Add methods for model training

[timesler]

The repo currently only includes models (MTCNN and InceptionResnetV1). It would be good to add code for updating pretrained models given new data or different training hyperparameters.

[shaikhaman786vector]

How to train on my own custom dataset

[timesler]

@shaikhaman786vector thank you for the prompt, I've been dragging my feet on getting this in the repo. I've just updated the repo and python package to include some utilities for finetuning the models. If you want to train on your own dataset, first update the package to the latest version on PyPI (0.2.2):

pip install --upgrade facenet-pytorch

After running the above command, follow this notebook to finetune a model on your own dataset. Depending on your dataset you may need to alter training parameters (e.g., learning rate, optimizer, weight decay) or data augmentation (e.g., random resizing, random horizontal flip, etc.).

If you get a chance, please let me know if you get the model training well. The finetuning notebook works well for me, but hasn't been tested much.

[shaikhaman786vector]

it will be good if you upload for triplet loss

[shaikhaman786vector]

Also if there will be code for recognizing new image, this will be complete flow for face recognition

[timesler]

@shaikhaman786vector I'm not sure what you mean by "recognizing new image". Do you mean an image for a person that was in the training set (in which case you can just pass the image tensor to the model), or an image of a new person?

There is an example showing how to infer using a trained model in the repo (examples/infer.ipynb).