#

Attention with Linear Biases (ALiBi) Experiment

This is an annotated PyTorch experiment to train a ALiBi model.

16import torch
17from torch.utils.data import DataLoader
18
19from labml import experiment, tracker
20from labml.configs import option, calculate
21from labml_helpers.datasets.text import SequentialUnBatchedDataset
22from labml_nn.transformers.alibi import AlibiMultiHeadAttention
23from labml_nn.experiments.nlp_autoregression import transpose_batch
24from labml_nn.transformers import TransformerConfigs
25from labml_nn.transformers.gpt import Configs as GPTConfigs

#

Configurations

We extend GPT configurations and change the attention mechanism.

28class Configs(GPTConfigs):

#

ALiBi based transformer (defined below)

36    transformer: TransformerConfigs = 'GPT_ALiBi'

#

Longer validation set

38    valid_seq_len: int = 128
39    valid_loader = 'shuffled_longer_valid_loader'

#

Log losses at the initial and final tokens

41    def other_metrics(self, output: torch.Tensor, target: torch.Tensor):

#

If there are more tokens that the training sequence length (during validation),

46        if self.seq_len < output.shape[0]:

#

Log the loss at training sequence length

48            tracker.add(f'loss.{self.seq_len - 1}.', self.loss_func(output[self.seq_len - 1], target[self.seq_len - 1]))

#

Log the loss at the first token

50            tracker.add(f'loss.0.', self.loss_func(output[0], target[0]))

#

Log the loss at the final token

52        tracker.add(f'loss.{int(output.shape[0]) - 1}.', self.loss_func(output[-1], target[-1]))

#

Create an ALiBi attention module

55def _alibi_mha(c: TransformerConfigs):

#

59    return AlibiMultiHeadAttention(c.n_heads, c.d_model, dropout_prob=c.dropout)

#

Set all attention mechanisms to ALiBi

63calculate(TransformerConfigs.encoder_attn, 'alibi_mha', _alibi_mha)
64calculate(TransformerConfigs.decoder_attn, 'alibi_mha', _alibi_mha)
65calculate(TransformerConfigs.decoder_mem_attn, 'alibi_mha', _alibi_mha)

#

Shuffled validation data loader with valid_seq_len sequence length

68@option(Configs.valid_loader)
69def shuffled_longer_valid_loader(c: Configs):

#

73    return DataLoader(SequentialUnBatchedDataset(text=c.text.valid,
74                                                 dataset=c.text,
75                                                 seq_len=c.valid_seq_len),
76                      batch_size=c.batch_size,
77                      collate_fn=transpose_batch,
78                      shuffle=True)

#

ALiBi based Transformer configurations

81@option(Configs.transformer, 'GPT_ALiBi')
82def _transformer_configs(c: Configs):

#

We use our configurable transformer implementation

89    conf = TransformerConfigs()

#

Set the vocabulary sizes for embeddings and generating logits

91    conf.n_src_vocab = c.n_tokens
92    conf.n_tgt_vocab = c.n_tokens

#

GPT uses GELU activation for position wise feedforward

94    conf.ffn.activation = 'GELU'

#

ALiBi doesn't use positional embeddings

97    conf.src_embed = 'no_pos'
98    conf.tgt_embed = 'no_pos'

#

Set all attention mechanisms to ALiBi

101    conf.encoder_attn = 'alibi_mha'
102    conf.decoder_attn = 'alibi_mha'
103    conf.decoder_mem_attn = 'alibi_mha'

#

106    return conf

#

109def main():

#

Create experiment

111    experiment.create(name="gpt_alibi")

#

Create configs

113    conf = Configs()

#

Override configurations

115    experiment.configs(conf, {

#

Use character level tokenizer

117        'tokenizer': 'character',

#

Prompt separator is blank

119        'prompt_separator': '',

#

Starting prompt for sampling

121        'prompt': 'It is ',

#

Use Tiny Shakespeare dataset

123        'text': 'tiny_shakespeare',

#

'text': 'tiny_shakespeare_no_split',

#

Use a context size of $128$

127        'seq_len': 64,

#

Use a context size of $128$

129        'valid_seq_len': 80,

#

Train for $32$ epochs

131        'epochs': 128,

#

Batch size $128$

133        'batch_size': 128,

#

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

136        'inner_iterations': 10,

#

Transformer configurations

139        'transformer.d_model': 128,
140        'transformer.ffn.d_ff': 512,
141        'transformer.n_heads': 8,
142        'transformer.n_layers': 4,
143        'transformer.dropout': 0.1,
144    })

#

Set models for saving and loading

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

#

Start the experiment

150    with experiment.start():

#

Run training

152        conf.run()

#

156if __name__ == '__main__':
157    main()