Enquiry of the debug method or alternatives to solve the inaccuracy

[sangwong416]

Hi,

I am using python program in jetson-inference to process the classification of 3 letters, i.e. A, B & C. DataSet are from “https://s3.amazonaws.com/nist-srd/SD19/by_class.zip”. I followed the ratio, 7:2:1, to place the images of A, B & C into train, val & test floders respectively.

Model was trained, exported and executed by the below commands.

1. python3 train.py --model-dir=models/abc --batch-size=4 --workers=1 --epochs=1 data/abc
2. python3 onnx_export.py --model-dir=models/abc
3. imagenet --model=models/abc/resnet18.onnx --labels=data/abc/labels.txt --input_blob=input_0 --output_blob=output_0 /dev/video0

The Top-5 and Top-1 % are 100% and 96% respectively after the testing.

During running the model, it seems that “A” become the default class and it can find class “B”. However, class “C” cannot be classified.

May I ask how I can debug? or find out the reasons of inaccuracy?

Thank you so much!

Regards,
Anthony

[dusty_nv]

Hi @sangwong416, have you tried running the imagenet program on test images from your training dataset to confirm that it’s able to classify the letters? Perhaps the difference is related to the imagery coming from your camera as opposed to what the dataset consists of.