My model is always predicting the same

[Hazard]

I’m trying to learn from first lessons of Fastai. Maybe they are not meant to be used with images like heatmaps. But anyway. What I don’t understand is why my model keeps predicting the same output. Here I share with you the code. Also, is there a great guide to understand the metrics of each epoch? I’m halfway through the 3rd video, I’m still pretty noob, this is my first “project” on DL sorry.

[Danielvs (Daniel van Strien)]

@Hazard

What I don’t understand is why my model keeps predicting the same output.

How large is your training/validation set? It looks like you might be overfitting to your training set since your training loss is going down but your validation loss is rising. You may need some more training data.

Alternatively, you might want to look into using image augmentations. You will need to spend a bit of time thinking about which make sense for the types of images you are working with, for example, is there a ‘right way up’ on your images or not (to decide whether to use rotation or not).

[Hazard]

You are right! My data set is quite small. I will look into using image aumentations too. Thanks! You really helped me a lot. Have a nice day!

[stefan-ai (Stefan Josef)]

Can it be that your training data is strongly imbalanced towards the class that your model ends up predicting most of the time? In this case try to add training data especially for the underrepresented classes in your training data.

[mrfabulous1]

Hi Hazard I hope hall is well!
I agree with the comments posted so far and would add that, your images are far less distinctive than say types of cat, so you will need say 50-200 maybe more images per class this would be a good start, these should be as distinctive as possible. The more distinctive the classes generally the less data you need.

Your confusion matrix is clearly saying that all five of your classes are similar to class 4 and the 3rd class is the most similar of all. (maybe overfiitting).

Having built approximately 70+ image classifer apps, the most difficult was a wristwatch classifier, it worked well for 2 classes but I couldn’t get it to work for 70 classes, because unless you read the name on the dial many watches look very similar.

Heat maps pose some problems that images don’t, here are some links that may help generate some ideas.

How to reduce overfitting of unconventional pictures

I have a dataset of heatmaps. After training it, there was a huge difference between the train and valid loss. I am assuming this is because I have turned data augmentation off as any flip, zooming, cropping doesn’t make sense as there will not be any instance of an image like that appearing. What other methods can I apply? Here is an example from the dataset: [Forms] Here are the metrics: [loss] Also, any idea on how I can use transfer learning for this particular dataset? Thanks,

Visualizing intermediate layers a la Zeiler and Fergus

In lesson 1 Jeremy introduced the Zeiler and Fergus paper that visualizes intermediate layers to help us develop intuition for how the layers of CNNs progressively learn the building blocks needed for the classification task at hand. I was curious whether there’s a ready to use library for visualizing intermediate layers to help us beginners develop intuition for how CNNs work and when they fail to work with a set of images and classification task. There have been multiple questions in the foru…

Image segmentation: leaving some pixels unlabeled?

Short version: Is it possible to leave some pixels unlabeled when training a segmentation model? If so, how would I specify “unlabeled” when creating the mask? Long version: I am hoping to train a segmentation model to classify vegetation cover types in aerial imagery. I will need to create my own labeled masks to train the model, and I am thinking about ways to do this quickly and efficiently. One option would be to only label some of the pixels in the image, such as the ones that are easiest …

Also segmentation seems to be used a lot with heat maps.

Cheers mrfabulous1