Hello everyone, I have problem with multi output regression.
I want to find a model that makes an image,output, with a voltage,input.
However, because the image shape is too various, I want to make a model with
a train set that only deal with a single each pixel.
Than I want it to use at the various shape of image.
For example, I want to make a model that maps v to X that is a vector having a shape of (1,0,0,0), (0,1,0,0) ,(0,0,0,1).. example.
With that model I want to map the test v to x that is like (1,0,1,0).
Through linear activation function based learning, it suceeded.
However it is hard to make it in a more precise way with nonlinear activation function, like ReLU.
Actually it seems that ReLU tends to pick a single pixel of the desired test image.
So if there is some tips related to this problem, I want to ask some help.
Because I am not majoring in computer science, I need your help, please.
Thank you.
Do you really need regression? You can see an example of regression with one-hot representation in week 2 videos of this course https://www.coursera.org/learn/neural-networks-deep-learning/home/info (videos are free).
The difference in image shape can be treated by padding. Convolutional neural networks are designed to capture features regardless their position (e.g. eyes, nose, hand). You can find pretty simple implementations of traditional neural architectures that proved their efficiency. The Stanford course is quite good https://www.coursera.org/learn/convolutional-neural-networks/home/info
You have several free deep learning frameworks (keras https://keras.io/, thensorflow, theano, pytorch...). Keras, in my opinion, is the simplest one, especially for standard architectures (https://keras.io/getting-started/sequential-model-guide/). You can also find free pre-trained models
https://www.learnopencv.com/keras-tutorial-using-pre-trained-imagenet-models/
You can also find a lot of examples of CNN in Keras just by googling "convolutional neural networks keras"
- Badges
- Science topic
- Similar topics
- Computer Science