මෝටර්රථ ප්රතිගාමී දුම්රිය ට්රාන්ස්ෆෝමර්

23class AutoregressiveModel(Module):

28    def __init__(self, src_embed: Module, encoder: Encoder, generator: Generator, *,
29                 is_save_ff_input: bool = False):
30        super().__init__()

32        self.src_embed = src_embed

34        self.encoder = encoder

38        self.encoder.layers[-1].is_save_ff_input = is_save_ff_input

41        self.generator = generator

43        self.src_mask = None

45    @property
46    def ff_input(self) -> torch.Tensor:

50        return self.encoder.layers[-1].ff_input

52    def forward(self, src: torch.Tensor):

54        if self.src_mask is None or self.src_mask.size(0) != len(src):
55            self.src_mask = subsequent_mask(len(src)).to(src.device)

57        res = self.encoder(self.src_embed(src), self.src_mask)

59        return self.generator(res), None

62class Configs(NLPAutoRegressionConfigs):

69    transformer: TransformerConfigs
70    model: AutoregressiveModel
71
72    is_save_ff_input = False

75@option(Configs.model)
76def autoregressive_model(c: Configs):

80    m = AutoregressiveModel(

83        src_embed=c.transformer.src_embed,
84        encoder=c.transformer.encoder,
85        generator=c.transformer.generator,

87        is_save_ff_input=c.is_save_ff_input)
88    return m.to(c.device)

91@option(Configs.transformer)
92def transformer_c(c: Configs):

96    tc = TransformerConfigs()
97    tc.n_src_vocab = c.n_tokens
98    tc.n_tgt_vocab = c.n_tokens
99
100    return tc

103def main():

105    experiment.create(name="knn_lm")

107    conf = Configs()

109    experiment.configs(conf,

111                       {'tokenizer': 'character',
112                        'prompt_separator': '',
113                        'prompt': 'It is ',
114                        'text': 'tiny_shakespeare',
115
116                        'optimizer.optimizer': 'Noam',
117                        'optimizer.learning_rate': 1.,
118                        'optimizer.d_model': 256,
119
120                        'seq_len': 1024,
121                        'epochs': 128,
122                        'batch_size': 6,
123                        'inner_iterations': 10,

126                        'transformer.d_model': 256,
127                        'transformer.ffn.d_ff': 1024,
128                        'transformer.n_heads': 8,
129                        'transformer.n_layers': 6})

132    conf.n_tokens = conf.text.n_tokens

135    experiment.add_pytorch_models(get_modules(conf))

138    with experiment.start():

140        conf.run()
141
142
143if __name__ == '__main__':
144    main()