electra pytorch
0.1.2
Un contenedor de trabajo simple para un entrenamiento previo rápido de modelos de lenguaje como se detalla en este documento. Acelera el entrenamiento (en comparación con el modelado de lenguaje enmascarado normal) en un factor de 4 veces y, finalmente, alcanza un mejor rendimiento si se entrena durante más tiempo. Un agradecimiento especial a Erik Nijkamp por tomarse el tiempo de replicar los resultados de GLUE.
$ pip install electra-pytorch El siguiente ejemplo utiliza reformer-pytorch , que está disponible para ser instalado con pip.
import torch
from torch import nn
from reformer_pytorch import ReformerLM
from electra_pytorch import Electra
# (1) instantiate the generator and discriminator, making sure that the generator is roughly a quarter to a half of the size of the discriminator
generator = ReformerLM (
num_tokens = 20000 ,
emb_dim = 128 ,
dim = 256 , # smaller hidden dimension
heads = 4 , # less heads
ff_mult = 2 , # smaller feed forward intermediate dimension
dim_head = 64 ,
depth = 12 ,
max_seq_len = 1024
)
discriminator = ReformerLM (
num_tokens = 20000 ,
emb_dim = 128 ,
dim = 1024 ,
dim_head = 64 ,
heads = 16 ,
depth = 12 ,
ff_mult = 4 ,
max_seq_len = 1024
)
# (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 electra
trainer = Electra (
generator ,
discriminator ,
discr_dim = 1024 , # the embedding dimension of the discriminator
discr_layer = 'reformer' , # the layer name in the discriminator, whose output would be used for predicting token is still the same or replaced
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)
)
# (4) train
data = torch . randint ( 0 , 20000 , ( 1 , 1024 ))
results = trainer ( data )
results . loss . backward ()
# after much training, the discriminator should have improved
torch . save ( discriminator , f'./pretrained-model.pt' )Si prefiere que el marco no intercepte automáticamente la salida oculta del discriminador, puede pasar el discriminador (con el lineal adicional [dim x 1]) usted mismo con lo siguiente.
import torch
from torch import nn
from reformer_pytorch import ReformerLM
from electra_pytorch import Electra
# (1) instantiate the generator and discriminator, making sure that the generator is roughly a quarter to a half of the size of the discriminator
generator = ReformerLM (
num_tokens = 20000 ,
emb_dim = 128 ,
dim = 256 , # smaller hidden dimension
heads = 4 , # less heads
ff_mult = 2 , # smaller feed forward intermediate dimension
dim_head = 64 ,
depth = 12 ,
max_seq_len = 1024
)
discriminator = ReformerLM (
num_tokens = 20000 ,
emb_dim = 128 ,
dim = 1024 ,
dim_head = 64 ,
heads = 16 ,
depth = 12 ,
ff_mult = 4 ,
max_seq_len = 1024 ,
return_embeddings = True
)
# (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 electra
discriminator_with_adapter = nn . Sequential ( discriminator , nn . Linear ( 1024 , 1 ))
trainer = Electra (
generator ,
discriminator_with_adapter ,
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)
)
# (4) train
data = torch . randint ( 0 , 20000 , ( 1 , 1024 ))
results = trainer ( data )
results . loss . backward ()
# after much training, the discriminator should have improved
torch . save ( discriminator , f'./pretrained-model.pt' )El generador debe tener aproximadamente entre un cuarto y como máximo la mitad del tamaño del discriminador para un entrenamiento eficaz. Si es mayor, el generador será demasiado bueno y el juego adversario colapsará. Esto se hizo reduciendo la dimensión oculta, la dimensión oculta de avance y el número de cabezas de atención en el documento.
$ python setup.py test $ mkdir data
$ cd data
$ pip3 install gdown
$ gdown --id 1EA5V0oetDCOke7afsktL_JDQ-ETtNOvx
$ tar -xf openwebtext.tar.xz
$ wget https://storage.googleapis.com/electra-data/vocab.txt
$ cd ..$ python pretraining/openwebtext/preprocess.py$ python pretraining/openwebtext/pretrain.py$ python examples/glue/download.py $ python examples/glue/run.py --model_name_or_path output/yyyy-mm-dd-hh-mm-ss/ckpt/200000 @misc { clark2020electra ,
title = { ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators } ,
author = { Kevin Clark and Minh-Thang Luong and Quoc V. Le and Christopher D. Manning } ,
year = { 2020 } ,
eprint = { 2003.10555 } ,
archivePrefix = { arXiv } ,
primaryClass = { cs.CL }
}
