#

Transformer Auto-Regression Experiment

This trains a simple transformer introduced in Attention Is All You Need on an NLP auto-regression task (with Tiny Shakespeare dataset).

16import torch
17from torch import nn
18
19from labml import experiment
20from labml.configs import option
21from labml_nn.experiments.nlp_autoregression import NLPAutoRegressionConfigs
22from labml_nn.transformers import TransformerConfigs, Encoder
23from labml_nn.transformers.utils import subsequent_mask

#

Auto-Regressive model

26class AutoregressiveTransformer(nn.Module):

#

encoder is the transformer Encoder
src_embed is the token embedding module (with positional encodings)
generator is the final fully connected layer that gives the logits.

30    def __init__(self, encoder: Encoder, src_embed: nn.Module, generator: nn.Module):

#

37        super().__init__()
38        self.src_embed = src_embed
39        self.encoder = encoder
40        self.generator = generator

#

The mask will be initialized on the first call

43        self.mask = None

#

45    def forward(self, x: torch.Tensor):

#

Create subsequent mask if mask is not initialized or if the size of the mask is different

48        if self.mask is None or self.mask.size(0) != len(x):

#

Subsequent mask, will mask out tokens from seeing future tokens

50            self.mask = subsequent_mask(len(x)).to(x.device)

#

Get the token embeddings with positional encodings

52        x = self.src_embed(x)

#

Transformer encoder

54        x = self.encoder(x, self.mask)

#

Get logits

56        x = self.generator(x)

#

Return results (second value is for state, since our trainer is used with RNNs also)

60        return x, None

#

Configurations

This inherits from NLPAutoRegressionConfigs

63class Configs(NLPAutoRegressionConfigs):

#

GPT model

72    model: AutoregressiveTransformer

#

Transformer

74    transformer: TransformerConfigs

#

Transformer configurations

77@option(Configs.transformer, 'Transformer')
78def _transformer_configs(c: Configs):

#

We use our configurable transformer implementation

85    conf = TransformerConfigs()

#

Set the vocabulary sizes for embeddings and generating logits

87    conf.n_src_vocab = c.n_tokens
88    conf.n_tgt_vocab = c.n_tokens

#

90    conf.d_model = c.d_model

#

93    return conf

#

Create GPT model and initialize weights

96@option(Configs.model)
97def _model(c: Configs):

#

101    m = AutoregressiveTransformer(c.transformer.encoder,
102                                  c.transformer.src_embed,
103                                  c.transformer.generator).to(c.device)
104
105    return m

#

108def main():

#

Create experiment

110    experiment.create(name="transformer")

#

Create configs

112    conf = Configs()

#

Override configurations

114    experiment.configs(conf, {

#

Use character level tokenizer

116        'tokenizer': 'character',

#

Prompt separator is blank

118        'prompt_separator': '',

#

Starting prompt for sampling

120        'prompt': 'It is ',

#

Use Tiny Shakespeare dataset

122        'text': 'tiny_shakespeare',

#

Use a context size of $256$

125        'seq_len': 512,

#

Train for 32 epochs

127        'epochs': 32,

#

Batch size $32$

129        'batch_size': 16,

#

Switch between training and validation for $10$ times per epoch

132        'inner_iterations': 10,

#

Model size

135        'd_model': 256,
136        'transformer.n_heads': 16,
137        'transformer.ffn.d_ff': 1024,

#

Use Noam optimizer

140        'optimizer.optimizer': 'Noam',
141        'optimizer.learning_rate': 1.,
142    })

#

Set models for saving and loading

145    experiment.add_pytorch_models({'model': conf.model})

#

Start the experiment

148    with experiment.start():

#

Run training

150        conf.run()

#

154if __name__ == '__main__':
155    main()