File size: 6,826 Bytes
434b0b0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 | # -*- coding: utf-8 -*-
# @Organization : Tongyi Lab, Alibaba
# @Author : Lingteng Qiu
# @Email : 220019047@link.cuhk.edu.cn
# @Time : 2025-10-20 10:00:00
# @Function : LHM++ Inference Utils Tool
import base64
import glob
import json
import os
import tempfile
from typing import Dict, List, Optional
import gradio as gr
import numpy as np
import torch
from PIL import Image
SMPLX_EXPRESSION_DIM = 100
from contextlib import contextmanager
def get_smplx_params(
data: Dict[str, torch.Tensor],
device: torch.device,
) -> Dict[str, torch.Tensor]:
"""Extract and prepare SMPL-X parameters for model input.
Filters relevant SMPL-X parameter keys from the input data dictionary
and moves them to the specified device with an added batch dimension.
Args:
data: Dictionary containing SMPL-X parameters and other data.
device: Target device (e.g., 'cuda' or 'cpu') for the parameters.
Returns:
Dictionary containing only SMPL-X parameters with shape (1, ...)
on the specified device.
"""
smplx_keys = [
"root_pose",
"body_pose",
"jaw_pose",
"leye_pose",
"reye_pose",
"lhand_pose",
"rhand_pose",
"expr",
"trans",
"betas",
]
smplx_params = {k: data[k].unsqueeze(0).to(device) for k in smplx_keys if k in data}
return smplx_params
def obtain_motion_sequence(motion_dir: str) -> List[Dict[str, torch.Tensor]]:
"""Load motion sequence data from SMPL-X and FLAME parameter files.
Reads SMPL-X parameter JSON files from the specified directory and optionally
merges them with corresponding FLAME parameters for facial expressions and poses.
Args:
motion_dir: Path to directory containing SMPL-X parameter JSON files.
Returns:
List of dictionaries containing SMPL-X parameters as torch tensors.
Each dictionary includes body pose, facial expressions, and eye poses.
"""
motion_files = sorted(glob.glob(os.path.join(motion_dir, "*.json")))
smplx_list = []
for motion_file in motion_files:
# Load SMPL-X parameters
with open(motion_file, "r") as f:
smplx_params = json.load(f)
# Try to load corresponding FLAME parameters
flame_path = motion_file.replace("smplx_params", "flame_params")
if os.path.exists(flame_path):
with open(flame_path, "r") as f:
flame_params = json.load(f)
# Override SMPL-X parameters with FLAME data
smplx_params["expr"] = torch.FloatTensor(flame_params["expcode"])
smplx_params["jaw_pose"] = torch.FloatTensor(flame_params["posecode"][3:])
smplx_params["leye_pose"] = torch.FloatTensor(flame_params["eyecode"][:3])
smplx_params["reye_pose"] = torch.FloatTensor(flame_params["eyecode"][3:])
else:
# Use zero expressions if FLAME params not available
# smplx_params["expr"] = torch.zeros(SMPLX_EXPRESSION_DIM)
# Using SMPLX expression
pass
smplx_list.append(smplx_params)
return smplx_list
def assert_input_image(input_image: Optional[np.ndarray]) -> None:
"""Validate input image is not None.
Args:
input_image: Input image array to validate.
Raises:
gr.Error: If input image is None.
"""
if input_image is None:
raise gr.Error("No image selected or uploaded!")
def prepare_working_dir() -> tempfile.TemporaryDirectory:
"""Create a temporary working directory.
Returns:
Temporary directory object.
"""
return tempfile.TemporaryDirectory()
def init_preprocessor() -> None:
"""Initialize the global preprocessor for image preprocessing."""
from core.utils.preprocess import Preprocessor
global preprocessor
preprocessor = Preprocessor()
def preprocess_fn(
image_in: np.ndarray,
remove_bg: bool,
recenter: bool,
working_dir: tempfile.TemporaryDirectory,
) -> str:
"""Preprocess input image with optional background removal and recentering.
Args:
image_in: Input image as numpy array.
remove_bg: Whether to remove background.
recenter: Whether to recenter the subject.
working_dir: Temporary directory for storing intermediate files.
Returns:
Path to the preprocessed image.
Raises:
AssertionError: If preprocessing fails.
"""
image_raw = os.path.join(working_dir.name, "raw.png")
with Image.fromarray(image_in) as img:
img.save(image_raw)
image_out = os.path.join(working_dir.name, "rembg.png")
success = preprocessor.preprocess(
image_path=image_raw, save_path=image_out, rmbg=remove_bg, recenter=recenter
)
if not success:
raise RuntimeError("Preprocessing failed")
return image_out
def get_image_base64(path: str) -> str:
"""Convert image file to base64 encoded string.
Args:
path: Path to image file.
Returns:
Base64 encoded image string with data URI prefix.
"""
with open(path, "rb") as f:
encoded_string = base64.b64encode(f.read()).decode()
return f"data:image/png;base64,{encoded_string}"
def get_available_device() -> str:
"""Returns the current CUDA device or ``"cpu"`` if CUDA is unavailable.
When CUDA is available, uses the device ID from ``torch.cuda.current_device()``
so the caller runs on the same device as the active CUDA context.
Returns:
str: Device string (e.g., ``"cuda:0"`` or ``"cpu"``).
"""
if torch.cuda.is_available():
current_device_id = torch.cuda.current_device()
device = f"cuda:{current_device_id}"
else:
device = "cpu"
return device
@contextmanager
def easy_memory_manager(model: torch.nn.Module, device: str = "cuda"):
"""Context manager that moves a model to GPU during use and back to CPU afterward.
Reduces GPU memory footprint by transferring the model off CUDA and clearing
the cache when the block exits. Use when running inference in memory-constrained
environments.
Args:
model (torch.nn.Module): The model to manage. Will be moved to the current
CUDA device (or CPU if unavailable) for the duration of the context.
device (str, optional): Unused; kept for API compatibility. Default: ``"cuda"``.
Yields:
torch.nn.Module: The model, ready for use on the target device.
Example:
>>> with easy_memory_manager(model):
... output = model(input_tensor)
"""
device = get_available_device()
model.to(device)
try:
yield model
finally:
model.to("cpu")
if torch.cuda.is_available():
torch.cuda.empty_cache()
|