
            pad_right = tf.math.floordiv(pad_size, 2) + tf.math.floordiv(INPUT_SIZE, 2)
            if tf.math.mod(pad_size, 2) > 0:
                pad_right += 1

            # Pad By Concatenating Left/Right Edge Values
            data = self.pad_edge(data, pad_left, 'LEFT')
            data = self.pad_edge(data, pad_right, 'RIGHT')

            # Pad Non Empty Frame Indices
            non_empty_frames_idxs = self.pad_edge(non_empty_frames_idxs, pad_left, 'LEFT')
            non_empty_frames_idxs = self.pad_edge(non_empty_frames_idxs, pad_right, 'RIGHT')

            # Reshape to Mean Pool
            data = tf.reshape(data, [INPUT_SIZE, -1, N_COLS, N_DIMS])
            non_empty_frames_idxs = tf.reshape(non_empty_frames_idxs, [INPUT_SIZE, -1])

            # Mean Pool
            data = tf.experimental.numpy.nanmean(data, axis=1)
            non_empty_frames_idxs = tf.experimental.numpy.nanmean(non_empty_frames_idxs, axis=1)

            # Fill NaN Values With 0
            data = tf.where(tf.math.is_nan(data), 0.0, data)

            return data, non_empty_frames_idxs


# Get the absolute path to the directory containing app.py
current_dir = os.path.dirname(os.path.abspath(__file__))
# Define the filename of the TFLite model
model_filename = "model.tflite"
# Construct the full path to the TFLite model file
model_path = os.path.join(current_dir, model_filename)
# Load the TFLite model using the interpreter
interpreter = tf.lite.Interpreter(model_path=model_path)
interpreter.allocate_tensors()

# Get input and output details of the TFLite model
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

index_to_class = {
    "TV": 0, "after": 1, "airplane": 2, "all": 3, "alligator": 4, "animal": 5, "another": 6, "any": 7, "apple": 8,
    "arm": 9, "aunt": 10, "awake": 11, "backyard": 12, "bad": 13, "balloon": 14, "bath": 15, "because": 16, "bed": 17,
    "bedroom": 18, "bee": 19, "before": 20, "beside": 21, "better": 22, "bird": 23, "black": 24, "blow": 25, "blue": 26,
    "boat": 27, "book": 28, "boy": 29, "brother": 30, "brown": 31, "bug": 32, "bye": 33, "callonphone": 34, "can": 35,
    "car": 36, "carrot": 37, "cat": 38, "cereal": 39, "chair": 40, "cheek": 41, "child": 42, "chin": 43,
    "chocolate": 44, "clean": 45, "close": 46, "closet": 47, "cloud": 48, "clown": 49, "cow": 50, "cowboy": 51,
    "cry": 52, "cut": 53, "cute": 54, "dad": 55, "dance": 56, "dirty": 57, "dog": 58, "doll": 59, "donkey": 60,
    "down": 61, "drawer": 62, "drink": 63, "drop": 64, "dry": 65, "dryer": 66, "duck": 67, "ear": 68, "elephant": 69,
    "empty": 70, "every": 71, "eye": 72, "face": 73, "fall": 74, "farm": 75, "fast": 76, "feet": 77, "find": 78,
    "fine": 79, "finger": 80, "finish": 81, "fireman": 82, "first": 83, "fish": 84, "flag": 85, "flower": 86,
    "food": 87, "for": 88, "frenchfries": 89, "frog": 90, "garbage": 91, "gift": 92, "giraffe": 93, "girl": 94,
    "give": 95, "glasswindow": 96, "go": 97, "goose": 98, "grandma": 99, "grandpa": 100, "grass": 101, "green": 102,
    "gum": 103, "hair": 104, "happy": 105, "hat": 106, "hate": 107, "have": 108, "haveto": 109, "head": 110,
    "hear": 111, "helicopter": 112, "hello": 113, "hen": 114, "hesheit": 115, "hide": 116, "high": 117, "home": 118,
    "horse": 119, "hot": 120, "hungry": 121, "icecream": 122, "if": 123, "into": 124, "jacket": 125, "jeans": 126,
    "jump": 127, "kiss": 128, "kitty": 129, "lamp": 130, "later": 131, "like": 132, "lion": 133, "lips": 134,
    "listen": 135, "look": 136, "loud": 137, "mad": 138, "make": 139, "man": 140, "many": 141, "milk": 142,
    "minemy": 143, "mitten": 144, "mom": 145, "moon": 146, "morning": 147, "mouse": 148, "mouth": 149, "nap": 150,
    "napkin": 151, "night": 152, "no": 153, "noisy": 154, "nose": 155, "not": 156, "now": 157, "nuts": 158, "old": 159,
    "on": 160, "open": 161, "orange": 162, "outside": 163, "owie": 164, "owl": 165, "pajamas": 166, "pen": 167,
    "pencil": 168, "penny": 169, "person": 170, "pig": 171, "pizza": 172, "please": 173, "police": 174, "pool": 175,
    "potty": 176, "pretend": 177, "pretty": 178, "puppy": 179, "puzzle": 180, "quiet": 181, "radio": 182, "rain": 183,
    "read": 184, "red": 185, "refrigerator": 186, "ride": 187, "room": 188, "sad": 189, "same": 190, "say": 191,
    "scissors": 192, "see": 193, "shhh": 194, "shirt": 195, "shoe": 196, "shower": 197, "sick": 198, "sleep": 199,
    "sleepy": 200, "smile": 201, "snack": 202, "snow": 203, "stairs": 204, "stay": 205, "sticky": 206, "store": 207,
    "story": 208, "stuck": 209, "sun": 210, "table": 211, "talk": 212, "taste": 213, "thankyou": 214, "that": 215,
    "there": 216, "think": 217, "thirsty": 218, "tiger": 219, "time": 220, "tomorrow": 221, "tongue": 222, "tooth": 223,
    "toothbrush": 224, "touch": 225, "toy": 226, "tree": 227, "uncle": 228, "underwear": 229, "up": 230, "vacuum": 231,
    "wait": 232, "wake": 233, "water": 234, "wet": 235, "weus": 236, "where": 237, "white": 238, "who": 239, "why": 240,
    "will": 241, "wolf": 242, "yellow": 243, "yes": 244, "yesterday": 245, "yourself": 246, "yucky": 247, "zebra": 248,
    "zipper": 249
}

