app.py

import os

os.system('cd TimeSformer;'
          'pip install -e .; cd ..')

os.system('ls -l')
os.system('pwd')

import timesformer


import torch
from torchvision import transforms


from transformers import AutoTokenizer


from PIL import Image
import json 
import os

from torchvision import transforms

from models.epalm import ePALM

import os

from transformers import AutoTokenizer

# import ruamel_yaml as yaml
from ruamel.yaml import YAML

import torch
import gradio as gr


yaml=YAML(typ='safe')



use_cuda = torch.cuda.is_available()
device = torch.deivce('cuda') if use_cuda else torch.deivce('cpu')

## Load model

config = 'configs/image/ePALM_caption.yaml'
# config = yaml.load(open(config, 'r'), Loader=yaml.Loader)
config = yaml.load(open(config, 'r'))


text_model = 'facebook/opt-2.7b' 
vision_model_name = 'vit_base_patch16_224'

# text_model = 'facebook/opt-6.7b' 
# vision_model_name = 'vit_large_patch16_224'


start_layer_idx = 19
end_layer_idx = 31
low_cpu = True 
model = ePALM(opt_model_name=text_model, 
               vision_model_name=vision_model_name, 
               use_vis_prefix=True, 
               start_layer_idx=start_layer_idx, 
               end_layer_idx=end_layer_idx, 
               return_hidden_state_vision=True, 
               config=config,
               low_cpu=low_cpu
)
print("Model Built")
model.to(device)


checkpoint_path = 'checkpoints/float32/ePALM_caption/checkpoint_best.pth'
# checkpoint_path = '/data/mshukor/logs/eplam/models/accelerate/ePALM_pt_L_acc_caption/checkpoint_best.pth'
checkpoint = torch.load(checkpoint_path, map_location='cpu') 
state_dict = checkpoint['model']
msg = model.load_state_dict(state_dict,strict=False)  


## Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(text_model, use_fast=False)
eos_token = tokenizer.eos_token
pad_token = tokenizer.pad_token




image_size = 224
normalize = transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))

transform = transforms.Compose([
            transforms.Resize((image_size,image_size),interpolation=Image.BICUBIC),
            transforms.ToTensor(),
            normalize,
            ])  








do_sample=False
num_beams=3
max_length=30





def inference(image, audio, video, task_type, instruction):

    if task_type == 'Image Captioning':
        text = ['']  
        text_input = tokenizer(text, padding='longest', return_tensors="pt").to(device) 
    else:
        raise NotImplemented

    if "Video" in task_type:
        pass
    elif "Audio" in task_type:
        pass
    else:
        image = transform(image)
        image = image.to(device,non_blocking=True).unsqueeze(0)




    with torch.autocast(device_type='cuda', dtype=torch.float16, enabled=True):

        out = model(image=image, text=text_input, mode='generate', return_dict=True, max_length=max_length, 
                    do_sample=do_sample, num_beams=num_beams)

    out_decode = []
    for i, o in enumerate(out):
        res = tokenizer.decode(o)
        response = res.split('</s>')[1].replace(pad_token, '').replace('</s>', '').replace(eos_token, '') # skip_special_tokens=True

    return response


inputs = [gr.inputs.Image(type='pil'), gr.Audio(source="upload", type="filepath"), gr.Video(source="upload", type="filepath"), gr.inputs.Radio(choices=['Image Captioning', 'Video Captioning', 'Audio Captioning', "Visual Question Answering", "Visual Grounding", "General", "General Video"], type="value", default="Image Captioning", label="Task"), gr.inputs.Textbox(lines=1, label="Instruction")]
outputs = ['text']
examples = [
    ['examples/images/soccer.jpg', None, None, 'Image Captioning', None],
    ['examples/images/ski.jpg', None, None, 'Visual Question Answering', 'what does the woman wearing black do?'],
    ['examples/images/banana.jpg', None, None, 'Visual Grounding', 'the detached banana'],
    ['examples/images/skateboard.jpg', None, None, 'General', 'which region does the text " a yellow bird " describe?'],
    ['examples/images/baseball.jpg', None, None, 'General', 'what color is the left car?'],
    [None, None, 'examples/videos/video7014.mp4', 'Video Captioning', None], 
    [None, None, 'examples/videos/video7017.mp4', 'Video Captioning', None], 
    [None, None, 'examples/videos/video7019.mp4', 'Video Captioning', None], 
    [None, None, 'examples/videos/video7021.mp4', 'Video Captioning', None], 
    [None, None, 'examples/videos/video7021.mp4', 'General Video', "What is this sport?"], 
    [None, 'examples/audios/6cS0FsUM-cQ.wav', None, 'Audio Captioning', None],
    [None, 'examples/audios/AJtNitYMa1I.wav', None, 'Audio Captioning', None],
]

title = "eP-ALM"
description = "Gradio Demo for eP-ALM: "
article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2303.11403' target='_blank'>Paper</a> | <a href='https://github.com/mshukor/eP-ALM' target='_blank'>Github Repo</a></p>"

io = gr.Interface(fn=inference, inputs=inputs, outputs=outputs,
                  title=title, description=description, article=article, examples=examples, cache_examples=False)
io.launch()