app.py

import os
import torch
import spaces
import matplotlib

import numpy as np
import gradio as gr

from PIL import Image
from transformers import pipeline
from huggingface_hub import hf_hub_download
from gradio_imageslider import ImageSlider

from depth_anything_v2.dpt import DepthAnythingV2
from loguru import logger

css = """
#img-display-container {
    max-height: 100vh;
}
#img-display-input {
    max-height: 80vh;
}
#img-display-output {
    max-height: 80vh;
}
#download {
    height: 62px;
}
"""

title = "# Depth Anything: Watch V1 and V2 side by side."
description1 = """Please refer to **Depth Anything V2** [paper](https://arxiv.org/abs/2406.09414) for more details."""

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
DEFAULT_V2_MODEL_NAME = "Base"
DEFAULT_V1_MODEL_NAME = "Base"

cmap = matplotlib.colormaps.get_cmap('Spectral_r')

# --------------------------------------------------------------------
# Depth anything V1 configuration
# --------------------------------------------------------------------
depth_anything_v1_name2checkpoint = {
    "Small": "LiheYoung/depth-anything-small-hf",
    "Base": "LiheYoung/depth-anything-base-hf",
    "Large": "LiheYoung/depth-anything-large-hf",
}

depth_anything_v1_pipelines = {}
# --------------------------------------------------------------------
# Depth anything V2 configuration
# --------------------------------------------------------------------

depth_anything_v2_configs = {
    'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]},
    'vitb': {'encoder': 'vitb', 'features': 128, 'out_channels': [96, 192, 384, 768]},
    'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]},
    'vitg': {'encoder': 'vitg', 'features': 384, 'out_channels': [1536, 1536, 1536, 1536]}
}
depth_anything_v2_encoder2name = {
    'vits': 'Small',
    'vitb': 'Base',
    'vitl': 'Large',
    # 'vitg': 'Giant', # we are undergoing company review procedures to release our giant model checkpoint
}
depth_anything_v2_name2encoder = {v: k for k, v in depth_anything_v2_encoder2name.items()}

depth_anything_v2_models = {}
# --------------------------------------------------------------------


def get_v1_pipe(model_name):
    return pipeline(task="depth-estimation", model=depth_anything_v1_name2checkpoint[model_name], device=DEVICE)


def get_v2_model(model_name):
    encoder = depth_anything_v2_name2encoder[model_name]
    model = DepthAnythingV2(**depth_anything_v2_configs[encoder])
    filepath = hf_hub_download(repo_id=f"depth-anything/Depth-Anything-V2-{model_name}", filename=f"depth_anything_v2_{encoder}.pth", repo_type="model")
    state_dict = torch.load(filepath, map_location="cpu")
    model.load_state_dict(state_dict)
    model = model.to(DEVICE).eval()
    return model


def predict_depth_v1(image, model_name):
    if model_name not in depth_anything_v1_pipelines:
        depth_anything_v1_pipelines[model_name] = get_v1_pipe(model_name)
    pipe = depth_anything_v1_pipelines[model_name]
    return pipe(image)


def predict_depth_v2(image, model_name):
    if model_name not in depth_anything_v2_models:
        depth_anything_v2_models[model_name] = get_v2_model(model_name)
    model = depth_anything_v2_models[model_name].cuda()
    return model.infer_image(image)


def compute_depth_map_v2(image, model_select: str):
    depth = predict_depth_v2(image[:, :, ::-1], model_select)
    depth = (depth - depth.min()) / (depth.max() - depth.min()) * 255.0
    depth = depth.astype(np.uint8)
    colored_depth = (cmap(depth)[:, :, :3] * 255).astype(np.uint8)
    return colored_depth


def compute_depth_map_v1(image, model_select):
    pil_image = Image.fromarray(image)
    depth = predict_depth_v1(pil_image, model_select)
    depth = np.array(depth["depth"]).astype(np.uint8)
    colored_depth = (cmap(depth)[:, :, :3] * 255).astype(np.uint8)
    return colored_depth


@spaces.GPU
@torch.no_grad()
def on_submit(image, model_v1_select, model_v2_select):
    logger.info(f"Computing depth for V1 model: {model_v1_select} and V2 model: {model_v2_select}")
    colored_depth_v1 = compute_depth_map_v1(image, model_v1_select)
    colored_depth_v2 = compute_depth_map_v2(image, model_v2_select)
    return colored_depth_v1, colored_depth_v2


with gr.Blocks(css=css) as demo:
    gr.Markdown(title)
    gr.Markdown(description1)
    gr.Markdown("### Depth Prediction demo")
    with gr.Row():
        model_select_v1 = gr.Dropdown(label="Depth Anything V1 Model", choices=list(depth_anything_v1_name2checkpoint.keys()), value=DEFAULT_V1_MODEL_NAME)
        model_select_v2 = gr.Dropdown(label="Depth Anything V2 Model", choices=list(depth_anything_v2_encoder2name.values()), value=DEFAULT_V2_MODEL_NAME)
    with gr.Row():
        gr.Markdown()
        gr.Markdown("Depth Maps: V1 <-> V2")
    with gr.Row():
        input_image = gr.Image(label="Input Image", type='numpy', elem_id='img-display-input')
        depth_image_slider = ImageSlider(elem_id='img-display-output', position=0.5)

    submit = gr.Button(value="Compute Depth")
    submit.click(on_submit, inputs=[input_image, model_select_v1, model_select_v2], outputs=[depth_image_slider])

    example_files = os.listdir('assets/examples')
    example_files.sort()
    example_files = [os.path.join('assets/examples', filename) for filename in example_files]
    examples = gr.Examples(examples=example_files, inputs=[input_image])


if __name__ == '__main__':
    demo.queue().launch(share=True)