Training with OCRNet 2.0 - Looking at old 1.0 jupyter notebook example

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hi everyone

I have configured OCRNet with TAO from this Jupyter notebook

tao_tutorials/notebooks/tao_launcher_starter_kit/ocrnet/ocrnet.ipynb at main · NVIDIA/tao_tutorials · GitHub

The notebook is using the version 1 of OCRNet, where the newest you can download is 2.1

I am using OCRNet trainable 2.0 and download it with ngc.
I also had to change the feature_channel: 512 tofeature_channel: 192
Running in nvcr.io/nvidia/tao/tao-toolkit:5.5.0-pyt

The notebook has the default epoch to be 10 which is fine verifying.

I am now training on the ICDAR15 dataset and I am seeing low accuracy. I am now running an experiment on 2000 epochs.

What are you others experience? What are good defaults? I tried searching on the forum and it there was not a lot of people talking about OCRNet with TAO 😄

2

I had better luck with going back to trainable_v1.0

I was stuck on accuracy with 0.05 with v2, with v1 I am starting with 0.78

When I was trying v2, it was complaining about feature_channel, saying 192 instead of 512
I also changed the input size since it is half the size for v1 than v2. 200x64x1 vs 100x32x1

3

For using OCRNet trainable 2.0 model, please see vit notebook: tao_tutorials/notebooks/tao_launcher_starter_kit/ocrnet/ocrnet-vit.ipynb at main · NVIDIA/tao_tutorials · GitHub.

1 Like
4

Thank you, that makes more sense now looking it.

It was a bit confusing when I looked at it first, but now with my experience it makes more sense. Could the README be updated to say where to look for what? :)

5

Yes, we can add some Notes/reminder into each notebook. Thanks for the suggestion.

1 Like