seq2seq-attn

seq2seq-attn copied to clipboard

# environment： ubuntu18 Torch7 gcc version 4.8.5 (Ubuntu 4.8.5-4ubuntu8) lua5.2 python2.7 only CPU # problem while i run command `th train.lua -data_file data/demo-train.hdf5 -val_data_file data/demo-val.hdf5 -savefile demo-model -gpuid -1` output:...

reBiocoder

code for seq-level distillation

1

Hi Yoon, As mentioned in the [Sequence-Level Knowledge Distillation](https://arxiv.org/pdf/1606.07947.pdf), implementation of the distillation model is released in this repo, but I didn't find the corresponding code (both word-level distillation and...

Oneplus

Hi, When training the model with guided alignment, I get the following error. ./s2sa/data.lua:174: bad argument #2 to 'sub' (out of range at /home/anupama/torch/pkg/torch/generic/Tensor.c:304) I have prepared the alignment file...

Anupama94

i read you code carefully, but i do not vary understand like 'decoder_clones = clone_many_times(decoder, opt.max_sent_l_targ) ' , why we need to copy opt.max_sent_l_targ times, and we share the parameters...

wanghechong

Hi, If my understanding is correct, this flag uses the same logic/architecture that `-use_chars_enc` does. The `use_chars_enc` uses a char-based CNN to generate an embedding matrix of size |V| x...

arendu-zz

I am using seq2seq-attn to go from AMRs to English. I am using this command: ``` th train.lua -data_file data/delexichar-train.hdf5 -val_data_file data/delexichar-val.hdf5 -savefile delexichar -use_chars_enc 1 -use_chars_dec 1 -gpuid 1...

anbasile

Do you see a better way to do this without nil checks each time?

twistedcubic

Thanks.

twistedcubic

for a large dataset about 10M QA pairs would it be a better performance on accuracy if we divide the dataset by the length of sentences. and feed it to...

SeekPoint

Fixes the following error when training on two GPUs: > Assertion `THCTensor_(checkGPU)(state, 3, self_, src1, src2)' Solution found here: https://github.com/torch/cutorch/issues/434#issuecomment-237281807

macournoyer