Re-training on the Cat/Dog Dataset

I encountered two problems in the process of retraining:

  1. During the training process, the accuracy of the model that is trained again until the end of 35 trainings is lower than that of the model that has been directly trained 35 times.
  2. The trained model has acceptable accuracy when testing cat and dog pictures, but the real-time camera output is wrong, and putting a cat picture in front of the camera shows the structure as a dog.

Hi, what is the accuracy that you get after 35 epochs?

Note that the model I provided was trained for 100 epochs I believe.

The cat-dog dataset was kept intentionally small to keep the training time down on Nano, but if desired you can read how to enlarge it here:

I also used the 100 epochs model, and the picture test did indeed work better, but the camera input is still indistinguishable from cats and dogs.

The phenomenon of Re-training on the PlantCLEF Dataset is consistent with the phenomenon of Re-training on the cat/dog Dataset.

Hi Hope, are you putting a picture in front of the camera? In my experience it works better from camera when looking at an actual object. For example, some testing I did from camera around my house:

Note that the datasets for these models are intentionally kept small to keep the training time down for Nano, so while they do work ok for me, they aren’t intended to be production-quality models. With the cat/dog example, if desired you can expand the dataset to be larger.

Hi I am having similar training issues to [] on both the “baseline” Cat_dog and Plant models.
My answers are far off i.e for
cat_dog model
Subsequently I downloaded the the 100 epoch model for cat_dog.
:dusty_nv = 94.825% cat, I get an answer of 70.01% dog, :-(wrong species and %
plant model
Also downloaded the 100 epoch model for plant
:dusty_nv = 83.10% elm, I get an answer of 25.64 % cedar, :-( wrong species and %
Clearly I have something totally wrong, any help is sincerely appreciated
I have the a nano model B01(latest) using latest Jetpack and pytorch, using python3.
Many thanks…

Problem solved: All result now match the results quoted in the examples for cat_dog and plant models.
One needs to delete the cached engine: resnet18.onnx.1.1.7103.GPU.FP16.engine which is stored in the models directory. It is automatically recreated and then works perfectly… :-)