GroundingDINO Model Error on GPU: Failure to Annotate with GPU Mode #389
Comments
我把我的onnxruntime-gpu从1.17.1降到1.16.0解决了这个问题，要根据你的cuda和cudnn版本来安装onnxruntime-gpu，详情参考https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html |
|
@Hermione-ff 谢谢，我将onnxruntime的版本调整为1.16.0，目前可以正常运行标注，但是仍然存在如下问题： |
CVHub520
changed the title
|
我也有遇到類似的問題。目前猜測是sequence length 的問題，解法的話是透過 padding 到 max_length。 主要要更改兩個部份，第一個是生成 onnx ，第二個是模型本身。 生成 onnx使用以下程式: import argparse
import os
import os.path as osp
import onnxruntime as ort
import torch
from groundingdino.models import build_model
from groundingdino.models.GroundingDINO.bertwarper import (
generate_masks_with_special_tokens_and_transfer_map,
)
from groundingdino.util.slconfig import SLConfig
from groundingdino.util.utils import clean_state_dict
"""
The ONNX Export of the Grounding DINO
Written by Wei Wang (CVHub)
Usage:
1. git clone https://github.com/IDEA-Research/GroundingDINO.git
2. cd GroundingDINO and pip install -r requirements.txt
3. export PYTHONPATH=/path/to/your/GroundingDINO
4. Place the current script in this directory.
5. Download the corresponding tokenizer.json and place it in this dir.
6. Run the script.
```bash
python export_grounding_dino_onnx.py \
--config_file groundingdino/config/GroundingDINO_SwinB_cfg.py \
--ckpt_file /path/to/your/groundingdino_swinb_cogcoor.pth or groundingdino_swint_ogc \
--device 'cpu' or 'gpu'
```
"""
class Args:
def __init__(self, **kwargs):
self.__dict__.update(kwargs)
class OnnxBaseModel:
def __init__(self, model_path, device_type: str = "gpu"):
self.sess_opts = ort.SessionOptions()
if "OMP_NUM_THREADS" in os.environ:
self.sess_opts.inter_op_num_threads = int(os.environ["OMP_NUM_THREADS"])
self.providers = ["CPUExecutionProvider"]
if device_type.lower() == "gpu":
self.providers = ["CUDAExecutionProvider"]
self.ort_session = ort.InferenceSession(
model_path,
providers=self.providers,
sess_options=self.sess_opts,
)
def get_ort_inference(self, blob, inputs=None, extract=True, squeeze=False):
if inputs is None:
inputs = self.get_input_name()
outs = self.ort_session.run(None, {inputs: blob})
else:
outs = self.ort_session.run(None, inputs)
if extract:
outs = outs[0]
if squeeze:
outs = outs.squeeze(axis=0)
return outs
def get_input_name(self):
return self.ort_session.get_inputs()[0].name
def get_input_shape(self):
return self.ort_session.get_inputs()[0].shape
def get_output_name(self):
return [out.name for out in self.ort_session.get_outputs()]
def load_pt_model(model_config_path, model_checkpoint_path, cpu_only=False):
args = SLConfig.fromfile(model_config_path)
args.device = "cuda" if not cpu_only else "cpu"
# modified config
args.use_checkpoint = False
args.use_transformer_ckpt = False
model = build_model(args)
checkpoint = torch.load(model_checkpoint_path, map_location="cpu")
model.load_state_dict(clean_state_dict(checkpoint["model"]), strict=False)
_ = model.eval()
return model
def export_onnx(model, output_file):
caption = "the running dog ." # ". ".join(input_text)
tokenized = model.tokenizer([caption], padding="max_length", return_tensors="pt")
(
text_self_attention_masks,
position_ids,
_,
) = generate_masks_with_special_tokens_and_transfer_map(
tokenized, model.specical_tokens, model.tokenizer
)
input_ids = tokenized["input_ids"]
token_type_ids = tokenized["token_type_ids"]
attention_mask = tokenized["attention_mask"]
text_token_mask = text_self_attention_masks
img = torch.randn(1, 3, 800, 1200)
dynamic_axes = {
"input_ids": {0: "batch_size", 1: "seq_len"},
"attention_mask": {0: "batch_size", 1: "seq_len"},
"position_ids": {0: "batch_size", 1: "seq_len"},
"token_type_ids": {0: "batch_size", 1: "seq_len"},
"text_token_mask": {0: "batch_size", 1: "seq_len", 2: "seq_len"},
"img": {0: "batch_size", 2: "height", 3: "width"},
"logits": {0: "batch_size"},
"boxes": {0: "batch_size"},
}
args = (
img,
input_ids,
attention_mask,
position_ids,
token_type_ids,
text_token_mask,
)
input_names = [
"img",
"input_ids",
"attention_mask",
"position_ids",
"token_type_ids",
"text_token_mask",
]
output_names = ["logits", "boxes"]
# export onnx model
torch.onnx.export(
model,
f=output_file,
args=args,
input_names=input_names,
output_names=output_names,
dynamic_axes=dynamic_axes,
opset_version=16,
)
print("Done!")
