Full food101 dataset accuracy at 75% with effnetb2

[longtongster]

Was anyone able to get a higher accuracy on the full food101 dataset test set? If so can you please share your ideas? I am looking to learn tips and tricks to further improve, if possible.

I just did some straightforward training of the head of 7 epochs and then some finetuning also for 7 epochs on 30% of the training data. Looking at my loss and accuracy the test acc is still higher than train accuracy so the model has some capacity to further improvement. I see that in finetuning certain layers need to be kept frozen (batchnorm). Maybe a suggestions for @mrdbourke to include a bit more details on how exactly to do fine-tuning in the course. Going to the full dataset we have too much data just to only train the classifier so it makes sense to move towards fine-tuning.

Kind Regards
Longtongster

[mrdbourke]

Hi @longtongster ,

Good to see you're trying to improve your models!

As for fine-tuning models, I'd say your next step is to try fine-tuning all the layers with code such as:

for param in model.parameters():
            param.requires_grad = True

This means all the parameters in the target model will have their gradients tracked and updated in the training code.

A good tip while fine-tuning is to change the learning rate to 10x smaller (because the weights are already trained on an existing dataset, you don't want to overtrain them), for example go from lr=0.001 -> lr=0.0001.

I'd try doing this for ~3-5 epochs and see how it goes.

[longtongster]

Thanks for sharing the idea @mrdbourke . I was doing already some fine-tuning but found that with a learning rate scheduler and a bit more training I am able to get to increase to 81% accuracy with the efficient net b2 with in the order of 8 mio parameters. Have you gone for the Full Monty and tried to squeeze more out of this dataset? If you have more ideas please let me know.

I also looked at some of the most confident wrong predictions and feel like I would have not been able to guess the real label. Some are obviously wrong. The model has quite some cheese cakes wrong once there is also a strawberry in the image :-)

[longtongster]

Just an update for anyone interested ... with using 40% of the data, I am able to get to an accuracy of 84.9% on the full test set by applying just fine-tuning for 10 epochs and a linearLR learning rate schedular that starts at 1e-4. On my machine that takes quite some time (e.g 1 hour). Some images have a meal with specific ingredients then the model might predict the for example the ingredients with highest prob and not the meal (which is the label).

[mrdbourke]

Nice work @longtongster !

I find that too, if a meal has more than one possible label (e.g. the image contains a burger and fries) the model often predicts both classes in the top 1-3 labels.

That's one of the tradeoffs of single-class classification, models can get very good but often images contain more than one item.