Issues viewing top losses for ULMFiT model?

I am currently having issues viewing the top losses for an NLP problem based on ULMFiT and Fastai v1’s text module.

The code below is modified from ClassificationInterpretation.plot_top_losses in the vision module’s, where I’ve substituted plotting the image data with a print the text.

n = 5
interp = ClassificationInterpretation.from_learner(learn)
tl_val, tl_idx = interp.top_losses(n)
print('prediction / actual / loss / probability', '\n')
for i,idx in enumerate(tl_idx):
    classes =
    text,cl =[idx]
    cl = int(cl)
    print(f'{classes[interp.pred_class[idx]]} / {classes[cl]} / {interp.losses[idx]:.2f} / {interp.probs[idx][cl]:.2f}  {text.text}')
    print('predictions:', *[(cat, np.round(np.float(t), 3)) for cat, t in zip(classes, interp.probs[idx])])

This produces the following output:

prediction / actual / loss / probability 

neutral / negative / 12.89 / 0.00  xxbos @southwestair xxmaj flight xxunk ( n xxunk d ) departs xxup @mco enroute to xxunk http : / / / xxunk 4 xxunk
predictions ('negative', 0.0) ('neutral', 1.0) ('positive', 0.0)

negative / negative / 7.14 / 0.99  xxbos @americanair i can not believe how long flight xxunk xxup phi is taking . i know it 's xxup us xxmaj airways but you own it . i would really like to get home .
predictions ('negative', 0.989) ('neutral', 0.01) ('positive', 0.001)

positive / neutral / 6.95 / 0.00  xxbos @americanair could you guys follow me so i can dm y all please
predictions ('negative', 0.017) ('neutral', 0.001) ('positive', 0.982)

positive / negative / 6.84 / 0.01  xxbos @usairways what is going on with the computers ? xxmaj why is my flight grounded ? xxmaj why does your airline suck so much ? xxhtstart xxmaj xxunk xxmaj questions xxhtend
predictions ('negative', 0.006) ('neutral', 0.001) ('positive', 0.993)

positive / positive / 6.38 / 1.00  xxbos xxmaj power xxmaj xxunk xxup rt @jetblue : xxmaj our fleet 's on fleek . http : / / / t 9 s 68 xxunk
predictions ('negative', 0.002) ('neutral', 0.002) ('positive', 0.996)

(This is the Twitter US Airline Sentiment Classification dataset.)

The first result looks like mislabeled data to me, however the second result is very strange. Notice that a strong negative prediction where the target was also negative resulted in the second highest loss.

I haven’t changed the default loss function, which is CrossEntropyLoss according to learn.loss_func.func.

What could be causing this? Any ideas on where to troubleshoot from here would be most appreciated!

Also, I expect that I am not the only one who would appreciate a similar ClassificationInterpretation class that is built for the text module. I would be happy to contribute to developing this if it is something others would want.


Hmm…I have not gone too far with the course. However, by briefly taking a look at the fastai documentation, I found that ClassificationInterpretation class is under the vision submodule. So I’m taking that this is only meant to be used for image classification.

You can use the losses attribute from your learner’s recorder attribute. And you possibly use that information to plot the top losses in some way. Maybe taking a look at the source code on how ClassificationInterpretation's plot_top_losses() function plots its top losses could get you started.

Hey @david_c is your problem solved? I am also planning to plot losses for ULMFIT model. Can I use your code posted or is there any updated version of it?

Unfortunately I have been working on other projects and don’t know the status of this issue, or even whether my own code caused the error. Do not use my previous code. The fastai codebase has continued to change. In particular, there is now a dedicated TextClassificationInterpretation class which I would expect to accomplish for text what the ClassificationInterpretation class does for vision.

Thanks for the reply @david_c. Yes, there is function show_top_losses in text interpreter.


Hello, anyone has got any intuition regarding why this is the case?

I have the same problem. Many of the top losses are examples that were actually classified correctly, and yet they show up in the top losses table. I have run a sentiment analysis like classification with data labeled in 3 categories. I trained the local language model first, which went very well.
I also understood that showing top losses is for vision, but it is supposed to work for text, too. If anyone can find a solution, that would be great. I assume that it’s a bug that it shows correctly predicted items in this table.