Runqi Yang

Did you check that these 4 are installed? - wget - unzip - python3 - pip Especially when python is installed, it should have a "python3" alias. If these are...

I just run the training on an RTX-4000: ``` 03/12/2022 04:51:19 [Data loaded.] 03/12/2022 04:51:26 Epoch 1 03/12/2022 04:59:32 > epoch [ 1] updates[ 2707] train loss[4.39872] remaining[0:00:00] 03/12/2022 05:00:20...

On CPU it's supposed to be slow. I haven't tested on CPUs before. And I do not have any experience with Colab. Did you complete your prepro.py?

If you interrupt the kernel, what's the stack trace? It may help by indicating where it was stuck.

Sorry for the late reply. Pull requests are definitely welcomed!

I haven't fully tested the code in pure-CPU training. It seems like you'll have to remove the line `'torch_cuda_state': torch.cuda.get_rng_state()` for CPU-only training.

It should be very easy to modify "interact.py" to change the CLI predicting to batch mode. Load you json and replace the evidence and questions to the ones you load...

- What's the error message (if there's any)? - What's the RAM size of your machine and how much RAM the program takes? - If you interrupt your program when...

So you encounter this just once?

The original paper does not include how to identify unanswerable questions, so it's not directly applicable. You may refer to models/papers tailored for SQuAD 2.0. If you find a way...