BradLarson
commented
5 years ago For TensorFlow weights, you can use Raw.restoreV2() to pull weights from a checkpoint file. I believe an example of that can be found in here:
https://github.com/tensorflow/swift-models/blob/master/Transformer/PythonCheckpointReader.swift
for the weights here, I'd exported them from a trained Caffe model via a model loader that I've not included in the repository. I'm working on a more robust way to pull from .caffemodel files, as well as SavedModel and other protobuf-based files. Haven't had the time to complete that, though.
johndpope
In swift-models repo - there's a question about creating deep dream sample. https://github.com/tensorflow/swift-models/issues/157
Looking through another repo - there's some code here that uses inception 5h trained model. https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/14_DeepDream.ipynb
It's not apparently obvious in swift how to load a pre-trained (python tensorflow) model - but I'm thinking with your work here - maybe able to cherry pick the loading of weights.
ok - it seems this repo is re-doing all the training out of the box without leveraging any existing dataset.
https://www.mathworks.com/matlabcentral/fileexchange/59133-deep-learning-toolbox-model-for-alexnet-network The model has 23 layers and can classify images into 1000 object categories (e.g. keyboard, mouse, coffee mug, pencil).
I had imagined Alexnet was equivalent to MNIST - pre-trained data you could just load. Digging deeper - I found this - bvlc_alexnet.npy -- pretrained numpy weights for tensorflow http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/
This won't magically load I guess due to architecture mismatch?
maybe that's an enhancement to this project for you to consider. If we can get a swift model sample loading trained weights - then that would help others.