ට්රාන්ස්ෆෝමර් ස්වයංක්රීය ප්රතිගාමී අත්හදා බැලීම

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

26class AutoregressiveTransformer(nn.Module):

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

43        self.mask = None

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

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

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

52        x = self.src_embed(x)

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

56        x = self.generator(x)

60        return x, None

63class Configs(NLPAutoRegressionConfigs):

72    model: AutoregressiveTransformer

74    transformer: TransformerConfigs

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

85    conf = TransformerConfigs()

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

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():

110    experiment.create(name="transformer")

112    conf = Configs()

114    experiment.configs(conf, {

116        'tokenizer': 'character',

118        'prompt_separator': '',

120        'prompt': 'It is ',

122        'text': 'tiny_shakespeare',

125        'seq_len': 512,

127        'epochs': 32,

129        'batch_size': 16,

132        'inner_iterations': 10,

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

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

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

148    with experiment.start():

150        conf.run()

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