Giters

SthPhoenix / InsightFace-REST

InsightFace REST API for easy deployment of face recognition services with TensorRT in Docker.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

scrfd_ * shape issue

gulldan opened this issue · comments

commented

Hello,

for models scrfd_* if i try out "test_images/Stallone.jpg"

docker run -p 18080:18080 -it --gpus '"device=0"' -e LOG_LEVEL=INFO -e PYTHONUNBUFFERED=0 -e
NUM_WORKERS=1 -e INFERENCE_BACKEND=trt -e FORCE_FP16=True -e DET_NAME=scrfd_10g_gnkps -e DET_THRESH=0.6 -e REC_NAME=glint36
0k_r100FC_1.0 -e REC_IGNORE=False -e REC_BATCH_SIZE=1 -e GA_NAME=genderage_v1 -e GA_IGNORE=False -e KEEP_ALL=True -e MAX_SIZE
=1024,780 -e DEF_RETURN_FACE_DATA=True -e DEF_EXTRACT_EMBEDDING=True -e DEF_EXTRACT_GA=True -e DEF_API_VER='1' --mount type=b
ind,source=/home/work/services/models,target=/models --health-cmd='curl -f http://localhost:18080/info || exit 1' --health-i
nterval=1m --health-timeout=10s --health-retries=3 insightface-rest
Preparing models...
mxnet version: 1.8.0
onnx version: 1.7.0
[08:23:40] INFO - Preparing 'glint360k_r100FC_1.0' model...
[08:23:40] INFO - Preparing 'scrfd_10g_gnkps' model...
[08:23:40] INFO - Reshaping ONNX inputs to: (1, 3, 780, 1024)
[08:23:40] INFO - Building TRT engine for scrfd_10g_gnkps...
[08:23:52] INFO - Building TensorRT engine with FP16 support.
[TensorRT] WARNING: /workspace/TensorRT/t/oss-cicd/oss/parsers/onnx/onnx2trt_utils.cpp:227: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[TensorRT] WARNING: No implementation obeys reformatting-free rules, at least 10 reformatting nodes are needed, now picking the fastest path instead.
[08:24:46] INFO - Building TRT engine complete!
[08:24:46] INFO - Preparing 'genderage_v1' model...
Starting InsightFace-REST using 1 workers.
mxnet version: 1.8.0
onnx version: 1.7.0
[08:24:49] INFO - Warming up face detector TensorRT engine...
[08:25:02] INFO - Engine warmup complete! Expecting input shape: (1, 3, 780, 1024)
[08:25:02] INFO - Warming up ArcFace TensorRT engine...
[08:25:03] INFO - Engine warmup complete! Expecting input shape: (1, 3, 112, 112). Max batch size: 1
[08:25:03] INFO - Warming up GenderAge TensorRT engine...
[08:25:03] INFO - Engine warmup complete! Expecting input shape: (1, 3, 112, 112). Max batch size: 1
INFO: Started server process [233]
[08:25:03] INFO - Started server process [233]
INFO: Waiting for application startup.
[08:25:03] INFO - Waiting for application startup.
INFO: Application startup complete.
[08:25:03] INFO - Application startup complete.
INFO: Uvicorn running on http://0.0.0.0:18080 (Press CTRL+C to quit)
[08:25:03] INFO - Uvicorn running on http://0.0.0.0:18080 (Press CTRL+C to quit)
INFO: 127.0.0.1:35790 - "GET /info HTTP/1.1" 200 OK
INFO: 192.168.49.168:65047 - "GET /docs HTTP/1.1" 200 OK
INFO: 192.168.49.168:65047 - "GET /openapi.json HTTP/1.1" 200 OK
inference cost: 0.0219266414642334
Traceback (most recent call last):
File "./modules/processing.py", line 236, in embed
faces = await self.model.get(image, max_size=max_size, threshold=threshold,
File "./modules/face_model.py", line 189, in get
boxes, probs, landmarks, mask_probs = self.det_model.detect(img.transformed_image, threshold=threshold)
File "./modules/face_model.py", line 40, in detect
bboxes, landmarks = self.retina.detect(data, threshold=threshold)
File "./modules/model_zoo/detectors/scrfd.py", line 173, in detect
scores_list, bboxes_list, kpss_list = self.forward(img, threshold)
File "./modules/model_zoo/detectors/scrfd.py", line 159, in forward
bboxes = distance2bbox(anchor_centers, bbox_preds)
File "./modules/model_zoo/detectors/scrfd.py", line 39, in distance2bbox
x1 = points[:, 0] - distance[:, 0]
ValueError: operands could not be broadcast together with shapes (24832,) (25088,)

what could be a solution?
NVIDIA-SMI 465.19.01
Driver Version: 465.19.01
CUDA Version: 11.3
NVIDIA Tesla T4

if change to retinaface_r50_v1 everything works fine.

