mnist | Real-time digit classifier trained on MNIST | Machine Learning library

You could start off with a custom training loop using tf.GradientTape:

Try using plt.imsave to save each image separately:

Based on the paper you shared, it looks like you need to change the weight arrays per each output neuron per each layer. Unfortunately, this means that the implementation of your optimization routine is going to depend on the layer type, since an "output neuron" for a convolution layer is quite different than a fully-connected layer. In other words, just looping over Flux.params(model) is not going to be sufficient, since this is just a set of all the weight arrays in the model and each weight array is treated differently depending on which layer it comes from.

Fortunately, Julia's multiple dispatch does make this easier to write if you use separate functions instead of a giant loop. I'll summarize the algorithm using the pseudo-code below:

when you use cmap="gray" in plt.imshow() you must either unscale your output or set vmin and vmax. From what I see you scaled by dividing 255, so you must multiply your data by 255 or, alternativle set vmin=0, vmax=1 Option1:

Gradient tape triggers automatic differentiation which requires tracking gradients on all your weights and activations. Autodiff requires multiple more memory. This is normal. You'll have to manually tune your batch size until you find one that works, then tune your LR. Usually, the tune just means guess & check or grid search. (I am working on a product to do all of that for you but I'm not here to plug it).

So if you look at the source code for tensorly linked here you can see that the documentation for the function in question partial_tucker says:

If the hyperspectral dataset is given to you as a large image with many channels, I suppose that the classification of each pixel should depend on the pixels around it (otherwise I would not format the data as an image, i.e. without grid structure). Given this assumption, breaking up the input picture into 1x1 parts is not a good idea as you are loosing the grid structure.

I further suppose that the order of the channels is arbitrary, which implies that convolution over the channels is probably not meaningful (which you however did not plan to do anyways).

Instead of reformatting the data the way you did, you may want to create a model that takes an image as input and also outputs an "image" containing the classifications for each pixel. I.e. if you have 10 classes and take a (145, 145, 200) image as input, your model would output a (145, 145, 10) image. In that architecture you would not have any fully-connected layers. Your output layer would also be a convolutional layer.

That however means that you will not be able to keep your current architecture. That is because the tasks for MNIST/CIFAR10 and your hyperspectral dataset are not the same. For MNIST/CIFAR10 you want to classify an image in it's entirety, while for the other dataset you want to assign a class to each pixel (while most likely also using the pixels around each pixel).

Some further ideas:

• If you want to turn the pixel classification task on the hyperspectral dataset into a classification task for an entire image, maybe you can reformulate that task as "classifying a hyperspectral image as the class of it's center (or top-left, or bottom-right, or (21th, 104th), or whatever) pixel". To obtain the data from your single hyperspectral image, for each pixel, I would shift the image such that the target pixel is at the desired location (e.g. the center). All pixels that "fall off" the border could be inserted at the other side of the image.
• If you want to stick with a pixel classification task but need more data, maybe split up the single hyperspectral image you have into many smaller images (e.g. 10x10x200). You may even want to use images of many different sizes. If you model only has convolution and pooling layers and you make sure to maintain the sizes of the image, that should work out.

I was able to reproduce this bug too. It seems to be related to the plt.tight_layout() that you apply within the loop. Instead of doing this, use plt.subplots to produce the axes objects first, then iterate over those instead. Once you plot everything, use tight_layout on the opened figure:

Referring to one of the most recent TfLite android app examples might help: Model Personalization App. This demo app uses transfer learning model instead of LSTM, but the overall workflow should be similar.

As Farmaker mentioned in the comment, try using SNAPSHOT in the gradle dependency: