Custom data classification model Hardly able to classify anything even after training the model after following S3E3 Training Image Classification Models.
I have followed the instructions exactly as told in S3E3, Jetson AI Fundamentals - S3E3 - Training Image Classification Models - YouTube video for training this classification resne18 model on a custom dataset, before trying my custom dataset I re-trained the cat-dog to check if everything is fine and it worked fine.
but now I have trained the model with a custom dataset consisting of 26 classes and around 24000 images, for 35 epochs and 16 batch sizes using the python3 train.py file in the repo. however even after training if i feed the model with an image from the test folder or run the stream from my csi port. it can hardly detect anything and accuracy is stuck at 8.05 % no matter what I show in the camera-stream or input as a image.
i’d also like to mention the jetson nano lost power mid-training at around 27 epochs, and i resumed the training using --resume
argument and it continued and finished all 35 epochs then now has saved the model to checkpoint.pth. but when I run the imagenet command does it use the modelbest.pth or checkpoint.pth.? I’m still suffering with the same issue.
I also studied the graph given in the repo it states that 30 epochs for a dataset of 5000 images gives the best accuracy. does that mean I have to change any hyper-parameters?
my directory management is also exactly as stated in this repo instructions. and i haven’t used the docker container i run it locally on my jetson nano itself.
Just want to confirm first. Do you use Jetson Nano or Jetson Orin Nano?
If Jetson Nano is used, we are going to move your topic to the corresponding board.
For training, do you have the training loss log can share with us?
Thanks.