Category: Misc

I am a novice to Tensorflow but have significant experience in the scikit-learn environment. I am trying to shift some codes to the more flexible TF/Keras routines in Python. One modification I would like to do is to build a simple neural network that trains on an earth mover’s distance/Wasserstein distance between arrays representing densities, rather than on direct root-mean-square differences in the densities.

I have been able to modify a loss function, but do not know how to implement the EMD in TF. I use pyEMD to calculate the EMD between two of the arrays in Python, but is doesn’t translate to tensors in TF. It seems like I may have to “unstack” the tensors, evaluate by each row, then “restack” as a tensor. It doesn’t seem reasonable to do it this way, but I am willing to make it work. At this point, I am focused on making it work rather than be as-fast-as-possible.

Does anyone have advice on either a TF-friendly EMD approach or how I may use pyEMD in a custom loss function? I appreciate any and all suggestions.

For example I can have the last layer of my network be tf.keras.layers.Dense(10) even if my labels are just single values, so the y is shape (500,)

The last layer outputs 10 values but there is only one label per input vector.

How can this work? Which of the 10 values is compared to the label to calculate the loss? I expected keras to give me an error in this case but it trains just fine and even does well on the toy dataset I’m using.

So I have been trying at this for ages now, my GPU is GTX 1650 SUPER and I have Python 3.9.5 with tensorflow version 2.7.0 and CUDA 11.2, cuDNN 8.1.0. Here is the issue: sometimes, my models train but most of the time it errors with:

E tensorflow/stream_executor/cuda/cuda_event.cc:29] Error polling for event status: failed to query event: CUDA_ERROR_ILLEGAL_ADDRESS: an illegal memory access was encountered 

or some other variants (again, they kinda take turns fighting to be errors):

E tensorflow/stream_executor/cuda/cuda_event.cc:29] Error polling for event status: failed to query event: CUDA_ERROR_MISALIGNED_ADDRESS: misaligned address E tensorflow/stream_executor/cuda/cuda_event.cc:29] Error polling for event status: failed to query event: CUDA_ERROR_ILLEGAL_INSTRUCTION: an illegal instruction was encountered 

and I have absolutely no idea why its doing this. I don’t think its my code because I was able to train my code 100% of the time on CPU, but it takes like half an hour to do that and GPUs take only 3 minutes, 10 times faster!!! Who else has this error and why is it occuring randomly, also: could MSI Afterburner mess with it?

I trained a custom image classification model using Python and now want to deploy it on a native application which I’ve made using React Native. My initial approach to this was to use tfjs. However, tfjs has poor support, unresolved complaints on the forums and a ton of dependency issues when it comes to using custom models on react native, from my experience. I couldn’t figure out any of those by myself. So I switched up and am now trying to load the model on a Python REST API, and am planning to call it from my React Native app. Is this the best way to approach this problem? I’m new to machine learning and programming overall in general, and would love to hear some insight. Is it sensible to upload an image from the React Native front end and then get the prediction using the model on the back end? Also if I’m wrong about the problems with tfjs and someone has successfully used it I’d love to hear how you do it, please. Any and everything helpful is much appreciated, thank you very much.