if __name__ == "__main__":
parser = argparse.ArgumentParser("Export Grounding DINO Model to ONNX", add_help=True)
parser.add_argument(
"--config_file",
"-c",
type=str,
required=True,
help="path to config file",
)
parser.add_argument(
"--ckpt_file",
"-p",
type=str,
required=True,
help="path to checkpoint file",
)
parser.add_argument("--output_dir", "-o", type=str, help="output directory")
parser.add_argument(
"--img_path",
"-i",
type=str,
default="asset/demo2.jpg",
help="Test image",
)
parser.add_argument(
"--text_prompt",
"-t",
type=str,
default="The running dog",
help="Text prompt",
)
parser.add_argument(
"--device",
type=str,
default="cpu",
choices=["cpu", "gpu"],
help="Device",
)
parser.add_argument("--box_threshold", type=float, default=0.3, help="Box prediction score")
parser.add_argument("--text_threshold", type=float, default=0.25, help="Text prompt score")
args = parser.parse_args()
# cfg
config_file = args.config_file # change the path of the model config file
ckpt_file = args.ckpt_file # change the path of the model
output_dir = args.output_dir
img_path = args.img_path
text_prompt = args.text_prompt
device = args.device
box_threshold = args.box_threshold
text_threshold = args.text_threshold
onnx_file = osp.splitext(osp.basename(ckpt_file))[0] + ".onnx"
# make dir
if output_dir:
os.makedirs(output_dir, exist_ok=True)
onnx_file = osp.join(output_dir, onnx_file)
print(f"onnx_file = {onnx_file}")
if not osp.exists(onnx_file):
# load model
model = load_pt_model(config_file, ckpt_file, cpu_only=True)
# export model
export_onnx(model, onnx_file)目前是簡單改寫，可以把官方的 quant 加上去，用法也一樣。 替換模型首先替換掉 @staticmethod
def get_tokenlizer(text_encoder_type="bert-base-uncased"):
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained(text_encoder_type)
return tokenizer接著將 def preprocess(self, image, text_prompt, img_mask=None):
# encoder texts
captions = self.get_caption(str(text_prompt))
tokenized_raw_results = self.net.tokenizer([captions], padding="max_length", max_length=256, return_tensors="np")
tokenized = {
"input_ids": tokenized_raw_results.input_ids.astype(np.int64),
"token_type_ids": tokenized_raw_results.token_type_ids.astype(np.int64),
"attention_mask": tokenized_raw_results.attention_mask,
}
specical_tokens = [101, 102, 1012, 1029]
(
text_self_attention_masks,
position_ids,
_,
) = self.generate_masks_with_special_tokens_and_transfer_map(tokenized, specical_tokens)
if text_self_attention_masks.shape[1] > self.net.max_text_len:
text_self_attention_masks = text_self_attention_masks[
:, : self.net.max_text_len, : self.net.max_text_len
]
position_ids = position_ids[:, : self.net.max_text_len]
tokenized["input_ids"] = tokenized["input_ids"][:, : self.net.max_text_len]
tokenized["attention_mask"] = tokenized["attention_mask"][:, : self.net.max_text_len]
tokenized["token_type_ids"] = tokenized["token_type_ids"][:, : self.net.max_text_len]
inputs = {}
inputs["img"] = image
inputs["input_ids"] = np.array(tokenized["input_ids"], dtype=np.int64)
inputs["attention_mask"] = np.array(tokenized["attention_mask"], dtype=bool)
inputs["token_type_ids"] = np.array(tokenized["token_type_ids"], dtype=np.int64)
inputs["position_ids"] = np.array(position_ids, dtype=np.int64)
inputs["text_token_mask"] = np.array(text_self_attention_masks, dtype=bool)
return image, inputs, captions最後將 def postprocess(self, outputs, caption, with_logits=True, token_spans=None):
logits, boxes = outputs
logits_filt = np.squeeze(logits, 0) # [0] # prediction_logits.shape = (nq, 256)
logits_filt = self.sig(logits_filt)
boxes_filt = np.squeeze(boxes, 0) # [0] # prediction_boxes.shape = (nq, 4)
# filter output
if token_spans is None:
filt_mask = logits_filt.max(axis=1) > self.box_threshold
logits_filt = logits_filt[filt_mask] # num_filt, 256
boxes_filt = boxes_filt[filt_mask] # num_filt, 4
# get phrase
tokenlizer = self.net.tokenizer
tokenized_raw_results = tokenlizer(caption, return_tensors="np")
tokenized = {
"input_ids": tokenized_raw_results.input_ids[0].astype(np.int64),
"token_type_ids": tokenized_raw_results.token_type_ids[0].astype(np.int64),
"attention_mask": tokenized_raw_results.attention_mask[0],
}
# build pred
pred_phrases = []
for logit in logits_filt:
posmap = logit > self.text_threshold
pred_phrase = self.get_phrases_from_posmap(posmap, tokenized, tokenlizer)
if with_logits:
pred_phrases.append([pred_phrase.strip(), logit.max()])
else:
pred_phrases.append([pred_phrase.strip(), 1.0])
else:
