[BUG!!] - OPENVINO / LUXONIS OAKD
Hello everyone,
I am experiencing this problem with running DeeLabV3+ on oak-d lite with blob format.
I converted ONNX-> XML/BINOpenVino format and then compiled xml to blob format with openvino-2022.2 compiler.
when i run my application i got this problem and i didnt manage to get my input layer.
[NeuralNetwork(0)] [warning] Input image (224x224) does not match NN (3x224)
SebastianH Mine is more specificly [DetectionNetwork(1)] [warning] Input image (512x512) does not match NN (3x512)
Hi SebastianH ,
2 options:
- Preferred: Specify correct layout using openvino's model optimizer, using
--layout
, docs here. - Change frame layout with depthai, using
colorCamera.setInterleaved(bool)
I hope this helps!
Thanks, Erik
erik
I tryed using colorCamera.setInterleaved(True) but i didnt help i will try your first recommendation and let you know
erik What layout should the model have?
Hi SebastianH , Per docs, model should be Planar / CHW layout.
Hi erik
Will there be any performance issue using colorCamera.setInterleaved(True) instead of chaning the model?
@erik Hi Erik, I have setinterleaved(True) (HWC) and I set the model input and outpuat layout to NHWC and looks like the openvino compiler changed it to NCHW alone and i get this message now:
[NeuralNetwork(0)] [warning] Input image is interleaved (HWC), NN specifies planar (CHW) data order
any fix?
Nadi This means you need to set setinterleaved(False)
, so it's planar, which is what NN specifies.
@SebastianH no performance issues on the model (or minimal, simple conversion layer), but some nodes that process frames (eg ImageManip) are slower if they process interleaved frames (see docs here)
Thanks, Erik
I am not talking about the output layer, i am talking about getalllayers -> i expect it to return input and output layers
and now it returns only the output
[184430107165680F00] [3.3] [89.777] [NeuralNetwork(0)] [warning] Input image (300x300) does not match NN (3x300)
Model: MobileNetV2
I am getting the same error and I tried the above two options which you have mentioned, but no luck.
I am assuming that it is just a warning, will it affect in any way.
heebahsaleem
It will not work.
You need to change:
- either the Openvino version you are using (trial and error), they seem to change their default shapes between versions.
- when compiling the blob with blobconverter, use these https://docs.openvino.ai/2022.3/openvino_inference_engine_tools_compile_tool_README.html parameters (primarily
-il
) to set the desired shape.
Thanks,
Jaka
I'm having a similiar issue
[18443010D18F411300] [2.2] [9.560] [DetectionNetwork(1)] [warning] Input image (244x244) does not match NN (3x244)
My blob conversion:
blob_path = blobconverter.from_tf(
frozen_pb="./frozen_graph.pb",
data_type="FP16",
shaves=6,
optimizer_params=[
f"--input_shape=[1,{SHAPE},{SHAPE},3]"
]
)
I'm confused about how I should modify this blob_path to make sure it outputs the desired shape/right convention @jakaskerl
- Edited
jakaskerl
thank you for your reply.
- what is trial and error Openvino version? I am using openvino 2022.2 version.
- I am using blobconverter app to convert [https://blobconverter.luxonis.com/]. How to use the one you have mentioned?
FYI I am using TF model
Your help will be highly appreciated. TIA