Export Problems

[alex-itico]

Search before asking

• I have searched the AnomaVision issues and found no similar bug report

Component

Export

Bug Description

I'm trying to export ONNX model or Engine by CLI. But ONNX-RunTime library had issues with the data. Also, when I try to export format=engine didn't anything the function.

anomavision_20260626_171850.log
anomavision_20260626_172025.log

Operating System

Ubuntu / Debian / Linux

Python Version

3.10.12

Package Manager Used

uv (e.g., uv pip install)

Target Hardware / Extra Bracket

Custom / Other

Minimal Reproducible Example

Adjunct the requirements of the ambient. Also, my config.yml used for Export
Format = "onnx" / "engine" --> I used the two formats with the same configExport.yml

configExport.yml
requirements.txt

Additional Context

No response

Contribution

• I'm willing to submit a PR to fix this issue

[github-actions]

👋 Hello @alex-itico, thank you for opening this issue! Welcome to the AnomaVision community.

Our triage bot has categorized this as a Bug Report. The engineering team will review your environment and hardware trace shortly.

Please hold tight while we route this to the right team member. 🚀

[DeepKnowledge1]

Thanks for reporting this and for providing the details.

Regarding the engine export, I haven't been able to reproduce or investigate that part because I don't have access to an NVIDIA GPU. TensorRT engine generation requires a compatible GPU, so I can't verify whether the issue is with the export process itself.

I'll focus on reproducing and fixing the ONNX export/ONNX Runtime issue first, since that's something I can test in my current environment. Once I have access to suitable GPU hardware (or if someone in the community can help test), I'll also investigate the engine export behavior.

Thanks again for the report and your patience!

[DeepKnowledge1]

@alex-itico the export is woking,
can you please give the console output log of the export

[alex-itico]

command: uv run anomavision export --config configExport.yml

yml configuration:

class_name: sellos_v2
run_name: sello_perfecto
algorithm: padim
backbone: resnet18
feat_dim: 100
layer_indices:

• 0
• 1
• 2
  resize:
• 160
• 160
  crop_size: null
  normalize: true
  norm_mean:
• 0.485
• 0.456
• 0.406
  norm_std:
• 0.229
• 0.224
• 0.225

model_data_path: "/home/meterbee_nvidia/Escritorio/anomavision/distributions/padim/sellos_v2/sello_perfecto"
dataset_path: "/home/meterbee_nvidia/Escritorio/anomavision/dataset/"
model: "model.pt"

format: "onnx"
precision: "fp16"
opset: 17
static_batch: true

device: "cuda"
log_level: "INFO"

OutPut Log :

🌿 Checking branch: UNKNOWN
⚠️ Branch 'UNKNOWN' does not exist locally.

🌿 Checking branch: main
⚠️ Branch 'main' does not exist locally.
2026-06-30 09:46:10.182126583 [W:onnxruntime:Default, device_discovery.cc:164 DiscoverDevicesForPlatform] GPU device discovery failed: device_discovery.cc:89 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor"
2026-06-30 09:46:10,386 - anomavision - INFO - Anomavision logging initialized - Level: INFO
2026-06-30 09:46:10,386 - anomavision - INFO - Log file: logs/anomavision_20260630_094610.log
2026-06-30 09:46:10,386 - anomavision.export - INFO - === export start === model=distributions/padim/sellos_v2/sello_perfecto/model.pt fmt=onnx in_shape=(1, 3, 160, 160) dyn_batch=True opset=18 precision=auto device=cuda opt=False
2026-06-30 09:46:10,387 - anomavision.export - INFO - Auto-detected device: CUDA
2026-06-30 09:46:10,387 - anomavision.export - INFO - load: distributions/padim/sellos_v2/sello_perfecto/model.pt
2026-06-30 09:46:10,527 - anomavision.export - INFO - Loaded object type: <class 'anomavision.padim.Padim'>
2026-06-30 09:46:10,528 - anomavision.export - INFO - GOING INTO FULL MODEL PATH
2026-06-30 09:46:10,781 - anomavision.export - INFO - Using FP16 precision (auto-detected for CUDA)
/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/export.py:293: UserWarning: Exporting a model while it is in training mode. Please ensure that this is intended, as it may lead to different behavior during inference. Calling model.eval() before export is recommended.
torch.onnx.export(
/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/export.py:293: UserWarning: # 'dynamic_axes' is not recommended when dynamo=True, and may lead to 'torch._dynamo.exc.UserError: Constraints violated.' Supply the 'dynamic_shapes' argument instead if export is unsuccessful.
torch.onnx.export(
W0630 09:46:13.424000 3516 torch/onnx/_internal/exporter/_schemas.py:455] Missing annotation for parameter 'input' from (input, rois, spatial_scale: 'float', pooled_height: 'int', pooled_width: 'int', sampling_ratio: 'int' = -1, aligned: 'bool' = False). Treating as an Input.
W0630 09:46:13.424000 3516 torch/onnx/_internal/exporter/_schemas.py:455] Missing annotation for parameter 'rois' from (input, rois, spatial_scale: 'float', pooled_height: 'int', pooled_width: 'int', sampling_ratio: 'int' = -1, aligned: 'bool' = False). Treating as an Input.
W0630 09:46:13.425000 3516 torch/onnx/_internal/exporter/_schemas.py:455] Missing annotation for parameter 'input' from (input, boxes, output_size: 'Sequence[int]', spatial_scale: 'float' = 1.0). Treating as an Input.
W0630 09:46:13.425000 3516 torch/onnx/_internal/exporter/_schemas.py:455] Missing annotation for parameter 'boxes' from (input, boxes, output_size: 'Sequence[int]', spatial_scale: 'float' = 1.0). Treating as an Input.
/usr/lib/python3.10/copyreg.py:101: FutureWarning: isinstance(treespec, LeafSpec) is deprecated, use isinstance(treespec, TreeSpec) and treespec.is_leaf() instead.
return cls.new(cls, *args)
2026-06-30 09:46:19,159 - anomavision.export - INFO - onnx: ok (8.77s) file=distributions/padim/sellos_v2/sello_perfecto/model.onnx size=0.1MB dynamic_batch=True opset=18 precision=FP16 device=cuda
2026-06-30 09:46:19,160 - anomavision.export - INFO - === export done in 8.77s ===

