app.py

import os
os.environ['TF_USE_LEGACY_KERAS'] = '1'

import gradio as gr
import numpy as np
import zipfile
import imageio

import tensorflow as tf
from tensorflow import keras

from utils import read_video, frame_sampling
from utils import num_frames, patch_size, input_size
from labels import K400_label_map, SSv2_label_map

LABEL_MAPS = {
    'K400': K400_label_map,
    'SSv2': SSv2_label_map,
}

ALL_MODELS = [
    'TFVideoSwinT_K400_IN1K_P244_W877_32x224',
    'TFVideoSwinB_SSV2_K400_P244_W1677_32x224',
]

sample_example = [
    ["examples/k400.mp4",  ALL_MODELS[0]],
    ["examples/ssv2.mp4",  ALL_MODELS[1]],
]

def get_model(model_type):
    model_path = keras.utils.get_file(
        origin=f'https://github.com/innat/VideoSwin/releases/download/v1.1/{model_type}.zip',
    )
    with zipfile.ZipFile(model_path, 'r') as zip_ref:
        zip_ref.extractall('./')
    
    model = keras.models.load_model(model_type)

    if 'W877' in model_type:
        data_type = 'K400'
    else:
        data_type = 'SSv2'

    label_map = LABEL_MAPS.get(data_type)
    label_map = {v: k for k, v in label_map.items()}
    
    return model, label_map


def inference(video_file, model_type):
    # get sample data
    container = read_video(video_file)
    frames = frame_sampling(container, num_frames=num_frames)

    # get models
    model, label_map = get_model(model_type)
    model.trainable = False

    # inference on model
    outputs = model(frames[None, ...], training=False)
    probabilities = tf.nn.softmax(outputs).numpy().squeeze(0)
    confidences = {
        label_map[i]: float(probabilities[i]) for i in np.argsort(probabilities)[::-1]
    }
    return confidences


def main():
    iface = gr.Interface(
        fn=inference,
        inputs=[ 
            gr.Video(type="file", label="Input Video"),
            gr.Dropdown(
                choices=ALL_MODELS, 
                default="TFVideoSwinT_K400_IN1K_P244_W877_32x224", 
                label="Model"
            )
        ],
        outputs=gr.Label(num_top_classes=3, label='scores'),
        examples=sample_example,
        title="VideoSwin: Video Swin Transformer",
        description="Keras reimplementation of <a href='https://github.com/innat/VideoSwin'>VideoSwin</a> is presented here."
    )
    iface.launch()

if __name__ == '__main__':
    main()