COCO LM Pretraining (wip)
Implementation of COCO-LM, Correcting and Contrasting Text Sequences for Language Model Pretraining, in Pytorch. They were able to make contrastive learning work in a self-supervised manner for language model pretraining. Seems like a solid successor to Electra.
Install
$ pip install coco-lm-pytorch
Usage
An example using the x-transformers library
$ pip install x-transformers
Then
import torch
from coco_lm_pytorch import COCO
# (1) instantiate the generator and discriminator, making sure that the generator is roughly a quarter to a half of the size of the discriminator
from x_transformers import TransformerWrapper, Encoder
generator = TransformerWrapper(
num_tokens = 20000,
emb_dim = 128,
max_seq_len = 1024,
attn_layers = Encoder(
dim = 256, # smaller hidden dimension
heads = 4, # less heads
ff_mult = 2, # smaller feedforward dimension
depth = 1
)
)
discriminator = TransformerWrapper(
num_tokens = 20000,
emb_dim = 128,
max_seq_len = 1024,
attn_layers = Encoder(
dim = 1024,
heads = 16,
ff_mult = 4,
depth = 12
)
)
# (2) weight tie the token and positional embeddings of generator and discriminator
generator.token_emb = discriminator.token_emb
generator.pos_emb = discriminator.pos_emb
# weight tie any other embeddings if available, token type embeddings, etc.
# (3) instantiate COCO
trainer = COCO(
generator,
discriminator,
discr_dim = 1024, # the embedding dimension of the discriminator
discr_layer = 'norm', # the layer name in the discriminator, whose output would be used for predicting token is still the same or replaced
cls_token_id = 1, # a token id must be reserved for [CLS], which is prepended to the sequence for contrastive learning
mask_token_id = 2, # the token id reserved for masking
pad_token_id = 0, # the token id for padding
mask_prob = 0.15, # masking probability for masked language modeling
mask_ignore_token_ids = [], # ids of tokens to ignore for mask modeling ex. (cls, sep)
cl_weight = 1., # weight for the contrastive learning loss
disc_weight = 1., # weight for the corrective learning loss
gen_weight = 1. # weight for the MLM loss
)
# (4) train
data = torch.randint(0, 20000, (1, 1024))
loss = trainer(data)
loss.backward()
# after much training, the discriminator should have improved
torch.save(discriminator, f'./pretrained-model.pt')
Citations
@misc{meng2021cocolm,
title = {COCO-LM: Correcting and Contrasting Text Sequences for Language Model Pretraining},
author = {Yu Meng and Chenyan Xiong and Payal Bajaj and Saurabh Tiwary and Paul Bennett and Jiawei Han and Xia Song},
year = {2021},
eprint = {2102.08473},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}
457 Dec 23, 2022
338 Dec 2, 2022
21 Aug 6, 2022
1k Dec 30, 2022
45 Oct 19, 2022
26 Dec 14, 2022
726 Dec 30, 2022
22 Nov 13, 2022
612 Jan 4, 2023
161 Aug 07, 2022
37 Dec 07, 2022
19 Nov 26, 2022
47 Nov 18, 2022
2 Aug 06, 2022
2 Oct 10, 2022
8 Oct 25, 2022
42 Dec 31, 2022
3.1k Dec 29, 2022
1 Mar 04, 2022
55 Nov 17, 2022
16 Mar 10, 2022
1 Apr 03, 2022
1 Apr 05, 2022
0 Dec 11, 2022
317 Dec 18, 2022
153 Jun 22, 2022
142 Jan 06, 2023
32 Jul 29, 2022
587 Dec 30, 2022