To achieve image regression problem


I am trying to implement some image regression problem. Basically, I want to use MINST data as the dataset, input the image and the output is the corresponding number, maybe not necessarily the exact number, but some closed one.

In tensorflow, autokeras has the function ImageRegressor, which have the model that can achieve this ( I attached the model below).

My current idea so far is to “copy” the ResNet152 function to the PyTorch. However, it doesn’t train well. I am wondering whether there are some other alternative ways to achieve this. Thanks!!

MNIST images are very small in size (28x28), I don’t think using ResNet152 is a good idea. I would suggest to create your own model from scratch.