Working with .blob file in python

nikocreatina

Good Morning i would like to have an understanding of this output i have when i run my model using the blob file. I saved the onnx file from a TensorFlow model and specified the shape as such [1, 15, 224, 224, 3]. By using this shape i was able to convert the onnx into blob file. I begun to work with this file and tried to implemented to run the inference. My task is video classification, with just 2 categories on the whole frame so not object detection or face/person detection. When i run the python code i got the following output looping and referring to the size of my nn being no the right shape. However it seems odd cause it give me the shape as [3, 224] and it seems that is taking the 3 which 9is referred to RGB channel as one shape. what is a possible solution? I would like to point out that this screenshot refers to shape [300,300], even by changing this to [224,224] will output the same error.

Over this section it talks about the shape arrangement. however how would i rearrange this? Does this apply also to my custom model?

jakaskerl

Hi nikocreatina
https://docs.luxonis.com/en/latest/pages/model_conversion/#model-layout-parameter looks like a different layout.
You need to specify it using --layout NCHW:NHWC (i think).

Thanks,
Jaka

nikocreatina

jakaskerl i Will try wit the layout option and I will also try the other opinion using the API directly like mentioned in this piece of code just after.

import depthai as dai
pipeline = dai.Pipeline()
colorCam = pipeline.createColorCamera()
colorCam.setInterleaved(False) # False = Planar

I will let you know which one solve the problem

jakaskerl

nikocreatina
The interleaved must be false anyway I think, otherwise you get weird results. Also, it speeds up image manipulations.

nikocreatina

jakaskerl

Inputs

Name: x, Type: DataType.U8F, Shape: [3, 224, 224, 15, 1] (StorageOrder.???)

Outputs

Name: output_node, Type: DataType.FP16, Shape: [2, 1] (StorageOrder.NC)

I just double checked the input shape and it is as follow. My onnx file definetly have an input of 1,15,224,224,3. 15 is the number of frames and 1 is the batches. Any ideas?

nikocreatina

Hi There i still got the problem even after following those steps.

jakaskerl

Hi nikocreatina
Since you are using tensorflow, how about changing the model before trying to convert it? Your onnx file is NHWC, but should be NCHW - if I follow the docs. Could this help: https://stackoverflow.com/questions/37689423/convert-between-nhwc-and-nchw-in-tensorflow?

Thanks,
Jaka