Unable to use learner.fit() because of Apex dependencies

[DGMC90]

Hi, I'm trying to follow the notebook example provided in this repo with some of my own data. However, when I go to fit the model, I get the following:

---

ModuleNotFoundError Traceback (most recent call last)
~/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/fast_bert/learner.py in get_optimizer(self, lr, num_train_steps, schedule_type)
197 try:
--> 198 from apex.optimizers import FP16_Optimizer
199 from apex.optimizers import FusedAdam

ModuleNotFoundError: No module named `'apex.optimizers'

---

I have installed Apex correctly using NVIDIA's documentation, and the Apex directory appears the same as in their repo, which leads me to think it's a fast-bert issue. I am using an AWS instance (ml.p3.8xlarge), and my environment is conda_pytorch_p36.

Thanks in advance for any help,

Darren

[DGMC90]

Closed as this seems to be an issue with Apex itself, not fast-bert. Sorry!

[kaushaltrivedi]

Did you install it with cpp_ext and cuda_ext flags.

[DGMC90]

@kaushaltrivedi yes, and the cloned repo seems to not be missing anything.

[valeriobasile]

Hi, I'm trying to run fast-bert in a colab notebook. After several trials, I was able to load apex. I installed it from the github repository with:

! git clone https://github.com/NVIDIA/apex.git
% cd apex
! pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" . --user
% cd ..

However, upon running learner.fit() the apex module "amp" cannot be found:

ImportError Traceback (most recent call last)
/usr/local/lib/python3.6/dist-packages/fast_bert/learner_cls.py in fit(self, epochs, lr, validate, schedule_type)
225 try:
--> 226 from apex import amp
227 except ImportError:

ImportError: cannot import name 'amp'

There is a related issue on the apex repo: NVIDIA/apex#259 but I couldn't make it work following those instructions, probably because they are not specific to colab.

Perhaps a solution like this could alleviate the issue? https://medium.com/the-artificial-impostor/use-nvidia-apex-for-easy-mixed-precision-training-in-pytorch-46841c6eed8c

UPDATE:

I made it work. I suspect something from the previous installations of apex was lingering, so I reset the runtimes and reconnected (actually I closed and refreshed the page altogether.
In the end what worked is the following lines to install apex from github:

!pip uninstall apex
% rm -rf /content/apex
! git clone https://github.com/NVIDIA/apex.git
! pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" /content/apex