app.py

import os
try:
    # Try to install detectron2 from source. Needed for semseg plotting functionality.
    os.system("python -m pip install 'git+https://github.com/facebookresearch/detectron2.git'")
except Exception as e:
    print('detectron2 cannot be installed. Falling back to simple semseg visualization.')
    print(e)

import torch
# We recommend running this demo on an A100 GPU
if torch.cuda.is_available():
    device = "cuda"
    gpu_type = torch.cuda.get_device_name(torch.cuda.current_device())
    power_device = f"{gpu_type} GPU"
    torch.cuda.max_memory_allocated(device=device)
else:
    device = "cpu"
    power_device = "CPU"
    os.system("pip uninstall -y xformers") # Only use xformers on GPU

import spaces
import gradio as gr
import random
import numpy as np
from torchvision.transforms.functional import center_crop
from fourm.demo_4M_sampler import Demo4MSampler
from fourm.data.modality_transforms import RGBTransform


# The flag below controls whether to allow TF32 on matmul. This flag defaults to False in PyTorch 1.12 and later.
torch.backends.cuda.matmul.allow_tf32 = True
# The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
torch.backends.cudnn.allow_tf32 = True

MAX_SEED = np.iinfo(np.int32).max

FM_MODEL_ID = 'EPFL-VILAB/4M-21_B'
MODEL_NAME = FM_MODEL_ID.split('/')[1].replace('_', ' ')

# Human poses visualization is disabled, since it needs SMPL weights. To enable human pose prediction and rendering:
# 1) Install via `pip install timm yacs smplx pyrender pyopengl==3.1.4`
#    You may need to follow the pyrender install instructions: https://pyrender.readthedocs.io/en/latest/install/index.html
# 2) Download SMPL data from https://smpl.is.tue.mpg.de/. See https://github.com/shubham-goel/4D-Humans/ for an example
# 3) Copy the required SMPL files (smpl_mean_params.npz, SMPL_to_J19.pkl, smpl/SMPL_NEUTRAL.pkl) to fourm/utils/hmr2_utils/data .
MANUAL_MODS_OVERRIDE = [
    'color_palette', 'tok_depth@224', 'tok_imagebind@224', 'sam_instance', 'tok_dinov2_global', 
    'tok_normal@224', 'tok_sam_edge@224', 'det', 'tok_canny_edge@224', 'tok_semseg@224', 'rgb@224', 
    'caption', 't5_caption', 'tok_imagebind_global', 'tok_rgb@224', 'tok_clip@224', 'metadata', 'tok_dinov2@224'
]

sampler = Demo4MSampler(
    fm=FM_MODEL_ID, 
    fm_sr=None, 
    tok_human_poses=None, 
    tok_text='./text_tokenizer_4m_wordpiece_30k.json',
    mods=MANUAL_MODS_OVERRIDE,
).to(device)


def img_from_path(img_path: str):
    rgb_transform = RGBTransform(imagenet_default_mean_and_std=True)
    img_pil = rgb_transform.load(img_path)
    img_pil = rgb_transform.preprocess(img_pil)
    img_pil = center_crop(img_pil, (min(img_pil.size), min(img_pil.size))).resize((224,224))
    img = rgb_transform.postprocess(img_pil).unsqueeze(0)
    return img

@spaces.GPU
def infer(img_path, seed=0, randomize_seed=False, target_modalities=None, top_p=0.8, top_k=0.0):
    if randomize_seed:
        seed = None
    img = img_from_path(img_path).to(device)
    preds = sampler({'rgb@224': img}, seed=seed, target_modalities=target_modalities, top_p=top_p, top_k=top_k) 
    sampler.plot_modalities(preds, ncols_max=4, use_fixed_plotting_order=True, save_path='./output.png')
    return './output.png'


examples = [
    'examples/example_0.png', 'examples/example_1.png', 'examples/example_2.png',
    'examples/example_3.png', 'examples/example_4.png', 'examples/example_5.png',
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 1500px;
}
#col-input-container {
    margin: 0 auto;
    max-width: 400px;
}
#run-button {
    margin: 0 auto;
}
"""

with gr.Blocks(css=css, theme=gr.themes.Base()) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
        # 4M: Massively Multimodal Masked Modeling
        """)
        
        with gr.Row():
            with gr.Column(elem_id="col-input-container"):
                gr.Markdown(f"""
                *A framework for training any-to-any multimodal foundation models. Scalable. Open-sourced. Across tens of modalities and tasks.*
                
                [`Website`](https://4m.epfl.ch) | [`GitHub`](https://github.com/apple/ml-4m) <br>[`4M Paper (NeurIPS'23)`](https://arxiv.org/abs/2312.06647) | [`4M-21 Paper (arXiv'24)`](https://arxiv.org/abs/2406.09406)

                This demo predicts all modalities from a given RGB input, using [{FM_MODEL_ID}](https://huggingface.co/{FM_MODEL_ID}), running on *{power_device}*.
                For more generative examples, and to enable human pose visualizations, please see our [GitHub repo](https://github.com/apple/ml-4m).
                
                (Disclaimer: The demo is a work in progress. We will switch it to using 4M-21 XL when running on GPU. Until then, this space runs on CPU and takes several minutes for inference.)
                """)
                
                img_path = gr.Image(label='RGB input image', type='filepath')
                run_button = gr.Button(f"Predict with {MODEL_NAME}", scale=0, elem_id="run-button")

                with gr.Accordion("Advanced Settings", open=False):
                    target_modalities = gr.CheckboxGroup(
                        choices=[
                            ('CLIP-B/16', 'tok_clip@224'), ('DINOv2-B/14', 'tok_dinov2@224'), ('ImageBind-H/14', 'tok_imagebind@224'), 
                            ('Depth', 'tok_depth@224'), ('Surface normals', 'tok_normal@224'), ('Semantic segmentation', 'tok_semseg@224'), 
                            ('Canny edges', 'tok_canny_edge@224'), ('SAM edges', 'tok_sam_edge@224'), ('Caption', 'caption'), 
                            ('Bounding boxes', 'det'), ('SAM instances', 'sam_instance'), ('Color palette', 'color_palette'), 
                            ('Metadata', 'metadata'),
                        ],
                        value=[
                            'tok_clip@224', 'tok_dinov2@224', 'tok_imagebind@224', 
                            'tok_depth@224', 'tok_normal@224', 'tok_semseg@224', 
                            'tok_canny_edge@224', 'tok_sam_edge@224', 'caption', 
                            'det', 'sam_instance', 'color_palette', 'metadata'
                        ],
                        label="Target modalities", 
                        info='Choose which modalities are predicted (in this order).'
                    )
                    seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0)
                    randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
                    top_p = gr.Slider(label="Top-p", minimum=0.0, maximum=1.0, step=0.01, value=0.8)
                    top_k = gr.Slider(label="Top-k", minimum=0.0, maximum=1.0, step=0.01, value=0.0)
        
            result = gr.Image(label="Predictions", show_label=False)

        gr.Examples(
            examples = examples,
            fn = infer,
            inputs = [img_path],
            outputs = [result],
            cache_examples='lazy',
        )

    run_button.click(
        fn = infer,
        inputs = [img_path, seed, randomize_seed, target_modalities, top_p, top_k],
        outputs = [result]
    )

demo.queue(max_size=10).launch()