MMaDA/inference_v2t.py

# coding=utf-8
# Copyright 2025 AIDAS Team
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import os
os.environ["TOKENIZERS_PARALLELISM"] = "true"
from PIL import Image
from tqdm import tqdm
import numpy as np
import torch
import wandb
import cv2
from models import MAGVITv2, MMadaConfig, MMadaModelLM
from training.prompting_utils import UniversalPrompting
from training.utils import get_config, flatten_omega_conf, image_transform
from transformers import AutoTokenizer, AutoConfig

def resize_vocab(model, config):
    print(f"Resizing token embeddings to {config.new_vocab_size}")
    model.resize_token_embeddings(config.new_vocab_size)

def get_vq_model_class(model_type):
    if model_type == "magvitv2":
        return MAGVITv2
    else:
        raise ValueError(f"model_type {model_type} not supported.")

def inference_video():
    
    pass

def load_video(
        video_path, 
        config, 
        uni_prompting,
        vq_model=None,
        device='cuda',
        sample='uniform', 
        num_frames=8
    ):
    """
    args:
        video_path: path to the video file
        return: video frames as a list of images
    """
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise IOError(f"Could not open video file {video_path}")
    
    frames = []
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        # Convert BGR to RGB
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        frames.append(Image.fromarray(frame))
        
    cap.release()
    
    total_frames = len(frames)
    
    if total_frames < num_frames:
        raise ValueError(f"Video {video_path} has less than 8 frames, got {total_frames} frames.")
    
    if sample == 'uniform':
        indices = np.linspace(0, total_frames - 1, num_frames).astype(int)
    elif sample == 'random':
        raise NotImplementedError("Random sampling not implemented yet.")
    else:
        raise ValueError(f"Sampling method {sample} not supported.")
    
    sampled_frames = []
    sampled_frames_tokens = []
    for idx in indices:
        frame = frames[idx]
        frame = image_transform(frame, resolution=config.dataset.params.resolution).to(device)
        sampled_frames.append(frame.unsqueeze(0))
        sampled_frames_tokens.append(
            vq_model.get_code(frame.unsqueeze(0)) + len(uni_prompting.text_tokenizer)
        )
        
    # num_frames * [num_frames, seq_len] -> [1, num_frames * seq_len]
    video_tokens = torch.cat(sampled_frames_tokens, dim=1) 
    
    return sampled_frames, video_tokens
        
        
        
def main():
    config = get_config()
    resume_wandb_run = config.wandb.resume
    run_id = config.wandb.get("run_id", None)
    if run_id is None:
        resume_wandb_run = False
        run_id = wandb.util.generate_id()
        config.wandb.run_id = run_id

    wandb_config = {k: v for k, v in flatten_omega_conf(config, resolve=True)}

    wandb.init(
        project="demo",
        name=config.experiment.name + '_video',
        config=wandb_config,
    )

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    tokenizer = AutoTokenizer.from_pretrained(config.model.mmada.pretrained_model_path, padding_side="left")

    uni_prompting = UniversalPrompting(tokenizer, max_text_len=config.dataset.preprocessing.max_seq_length,
                                       special_tokens=("<|soi|>", "<|eoi|>", "<|sov|>", "<|eov|>", "<|t2i|>", "<|mmu|>", "<|t2v|>", "<|v2v|>", "<|lvg|>", "<|v2t|>"),
                                       ignore_id=-100, cond_dropout_prob=config.training.cond_dropout_prob, use_reserved_token=True)

    vq_model = get_vq_model_class(config.model.vq_model.type)
    vq_model = vq_model.from_pretrained(config.model.vq_model.vq_model_name).to(device)
    vq_model.requires_grad_(False)
    vq_model.eval()

    train_step = config.step
    trained_checkpoint_path = f"/home/work/AIDAS/ckpts/omada/omada-training-stage1/checkpoint-{train_step}/unwrapped_model"

    model = MMadaModelLM.from_pretrained(trained_checkpoint_path, trust_remote_code=True, torch_dtype=torch.bfloat16, config="/home/work/AIDAS/ckpts/omada/omada-training-stage1/config.json")
    # model = MMadaModelLM.from_pretrained("Gen-Verse/MMaDA-8B-MixCoT", trust_remote_code=True, torch_dtype=torch.bfloat16)
    model.to(device)

    mask_token_id = model.config.mask_token_id

    temperature = 0.8  # 1.0 = no change, < 1.0 = less random, > 1.0 = more random, in predictions
    top_k = 1  # retain only the top_k most likely tokens, clamp others to have 0 probability
    file_list = os.listdir(config.video_image_root)
    file_list = [f for f in file_list if f.lower().endswith(('.mp4'))]
    responses = ['' for i in range(len(file_list))]
    videos = []
    config.question = config.question.split(' *** ')
    for i, file_name in enumerate(tqdm(file_list)):
        video_path = os.path.join(config.video_image_root, file_name)
        print("current video path:", video_path)
        video_frames, video_tokens = load_video(
            video_path, 
            config, 
            uni_prompting,
            vq_model=vq_model,
            device=device,
            sample='uniform', 
            num_frames=8
        )
        print("video tokens shape:", video_tokens.shape)
        batch_size = 1

        for question in config.question:
            input_ids = uni_prompting.text_tokenizer(['<|start_header_id|>user<|end_header_id|>\n' + question +'<eot_id><|start_header_id|>assistant<|end_header_id|>\n'])['input_ids']
            input_ids = torch.tensor(input_ids).to(device)

            input_ids = torch.cat([
                (torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|v2t|>']).to(device),
                (torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|soi|>']).to(device),
                video_tokens,
                (torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|eoi|>']).to(device),
                (torch.ones(input_ids.shape[0], 1) * uni_prompting.sptids_dict['<|sot|>']).to(device),
                input_ids
            ], dim=1).long()
            print(f"input_ids shape: {input_ids.shape}")
    
            output_ids = model.mmu_generate(input_ids, max_new_tokens=128, steps=128, block_length=128)
            text = uni_prompting.text_tokenizer.batch_decode(output_ids[:, input_ids.shape[1]:], skip_special_tokens=True)
            print(text)
            responses[i] += f'User: ' + question + f'\n Answer : ' + text[0] + '\n'

    # images = torch.cat(images, dim=0)
    # images = torch.clamp((images + 1.0) / 2.0, min=0.0, max=1.0)
    # images *= 255.0
    # images = images.permute(0, 2, 3, 1).cpu().numpy().astype(np.uint8)
    # pil_images = [Image.fromarray(image) for image in images]

    # wandb_images = [wandb.Image(image, caption=responses[i]) for i, image in enumerate(pil_images)]
    # wandb.log({"multimodal understanding": wandb_images}, step=0)


if __name__ == '__main__':
    main()