search

JasonDCox / ML-Mentorship-GovSchool

0 stars 0 forks source link

Create working General Tensorflow model on Desktop #28

Closed brandonC1234 closed 2 years ago

brandonC1234 commented 2 years ago

Description

This involves setting up Tensorflow gpu, Cuda/cuDNN, finding a dataset for general animal detection, compiling it to a usable tensorFlow record, and training the model. Follow the video in the description for help.

Acceptance Criteria

References

brandonC1234 commented 2 years ago

Progress:

I was able to download a cats and dog dataset from Oxford in Tensorflow record format, but I did have to change the extension from .tfrecord to .record.

The main issue was creating a correct label map, but since there are only 2 classes (rather than the 37 in the breed dataset), this wasn't a huge pain.

Performance

ABOSULUTELY NO FALSE POSITIVES, however, I'm rarely getting a confidence of over 70% on the test images

Improvements

After getting low confidence after training for ~15,000 steps, I'm pretty sure that the model I'm using is simply too low resolution. I'm currently using 320x320 mobile ssd, but I'm going to test the higher resolution and potentially a slower but more accurate model as soon as I can.

brandonC1234 commented 2 years ago

Conclusion

After upgrading to the 640x640 mobile ssd model, I was able to achieve acceptable accuracies and a low rate of false positives. The frame rate may be questionable, but that is something to investigate later.