[alex-itico]

My fault the exportation its okey. The issue its with precision fp16 and CUDA because when I export in format fp32 and detect with CPU the model works. But with ONNX fp16 cuda had problems because doesn't find Tensorrt in the environment.

[alex-itico]

My version of ONNX-Runtime does not support CUDA/TensorRT!

[alex-itico]

Well @DeepKnowledge1 I fixed the library but now I export in format .engine, when I try to detect the CLI output log shows this:

2026-06-30 11:32:57,536 - anomavision - INFO - Anomavision logging initialized - Level: INFO
2026-06-30 11:32:57,537 - anomavision - INFO - Log file: logs/anomavision_20260630_113257.log
2026-06-30 11:32:57,537 - anomavision.detect - INFO - Streaming mode: False
2026-06-30 11:32:57,537 - anomavision.detect - INFO - Image processing: resize=(160, 160), crop=None, norm=True
2026-06-30 11:32:57,654 - anomavision.detect - INFO - Dataset path: /home/meterbee_nvidia/Escritorio/anomavision/fotos_buenas
2026-06-30 11:32:57,655 - anomavision.detect - INFO - Device: cuda
2026-06-30 11:32:57,655 - anomavision.detect - INFO - Loading model: /home/meterbee_nvidia/Escritorio/anomavision/distributions/padim/sellos_v2/sello_perfecto/model.engine
2026-06-30 11:32:57,655 - anomavision.inference.model.wrapper - INFO - Initializing ModelWrapper with /home/meterbee_nvidia/Escritorio/anomavision/distributions/padim/sellos_v2/sello_perfecto/model.engine on cuda
2026-06-30 11:32:57,656 - anomavision.inference.model.wrapper - INFO - Creating backend for model: /home/meterbee_nvidia/Escritorio/anomavision/distributions/padim/sellos_v2/sello_perfecto/model.engine
2026-06-30 11:32:57,656 - anomavision.inference.model.wrapper - INFO - Detected model type: ModelType.TENSORRT
2026-06-30 11:32:57,656 - anomavision.inference.model.backends.tensorrt_backend - WARNING - TensorRT backend is not implemented.
2026-06-30 11:32:57,656 - anomavision.detect - ERROR - Failed to load model: TensorRT support not implemented yet.
2026-06-30 11:32:57,656 - anomavision.detect - ERROR - Process failed: TensorRT support not implemented yet.
Traceback (most recent call last):
File "/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/detect.py", line 524, in main
metrics, results = run_inference(args)
File "/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/detect.py", line 281, in run_inference
model = ModelWrapper(model_path, device_str)
File "/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/inference/model/wrapper.py", line 112, in init
self.backend: InferenceBackend = make_backend(model_path, device)
File "/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/inference/model/wrapper.py", line 89, in make_backend
return TensorRTBackend(model_path, device)
File "/home/meterbee_nvidia/Escritorio/anomavision/.venv/lib/python3.10/site-packages/anomavision/inference/model/backends/tensorrt_backend.py", line 39, in init
raise NotImplementedError("TensorRT support not implemented yet.")
NotImplementedError: TensorRT support not implemented yet.

[DeepKnowledge1]

Impressive progress