pbelcak/fastfeedforward

A repository for log-time feedforward networks

fastfeedforward's Issues

GREAT WORK

Arguably the most impactful paper released this year. Great work guys.

what's the parameter 'leaf width' mean

Hi ！
May I ask what's the parameter 'leaf width' mean? Looking forward for your reply!

FFF-BERT seems to run slower than a vanilla BERT model

EDIT: Moved to pbelcak/UltraFastBERT#1

Add Arxiv link to readme

Please mention the corresponding Arxiv article link https://arxiv.org/abs/2308.14711 in the readme.

Error: Cannot use soft decisions during evaluation.

Hello,
thanks for publishing the code along with your paper. While reading your paper, I tried to run the demo notebook from this repository. I modified it slightly for colab to use the pip package instead of the local repository. However, I run into this issue:

ValueError                                Traceback (most recent call last)
<ipython-input-6-809883d6966a> in <cell line: 10>()
     32         testing_accuracies = []
     33         for batch_images, batch_labels in tqdm(dataloader_testing):
---> 34                 output = model(batch_images.view(-1, 784))
     35                 loss = criterion(output, batch_labels)
     36                 accuracy = (output.argmax(dim=1) == batch_labels).detach().float().mean()

1 frames
/usr/local/lib/python3.10/dist-packages/fastfeedforward/fff.py in forward(self, x, return_entropies, use_hard_decisions)
    332                                 raise ValueError("Cannot return entropies during evaluation.")
    333                         if not use_hard_decisions:
--> 334                                 raise ValueError("Cannot use soft decisions during evaluation.")
    335                         return self.eval_forward(x)
    336 

ValueError: Cannot use soft decisions during evaluation.

Here is the notebook to reproduce the error:

https://colab.research.google.com/drive/1apDlwbJPGMMoNodTOUx5m8Klnp8zE7T-?usp=sharing

The eval forward speed dropped significantly

The network's running speed in the eval state has dropped significantly. I don't know what's going on. There are no errors reported. During training, only 30 seconds were used for 1140 steps of iteration, while one step in the eval state took 1-2 seconds.
i took a FFF(40,40,2,4) at the end of my backbone net on cuda.
I'm working on a task, and your work is suits me very well, but it's too difficult to use,sad.

Left branches getting ignored in PyTorch implementation

Moved to pbelcak/UltraFastBERT#2