inv_index_to_class = {v: k for k, v in index_to_class.items()}

mp_holistic = mp.solutions.holistic


def mediapipe_detection(image, model):
    # COLOR CONVERSION BGR 2 RGB
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image.flags.writeable = False  # Image is no longer writeable
    results = model.process(image)  # Make prediction
    image.flags.writeable = True  # Image is now writeable
    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)  # COLOR COVERSION RGB 2 BGR
    return image, results


def extract_keypoints(results):
    face = np.array([[res.x, res.y, res.z] for res in results.face_landmarks.landmark]).flatten()
    lh = np.array([[res.x, res.y, res.z] for res in results.left_hand_landmarks.landmark]).flatten()
    rh = np.array([[res.x, res.y, res.z] for res in results.right_hand_landmarks.landmark]).flatten()
    pose = np.array([[res.x, res.y, res.z, res.visibility] for res in results.pose_landmarks.landmark]).flatten()

    # Pad or truncate the arrays to the expected length (543)
    face = np.pad(face, (0, max(0, 543 - len(face))), mode='constant')
    lh = np.pad(lh, (0, max(0, 543 - len(lh))), mode='constant')
    rh = np.pad(rh, (0, max(0, 543 - len(rh))), mode='constant')
    pose = np.pad(pose, (0, max(0, 543 - len(pose))), mode='constant')

    # Concatenate the arrays in the correct order and return the result
    return np.concatenate([face, lh, rh, pose])


# Make prediction
def make_prediction(processed_landmarks):
    inputs = np.array(processed_landmarks, dtype=np.float32)

    # Set the input tensor for the TFLite model
    interpreter.set_tensor(input_details[0]['index'], inputs)

    # Invoke the TFLite interpreter to perform inference
    interpreter.invoke()

    # Get the output tensor of the TFLite model
    output_data = interpreter.get_tensor(output_details[0]['index'])

    # Find the index of the predicted class
    index = np.argmax(output_data)

    # Map the index to the corresponding class label using the index_to_class dictionary
    prediction = inv_index_to_class[index]

    return prediction


# ...

with mp_holistic.Holistic(min_detection_confidence=0.5, min_tracking_confidence=0.5) as holistic:
    import cv2
    import numpy as np
    import gradio as gr
    import tensorflow as tf


    # Modify the predict_with_webcam function to take an image as input and return the prediction string
    def predict_with_webcam(frame):
        if frame is None:
            raise ValueError("Frame is None. Make sure your webcam is working properly.")

        # Make detections using mediapipe
        image, results = mediapipe_detection(frame, holistic)
        print(results)

        if results is not None and results.face_landmarks is not None:
            landmarks = extract_keypoints(results)
            if landmarks is not None:
                # Calculate the number of landmarks per frame
                landmarks_per_frame = len(landmarks) // (N_ROWS * N_DIMS)
                # Reshape the landmarks to have shape (None, N_ROWS, N_DIMS)
                landmarks = landmarks.reshape(-1, landmarks_per_frame, N_DIMS)
                # Initialize PreprocessLayer
                preprocess_layer = PreprocessLayer()
                # Call the PreprocessLayer to preprocess the landmarks
                processed_landmarks, _ = preprocess_layer.call(landmarks)
                prediction = make_prediction(processed_landmarks)  # Pass the preprocessed landmarks to make_prediction
                print("Prediction:", prediction)
                return prediction
            else:
                return "Could not detect landmarks or extract keypoints. Make sure your webcam is working properly."
        else:
            return "Could not detect face landmarks. Make sure your webcam is working properly."


    # Define the Gradio interface
    iface = gr.Interface(
        fn=predict_with_webcam,  # The function to use for prediction
        inputs="webcam",  # Use Gradio's "webcam" input to capture frames from the webcam
        outputs=gr.outputs.Textbox()  # Display the prediction as text
    )

    # Launch the interface
    iface.launch()