dino/app.py

import os
os.system('pip install gradio==2.3.0a0')
os.system('pip freeze')
from video_generation import VideoGenerator

import gradio as gr
import shutil
from argparse import Namespace
import subprocess

os.environ['KMP_DUPLICATE_LIB_OK']='True'


def func(resize, video):
    output_dir = '/tmp/outputs'
    input_video = '/tmp/input.mp4'
    if os.path.exists(input_video):
        os.remove(input_video)
    subprocess.call(f"ffmpeg -ss 00:00:00 -i {video} -to 00:00:05 -c copy {input_video}".split())
    if os.path.exists(output_dir):
        shutil.rmtree(output_dir)

    os.makedirs(output_dir)

    args = Namespace(
        arch="vit_small",
        patch_size=8,
        pretrained_weights="dino_deitsmall8_pretrain.pth",
        checkpoint_key="teacher",
        input_path=input_video,
        output_path=output_dir,
        threshold=0.6,
        resize=resize,
        video_only=False,
        fps=30.0,
        video_format="mp4"
    )

    vid_generator = VideoGenerator(args)
    vid_generator.run()

    # Make a video that puts the resized input video + the attn output video together as one
    ffmpeg_cmd = f"""
        ffmpeg
        -i {output_dir}/original-reshaped.mp4
        -i {output_dir}/video.mp4
        -filter_complex hstack
        {output_dir}/stacked.mp4
    """
    subprocess.call(ffmpeg_cmd.split())
    return f'{output_dir}/stacked.mp4'

title = "Interactive demo: DINO"
description = "Demo for Facebook AI's DINO, a new method for self-supervised training of Vision Transformers. Using this method, they are capable of segmenting objects within an image without having ever been trained to do so. This can be observed by displaying the self-attention of the heads from the last layer for the [CLS] token query. This demo uses a ViT-S/8 trained with DINO. To use it, upload a video file. Right now we only suggest using .mp4 files."
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2104.14294'>Emerging Properties in Self-Supervised Vision Transformers</a> | <a href='https://github.com/facebookresearch/dino'>Github Repo</a></p>"
iface = gr.Interface(fn=func, 
                     inputs=[gr.inputs.Slider(120, 480, 20, label="resize"), gr.inputs.Video(label="input video")],
                     outputs='video',
                     title=title,
                     description=description,
                     enable_queue=True,
                     # examples=[[420, 'skate_jump.mp4']],  # Not working for some reason...
                     article=article)

iface.launch(debug=True)