one more transformers lib
Project description
Simple way to use transformer models
Features
- Multimodal Transformer (text, text -> text; for example for Automatic Post-Editing task);
- Label smoothing loss;
- Unlikelihood loss (sentence-level and for full context), arxiv;
- LayerDrop technique for transformer encoder/decoder layers, arxiv;
- Pre/Post LayerNorm encoders (Decoders in process), arxiv;
- ADMIN Initialization (in future), arxiv;
- Top-k/Top-p sampling with temperature, arxiv;
Installation
pip install plain-transformers
Usage
Multimodal transformer example with two tokenizers:
Step one: import model and some usefull staff;
import torch
from plain_transformers.models import MultimodalTransformer
from plain_transformers.layers import MultimodalTransformerDecoder
from plain_transformers.layers import TransformerEncoder
from plain_transformers import BPEWrapper
from plain_transformers.initializations import normal_initialization, initialize_weights
from plain_transformers.samplers.nucleus_sampler import NucleusSampler
import youtokentome as yttm
Step two: train and load tokenizers;
# train your encoder tokenizer
yttm.BPE.train(..., model='encoder_tokenizer.model')
# train your decoder tokenizer
yttm.BPE.train(..., model='decoder_tokenizer.model')
# load tokenizers
encoder_tokenizer = BPEWrapper(model='encoder_tokenizer.model')
decoder_tokenizer = BPEWrapper(model='decoder_tokenizer.model')
Step three: init out model configuration;
cfg = {
'd_model': 768,
'first_encoder': {
'first_encoder_vocab_size': encoder_tokenizer.vocab_size(),
'first_encoder_max_length': 512,
'first_encoder_pad_token_id': encoder_tokenizer.pad_id,
'first_encoder_token_type_vocab_size': 2,
'first_encoder_n_heads': 8,
'first_encoder_dim_feedforward': 2048,
'first_encoder_num_layers': 3,
'first_encoder_type': 'post_ln'
},
'second_encoder': {
'second_encoder_vocab_size': encoder_tokenizer.vocab_size(),
'second_encoder_max_length': 512,
'second_encoder_pad_token_id': encoder_tokenizer.pad_id,
'second_encoder_token_type_vocab_size': 2,
'second_encoder_n_heads': 8,
'second_encoder_dim_feedforward': 2048,
'second_encoder_num_layers': 3,
'second_encoder_type': 'post_ln'
},
'decoder': {
'decoder_max_length': 512,
'decoder_vocab_size': decoder_tokenizer.vocab_size(),
'decoder_pad_token_id': decoder_tokenizer.pad_id,
'decoder_token_type_vocab_size': 2,
'decoder_n_heads': 8,
'decoder_dim_feedforward': 2048,
'decoder_num_layers': 3,
'decoder_type': 'post_ln'
},
}
Step four: initialize model and apply weight initialisation (with default parameter std=0.02
);
model = MultimodalTransformer(
TransformerEncoder,
TransformerEncoder,
MultimodalTransformerDecoder,
cfg['d_model'],
**cfg['first_encoder'],
**cfg['second_encoder'],
**cfg['decoder'],
share_decoder_head_weights=True,
share_encoder_decoder_embeddings=False,
share_encoder_embeddings=True,
)
initialize_weights(model, normal_initialization, init_range=0.02)
Step five: train our model like ordinary seq2seq;
train(model, ...)
Step six: initialize Sampler and generate model answer;
sampler = NucleusSampler(model, encoder_tokenizer=(encoder_tokenizer, encoder_tokenizer), decoder_tokenizer=decoder_tokenizer)
sampler.generate('Hello Bob, what are you doing?', second_input_text='Fine, thanks!', top_k=5)
Example
You can find working example of NMT here.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
plain-transformers-0.0.1.5.tar.gz
(19.7 kB
view hashes)
Built Distribution
Close
Hashes for plain-transformers-0.0.1.5.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | 7c957ad1fe895b9d6134bb55e38e4be1a87c9113ce90a6f048702b572b0de9da |
|
MD5 | d66589b3fc10925d4472fa3dea7561b9 |
|
BLAKE2b-256 | 9ef4a76f1486e133dc62169114a4eb0978bb1a7fc1e44a9b7995cb3d18982dc2 |
Close
Hashes for plain_transformers-0.0.1.5-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | d1030813ec64928e669d6de461b2a04f4aafd47cb6911022ca8e6d4a47db9664 |
|
MD5 | 02f8443544dd619382afe09faf5e7e00 |
|
BLAKE2b-256 | 85124aa794b5edfc434b4205b283ba28fb229c7e8f64acb549695c2a57b71db4 |