U-Net | An reimplementation U-Net on MXNet

I am trying to learn build a U-NET architecture from scratch. I have written this code but the problem is that when I try to run to check the output of the encoder part, I am having issues with it. When you the run the code below , you'll get

I'm trying to train a 1D CNN to identify specific parts of a text string.

The inputs are arrays of shape (128,1) containing 128 characters, and the aim is for the network to classify each of the characters into a particular class. For purposes of illustration, an input array could look like this:

I'm trying to implement the UNET at the keras website:

Image segmentation with a U-Net-like architecture

With only one change. use Dice loss instead of "sparse_categorical_crossentropy". However, every time I try something, I get different error. I'm coding on google colab using Tensorflow 2.7.

For example, I tried using

env:

This is something of a meta question, detached from code or libraries.

Let's say we have a large array of very primitive elements, like from an image read directly from our hard drive. The important thing is, that the representation of our input data is very simple (bits, bytes, 8 bit integers, along those lines), but there are lots of data points (instead of having 256 32-bit integers we may have 256x32=8192 bits). Let's assume that we do neither know how (or if) these simple data points correlate (i.e. whether they are shorts, integers, floats,...), so their encoding is unknown. I would like to train a neural network to interpret and decode these data points, but am struggling to think of a way to represent/format my input data. Since I don't know anything about the encoding (I do have labels of some sort for the input arrays, for example grayscale images) I'd imagine it difficult and prone to errors to assume structure within my data. It would for example be easy to calculate bytes from the bits and reduce my dimensionality by a factor of 8, but I think that might cause a loss of information from structure. For the same reason it's propably not suitable to split up the data into smaller batches.

I looked into this and found articles like this, but any of the proposed methods would most likely harm the integrity of my data.

I did some experiments combining bits into bytes but this still leaves very large inputs (because they are only reduced by a factor of 8) and did not yield the results I was hoping for. Another Idea I had was to feed the all of the inputs into a CNN to extract features and to propagate those through a RNN (which had reasonable success) but this doesn't scale too well either and harms the data's integrity as well. Another approach was to do something like U-Net where I detect features with a CNN and then propagate these features along with the original inputs into a RNN, but this leads to an explosion of complexity (> 30.000.000 parameters for a mere ~2500 input bits).

Looking forward to suggestions, and I hope that the problem is explained clearly.

I run a tensorflow u-net model without dropout (but BN) with a custom metric called "average accuracy". This is literally the section of code. As you can see, datasets must be the same as I do nothing in between fit and evaluate.

Google said

I am running a 3D U-net model for segmentation, and came across this error:

Detailed error:

I have a U-Net model with pretrained weights from an Auto-encoder, The Auto-encoder was built an image dataset of 1400 images. I am trying to perform semantic segmentation with 1400 labelled images of a clinical dataset. The model performs well with an iou_score=0.97 on my test image dataset, but when I try to test it on a random image outside my dataset, I get a very bad segmentation result. I don't understand the reason for it. Please review my code and suggest me where I was wrong.

Training on my dataset & labels :