Iāll give it a try and will probably come back to the forums soon with a ton of questions! Thanks
FYI I found this recent baseline paper for Chinese language and the competition is opened till the end of this month.
2 Likes
Thanks for Jeremy to recommend this paper for Chinese classification problem.
1 Like
btw I speak Mandarin fluently and read/write Chinese (Simplified mostly - but also Traditional) so if you wanted any help tackling Chinese let me know! This looks like an interesting challenge.
I wonder if characters could be treated more like images for the CNN to recognize? I know when I am reading Chinese its kind of like an instant recognition of the whole character as one unique āobjectā which is quite different from reading English or other phonetic-based languages.
Most Chinese characters, especially with Traditional, have a meaning component aka the āRadicalā and also a phonetic component. In fact, to look up a word in a Chinese dictionary the process is to first search for the radical - and then in that subset of that radical you locate the word based on the number of strokes. So Iām also wondering if the characters could be broken down in some meaningful way to capture these features which are embedded in the characters.
1 Like
Iāve begun to train a French Language Model on Wikipedia.fr (well part of it). Tokenizing always made me run out of memory so Iāve changed a bit Jeremyās functions to save the tokens on each chunk instead of concatenating them in one big array. All the details are in this notebook if someone has the same problem.
I discovered afterward that I had 600 million tokens so be sure to listen to Jeremyās advice regarding this!
Iāve begun to train with high learning rates (it will go up to 8!) and Leslieās 1cycle policy. First results are looking good since the first epoch gave me a validation loss of 3.98 and 27,7% accuracy (one epoch is 33K iterations though). Iāve launched a cycle of 10 epochs (might be a bit of an overkill since it represents 330K iterations) and will share the results tomorrow!
7 Likes
Thatās exciting - no-one has looked at this before AFAIK so thereās probably a paper to be written there, if youāre interested (and Iām happy to help if wanted). I think it would basically be a case of doing lots of experiments to show how 1cycle impacts learning rates, gradient clipping, and regularization in AWD LSTM and NLP classification. Ideally youād find you can decrease the regularization quite a bit, which should help particularly where you have less data.
6 Likes
That sounds super interesting but Iām not sure I would know where to begin. Letās also see the final results first, the training has plenty of time to diverge yet 
To get an idea of what represents good results, what was the validation loss (or perplexity) and accuracy of your final model trained on wiki103 we finetune after?
You may find this paper interesting. Enjoy 
You really know how to appreciate the Traditional Chinese characters. I was born in Hong Kong and studied ancient Chinese literature. While I am still working on the āmodernā Traditional Chinese, hopefully, the future step is to build the language based on ååŗ«å
Øęø
1 Like
So I couldnāt resist launching another machine on the same training but with less epochs overnight. In two epochs (still 64K iterations), I got a validation loss of 3.53 (perplexity 34.28) and an accuracy of 31.5%.
To detail the steps that led me there, the model is exactly the same as in the imdb notebook, and I took the dropouts Jeremy shared
drops = np.array([0.25, 0.1, 0.2, 0.02, 0.15])
I didnāt use Adam since it doesnāt seem to work well with high learning rates, so my line for opt_fn is:
opt_fn = partial(optim.SGD, momentum=0.9)
I admit I forgot the weight decay, so I just ran the lr_finder (the version 2 to avoid running for a full epoch since itās sooooooo long) and here is the plot
10 sounded a little bit too dangerous so I picked 8 for my maximum learning rate, then ran my favorite line
learner.fit(lr, 1, cycle_len=2, use_clr_beta=(10,10,0.95,0.85))
And here are the results for a cycle of length 2:
Now waiting for the cycle of length 10 to be completed on my other machine, but itās already at 3.48 of loss on epoch 6, so it should be better.
Iād love to test this on the wiki103 set, since there are benchmarks available. Jeremy, if you have a mean to share the numericalized version of it it would be tremendously helpful since I wouldnāt have to spend the day tokenizing it.
6 Likes
My beginning step would be to ask Sebastian Ruder to help us 
I didnāt find a great way of comparing, because I found that depending on dropout I could get the same effectiveness after fine-tuning on imdb for a wide range of different ppl and accuracy on wt103. Ppl of anything under 50 is pretty great. Hereās a table from smerityās latest:
5 Likes
Letās ask him then.
Not sure if you saw my earlier post since we replied at the same time, but training didnāt diverge and gave some pretty good results.
Again, not sure if you saw my earlier post but if you happen to have your numericalized wiki103 available it would be great for me to begin testing on this dataset.
2 Likes
For Swahili, Iāve collected a set of 1573 pdfs from Tanzaniaās Parliament website. I wrote a bash script to read all of them in sequence, convert the pdfs to TIFFs, and extract the raw text using tesseract, and so far this has got me 101 files from the hansard, totaling about 4M words, obtained using cat * | wc -w. This is already bigger than Wikipedia, but there are OCR errors that make this estimate inaccurate, and mistaken tokens might be elided when tokenizing by using min_count=3 or something like that. The process is ongoing, and Iāll be updating the dataset in multiple phases until we get to 10M.
7 Likes
I didnāt see that - quite terrific! Looks like you should be able to drop your dropout by a lot and get much better ppl, FYI.
Iāve popped the numericalized file and vocab here for you: http://files.fast.ai/data/wt103_ids.tgz
3 Likes
that would be amazing if you could create your LM based on ååŗ«å
Øęø! Have you read all of å大åčļ¼I used to read a lot of those books when I was in school 
Also thanks for sharing the paper on using CNNs for classifying sentiment in Chinese text although it looks like they are just using 1D CNNs as an alternative to RNNs. I was still thinking to use RNNs but instead of just using a tokenization from Unicode I thought perhaps it could be possible to treat each unique character as an image (represented by CNN feature extraction) and so in the same way that video is a sequence of images here we are applying the same technique with the RNN only itās a sequence of feature-rich characters.
1 Like
Thanks for the files!
I hadnāt realized that the vocab size for the Language Model wasnāt truncated to 60k words like in our fine-tuning (which makes sense if we want it to be as general as possible) so I canāt use the same batch size (128) and itāll probably impact the loss since itās harder to predict one word out of that many.
Speaking of which, training on a longer cycle the French Model (10 epochs) with very high learning rate allowed the model to find a better minimum: the final validation loss was 3.33 (27.9 perplexity!) and the accuracy was around 33%.
On all my experiments, Leslieās 1cycle policy works best when we train a new model from scratch than when you finetune something, and it can give you a decent model very fast with super convergence, but itāll also gain in precision if you let the cycle be longer. And 1 long cycle is far better than two shorter ones in a row.
2 Likes
Oh I apologize I didnāt mean to give you that one - my attempts at using a larger vocab didnāt result in better models, and were much harder to train. Iāve uploaded a replacement now to the same place with 30k vocab. Thatās what I used in practice.
BTW when we donāt use the same vocab size as a paper weāre comparing to we canāt exactly compare, since it impacts the difficulty of the problem. Also, IIRC French is an easier language for an LM. Will be very interested to hear how the en model goes with similar settings - that fr model result is really encouraging.
2 Likes
Try using sentencepiece. Should give the a nearly-optimal segmentation.
1 Like
Do you recommend using sentencepiece in a situation where one does not have a readily-available tokenizer? Iām buliding a Swahili AWD-LSTM model on Wikipedia, fine-tuned to the Hansard I posted about earlier, but spacy doesnāt have a model for it, and there are no publicly available tokenizers.
30k words is much easier indeed, thanks for the new file. I can now go to a batch size of 256 and even higher learning rates! 
2 Likes
Ndiyo!
4 Likes