# TODO: Using token_spans.
raise NotImplementedError
return boxes_filt, pred_phrases這樣應該可以解決問題。 |
大佬您好，有个bug麻烦您过目。
我在使用GoundingDINO（SwinB-QInt8）IDEA这个模型进行人的头部自动标注，在
__preferred_device__ = "CPU"情况下的结果，可以标注且不会报错。我已经配置好了GPU运行的环境，并且在DAMO_YOLO等其它模型上也可以正常用GPU运行，但在GroundingDINO上使用GPU运行时，会出现以下报错：
在加载模型时：
� [ 0 ; 9 3 m 2 0 2 4 - 0 4 - 2 5 0 9 : 1 9 : 0 9 . 4 4 8 6 1 0 8 [ W : o n n x r u n t i m e : , t r a n s f o r m e r _ m e m c p y . c c : 7 4 o n n x r u n t i m e : : M e m c p y T r a n s f o r m e r : : A p p l y I m p l ] 1 2 6 4 M e m c p y n o d e s a r e a d d e d t o t h e g r a p h t o r c h _ j i t f o r C U D A E x e c u t i o n P r o v i d e r . I t m i g h t h a v e n e g a t i v e i m p a c t o n p e r f o r m a n c e ( i n c l u d i n g u n a b l e t o r u n C U D A g r a p h ) . S e t s e s s i o n _ o p t i o n s . l o g _ s e v e r i t y _ l e v e l = 1 t o s e e t h e d e t a i l l o g s b e f o r e t h i s m e s s a g e . � [ m � [ 0 ; 9 3 m 2 0 2 4 - 0 4 - 2 5 0 9 : 1 9 : 0 9 . 6 4 0 2 0 1 8 [ W : o n n x r u n t i m e : , s e s s i o n _ s t a t e . c c : 1 1 6 6 o n n x r u n t i m e : : V e r i f y E a c h N o d e I s A s s i g n e d T o A n E p ] S o m e n o d e s w e r e n o t a s s i g n e d t o t h e p r e f e r r e d e x e c u t i o n p r o v i d e r s w h i c h m a y o r m a y n o t h a v e a n n e g a t i v e i m p a c t o n p e r f o r m a n c e . e . g . O R T e x p l i c i t l y a s s i g n s s h a p e r e l a t e d o p s t o C P U t o i m p r o v e p e r f . � [ m � [ 0 ; 9 3 m 2 0 2 4 - 0 4 - 2 5 0 9 : 1 9 : 0 9 . 6 4 0 5 9 2 9 [ W : o n n x r u n t i m e : , s e s s i o n _ s t a t e . c c : 1 1 6 8 o n n x r u n t i m e : : V e r i f y E a c h N o d e I s A s s i g n e d T o A n E p ] R e r u n n i n g w i t h v e r b o s e o u t p u t o n a n o n - m i n i m a l b u i l d w i l l s h o w n o d e a s s i g n m e n t s . � [ m在自动标注时：
Error in predict_shapes: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Expand node. Name:'Expand_33526' Status Message: Expand_33526: left operand cannot broadcast on dim 2 LeftShape: {1,900,4}, RightShape: {1,900,256}从而无法正常标注。我没有找到解决该问题的有效途径，恳请指点
The text was updated successfully, but these errors were encountered: