Some questions about the code

[winner-4]

   # get class_label
    is_trans = data["is_trans"].cuda()
    trans = 1 - is_trans

Why am I getting an error message during reasoning?

File "/workspace/03_Matting/SDMatte/modeling/SDMatte/meta_arch.py", line 197, in forward
is_trans = data["is_trans"].cuda()
~~~~^^^^^^^^^^^^
KeyError: 'is_trans'

[LongfeiH]

Did you make any modifications to the data/dataset.py file?

[winner-4]

Did you make any modifications to the data/dataset.py file?

I did not make any changes to data/dataset.py file. Below is my inference code.

`import os
import cv2
import numpy as np
import torch
from tqdm import tqdm
from torch.utils.data import Dataset, DataLoader
from torchvision.transforms import functional as F
from os.path import join
from detectron2.checkpoint import DetectionCheckpointer
from detectron2.config import LazyConfig, instantiate
from detectron2.engine import default_argument_parser
import warnings

from data.dataset import DataGenerator

import time

class SDMatteInferencer:
def init(self, config_path, checkpoint_path):
# Initialize model
self.cfg = LazyConfig.load(config_path)
self.model = instantiate(self.cfg.model).to(self.cfg.train.device)
DetectionCheckpointer(self.model).load(checkpoint_path)
self.model.eval()

def _process_prompts(self, image, points=None, bbox=None, mask=None):
    """Convert different prompts to model's expected format"""
    prompt_mask = np.zeros_like(image[..., 0])
    h, w = prompt_mask.shape

    # Process points (foreground)
    if points is not None:
        print('process points')
        for x, y in points:
            cv2.circle(prompt_mask, (int(x*w), int(y*h)), 5, 255, -1)

    # Process bbox [x1,y1,x2,y2] (relative coordinates)
    if bbox is not None:
        print('process bbox')
        x1, y1, x2, y2 = [int(v*w) if i%2==0 else int(v*h) for i,v in enumerate(bbox)]
        cv2.rectangle(prompt_mask, (x1, y1), (x2, y2), 255, 2)

    # Process mask
    if mask is not None:
        print('process mask')
        prompt_mask = np.maximum(prompt_mask, mask*255)

    return torch.tensor(prompt_mask, dtype=torch.float32)[None, None,...]/255.0

@torch.no_grad()
def predict(self, image_path, points=None, bbox=None, mask_path=None):
    # Read input
    orig_image = cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)
    h, w = orig_image.shape[:2]

    # Read optional mask
    mask = cv2.imread(mask_path, 0) if mask_path else None

    # Preprocess
    image_tensor = F.to_tensor(orig_image).unsqueeze(0).to(self.cfg.train.device)
    prompt_tensor = self._process_prompts(orig_image, points, bbox, mask).to(self.cfg.train.device)

    # Inference
    pred_alpha = self.model({
        "image": image_tensor,
        "prompt": prompt_tensor
    })[0].squeeze()

    # Postprocess
    alpha = (pred_alpha * 255).byte().cpu().numpy()
    return Image.fromarray(alpha).resize((w, h), Image.BILINEAR)

Usage example

if name == "main":
# 初始化
inferencer = SDMatteInferencer(
config_path="./configs/SDMatte.py",
checkpoint_path="./checkpoints/SDMatte.pth"
)

# 输入示例：图片+两点+bbox
result = inferencer.predict(
    image_path="./one_person/images/02_0243_input.jpg",
    mask_path = './one_person/sam3_hair/02_0243_input.png'
)

result.save("02_0243_input.png")

`

The error message is as follows:

Traceback (most recent call last): File "/workspace/03_Matting/SDMatte/inference_mph.py", line 81, in <module> result = inferencer.predict( ^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/workspace/03_Matting/SDMatte/inference_mph.py", line 63, in predict pred_alpha = self.model({ ^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/03_Matting/SDMatte/modeling/SDMatte/meta_arch.py", line 211, in forward label_latent = self.unet( ^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/workspace/03_Matting/SDMatte/utils/replace.py", line 456, in forward aug_emb = self.bbox_embedding(coords_embeds) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/diffusers/models/embeddings.py", line 1310, in forward sample = self.linear_1(sample) ^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1739, in _wrapped_call_impl return self._call_impl(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl return forward_call(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/opt/conda/envs/sam2/lib/python3.11/site-packages/torch/nn/modules/linear.py", line 125, in forward return F.linear(input, self.weight, self.bias) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: mat1 and mat2 shapes cannot be multiplied (1x373248000 and 1280x1280)

[LongfeiH]

Your code is missing a hint about whether the target object is transparent, specifically is_trans. You can refer to the relevant code in data/dataset.py to update your code. Alternatively, you can simply set this quantity to tensor 1.