i try to change base container to nvcr.io/nvidia/tensorrt:21.06-py3 but nothing changes

Preparing models...
mxnet version: 1.8.0
onnx version: 1.7.0
[07:53:24] INFO - Preparing 'glint360k_r100FC_1.0' model...
[07:53:24] INFO - Preparing 'scrfd_10g_gnkps' model...
[07:53:24] INFO - Reshaping ONNX inputs to: (1, 3, 780, 1024)
[07:53:24] INFO - Building TRT engine for scrfd_10g_gnkps...
[07:53:36] INFO - Building TensorRT engine with FP16 support.
[TensorRT] WARNING: /home/jenkins/agent/workspace/OSS/OSS_L0_MergeRequest/oss/parsers/onnx/onnx2trt_utils.cpp:271: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_130 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_177 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_141 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_224 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_188 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_152 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_235 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_199 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation of layer InstanceNormalization_246 obeys the requested constraints in strict mode. No conforming implementation was found i.e. requested layer computation precision and output precision types are ignored, using the fastest implementation.
[TensorRT] WARNING: No implementation obeys reformatting-free rules, at least 10 reformatting nodes are needed, now picking the fastest path instead.
[07:54:32] INFO - Building TRT engine complete!
[07:54:32] INFO - Preparing 'genderage_v1' model...
Starting InsightFace-REST using 1 workers.
mxnet version: 1.8.0
onnx version: 1.7.0
[07:54:36] INFO - Warming up face detector TensorRT engine...
[07:54:48] INFO - Engine warmup complete! Expecting input shape: (1, 3, 780, 1024)
[07:54:48] INFO - Warming up ArcFace TensorRT engine...
[07:54:49] INFO - Engine warmup complete! Expecting input shape: (1, 3, 112, 112). Max batch size: 1
[07:54:49] INFO - Warming up GenderAge TensorRT engine...
[07:54:49] INFO - Engine warmup complete! Expecting input shape: (1, 3, 112, 112). Max batch size: 1
INFO:     Started server process [233]
[07:54:49] INFO - Started server process [233]
INFO:     Waiting for application startup.
[07:54:49] INFO - Waiting for application startup.
INFO:     Application startup complete.
[07:54:49] INFO - Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:18080 (Press CTRL+C to quit)
[07:54:49] INFO - Uvicorn running on http://0.0.0.0:18080 (Press CTRL+C to quit)
INFO:     192.168.49.168:64156 - "GET /docs HTTP/1.1" 200 OK
INFO:     192.168.49.168:64156 - "GET /openapi.json HTTP/1.1" 200 OK
inference cost: 0.02373218536376953
Traceback (most recent call last):
  File "./modules/processing.py", line 236, in embed
    faces = await self.model.get(image, max_size=max_size, threshold=threshold,
  File "./modules/face_model.py", line 189, in get
    boxes, probs, landmarks, mask_probs = self.det_model.detect(img.transformed_image, threshold=threshold)
  File "./modules/face_model.py", line 40, in detect
    bboxes, landmarks = self.retina.detect(data, threshold=threshold)
  File "./modules/model_zoo/detectors/scrfd.py", line 173, in detect
    scores_list, bboxes_list, kpss_list = self.forward(img, threshold)
  File "./modules/model_zoo/detectors/scrfd.py", line 159, in forward
    bboxes = distance2bbox(anchor_centers, bbox_preds)
  File "./modules/model_zoo/detectors/scrfd.py", line 39, in distance2bbox
    x1 = points[:, 0] - distance[:, 0]
ValueError: operands could not be broadcast together with shapes (24832,) (25088,)

layer InstanceNormalization was fix in new TRT but there is no prebuild nvidia container yet

commented

Hi! For SCRFD family models input dimensions should be divisible by 32.
deepinsight/insightface#1578

So you should just change 1024,780 to 1024,768

commented

Hi! For SCRFD family models input dimensions should be divisible by 32.
deepinsight/insightface#1578

oh ,thanks

commented

Actually I was struggling with same issue when SCRFD was only released, I think I should mention this in readme or dockerfile comments