Image Regression with Custom Head

[LotusGaurdian (Lotus Guardian)]

Hi,

I am new to fastai (getting started with the lib), fascinated to see the speed boost, tweaks and productivity it provides in creating models.

Am trying to create a model with custom head from a pre-trained model eg, resnet34 for an Image Regression problem - Price prediction based on Image, but doesn’t know how to achieve it using fastai library.

Looking forward to your help and discussion.

Thank you.

[LotusGaurdian (Lotus Guardian)]

Might be to make the ask clear, am looking forward to know how to setup custom head for a pertained model like resnet in a cleaner way!

[mike.moloch (github:aeamaea)]

HI @LotusGaurdian , what have you been able to do so far? can you draw a picture or some code samples of what you’ve done so far?

Isn’t fine-tuning essentially the same thing? Is the custom head going to be drastically different from the resnet34 model?

[LotusGaurdian (Lotus Guardian)]

Hi Mike,

Thanks for stepping in to guide.

Am very new to fast.ai lib and does’t recognise how to do a standard regression custom head on a pertained body, with concise apis.

Yes, fine-tuning is essentially going to be same, I expect.

The custom head won’t be any different but a standard one.

All I am trying to do is input a image and instead of classifying it, would like to predict it’s expected price in numbers.

Please refer the below image, and sorry for a quick scribble. The challenge is how to do construct via fast.ai lib and begin fine tuning.

Any reference snippets, inputs, reference material would be of great help! Thanks.

[mike.moloch (github:aeamaea)]

Thanks for the great diagrams @LotusGaurdian , it becomes much more clear now. Thanks to your post, I looked up image regression and learned a new concept!

Also, it’ beyond my current capabilities to suggest concrete fast.ai snippets etc , but just as a general direction, would it make sense to turn this into a multi-class prediction and each prediction represents a price range? So, instead of ‘goldfish’ we take ‘goldfish’ to mean “$120-130” ? To make it finer grained, make the number of categories larger and that’d squeeze the range accordingly?

There was an answer to a similar question previously, not sure if it fits your scenario exactly, but might help in moving things forward a bit?

Image Regression using fastai

You might be able to use it out of the box. learn = ConvLearner(data, models.resnet34, metrics=error_rate) gives a linear final layer and doesn’t have a softmax layer at the end. Then when you want to actually get predictions from the learner, I wonder if you could do something on the interpreter side (Not 100% sure if that’s how it’s designed) interp = ClassificationInterpretation.from_learner(learn, sigmoid=False) This would give you that linear layer of predictions out from the initial C…

[msivanes (Manikandan Sivanesan)]

@LotusGaurdian

I am framing the response since you are getting started with fastai library.

1. Create dataloaders using high level api - ImageDataLoaders.from_df ( Vision data | fastai ) . y_block should be chosen as RegressionBlock . If you want to tweak the data pipeline for more advanced use cases, definitely check out the Data Block tutorial in docs.fastai.com.
2. Create a vision learner with y_range chosen from (min - max value) - Learner for the vision applications | fastai

Complete tutorial from walkwithfastai - Scalar Regression | walkwithfastai

Above should help you to move forward. If you are stuck with something, please use this topic itself to post your follow up questions. We are happy to assist you.

Welcome to the community & Happy Learning!!

[LotusGaurdian (Lotus Guardian)]

@mike.moloch Great thoughts and will try the categorisation route as well.

@msivanes Thanks for the links and it will definitely help!