#

Switch Transformer Experiment

This is an annotated PyTorch experiment to train a switch transformer.

14import torch
15import torch.nn as nn
16
17from labml import experiment, tracker
18from labml.configs import option
19from labml_helpers.module import Module
20from labml_helpers.train_valid import BatchIndex
21from labml_nn.experiments.nlp_autoregression import NLPAutoRegressionConfigs

#

Auto regressive model

24class AutoregressiveModel(Module):

#

29    def __init__(self, n_vocab: int, d_model: int, transformer: Module):
30        super().__init__()

#

Token embedding module

32        self.src_embed = nn.Embedding(n_vocab, d_model)

#

Transformer

34        self.transformer = transformer

#

Final layer

36        self.generator = nn.Linear(d_model, n_vocab)
37        self.mask = None

#

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

#

Initialize the subsequent mask

41        if self.mask is None or self.mask.size(0) != len(x):
42            from labml_nn.transformers.utils import subsequent_mask
43            self.mask = subsequent_mask(len(x)).to(x.device)

#

Token embeddings

45        x = self.src_embed(x)

#

Run it through the transformer

47        res, counts, route_prob, n_dropped, route_prob_max = self.transformer(x, self.mask)

#

Generate logits of the next token

49        res = self.generator(res)

#

51        return res, counts, route_prob, n_dropped, route_prob_max

#

Configurations

This extends NLPAutoRegressionConfigs .

The default configs can and will be over-ridden when we start the experiment

54class Configs(NLPAutoRegressionConfigs):

#

63    model: AutoregressiveModel
64    transformer: Module

#

Token embedding size

67    d_model: int = 128

#

Number of attention heads

69    heads: int = 4

#

Dropout probability

71    dropout: float = 0.0

#

Number of features in FFN hidden layer

73    d_ff: int = 256

#

Number of transformer layers

75    n_layers: int = 6

#

Number of experts

77    n_experts: int = 4

#

Load balancing coefficient

79    load_balancing_loss_ceof = 0.01

#

Whether to scale the chosen expert outputs by the routing probability

81    is_scale_prob: bool = True

#

Whether to drop tokens

83    drop_tokens: bool = False

#

Capacity factor to determine capacity of each model

85    capacity_factor: float = 1.0

#

87    def init(self):
88        super().init()

#

Initialize tracking indicators

90        tracker.set_scalar("lb_loss.*", False)
91        tracker.set_scalar("route.*", False)
92        tracker.set_scalar("dropped.*", False)

#

Training or validation step

94    def step(self, batch: any, batch_idx: BatchIndex):

#

Move data to the device

100        data, target = batch[0].to(self.device), batch[1].to(self.device)

#

Update global step (number of tokens processed) when in training mode

103        if self.mode.is_train:
104            tracker.add_global_step(data.shape[0] * data.shape[1])

#

Whether to capture model outputs

107        with self.mode.update(is_log_activations=batch_idx.is_last):

#

Get model outputs.

109            output, counts, route_prob, n_dropped, route_prob_max = self.model(data)

#

Calculate and cross entropy loss

112        cross_entropy_loss = self.loss_func(output, target)

#

Total number of tokens processed, $T$ , in the current batch $B$

114        total = counts.sum(dim=-1, keepdims=True)

#

Fraction of tokens routed to each expert $f_{i} = \frac{1}{T} x \in B \sum 1 {a r g ma x p (x), i}$ $f_{i}$ is the count of tokens where the argmax of $p (x)$ is equal to $i$ .

118        route_frac = counts / total

#

Mean routing probability $P_{i} = \frac{1}{T} x \in B \sum p_{i} (x)$

121        route_prob = route_prob / total

#

Load balancing loss $L = N i = 1 \sum N f_{i} \cdot P_{i}$ $L$ is the loss for a single layer and here we are taking the sum of losses across all layers.

126        load_balancing_loss = self.n_experts * (route_frac * route_prob).sum()

#

Track stats

129        tracker.add('dropped.', total.new_tensor(n_dropped) / total)
130        tracker.add('route.min.', route_frac.min())
131        tracker.add('route.max.', route_frac.max())
132        tracker.add('route.std.', route_frac.std())
133        tracker.add('route.max_prob.', route_prob_max)
134        tracker.add("loss.", cross_entropy_loss)
135        tracker.add("lb_loss.", load_balancing_loss)

#

Combined loss. The load balancing loss is multiplied by a coefficient $α$ which is set to something small like $α = 0.01$ .

140        loss = cross_entropy_loss + self.load_balancing_loss_ceof * load_balancing_loss

#

Calculate and log accuracy

143        self.accuracy(output, target)
144        self.accuracy.track()

#

Train the model

147        if self.mode.is_train:

#

Calculate gradients

149            loss.backward()

#

Clip gradients

151            torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=self.grad_norm_clip)

#

Take optimizer step

153            self.optimizer.step()

#

Log the model parameters and gradients on last batch of every epoch

155            if batch_idx.is_last:
156                tracker.add('model', self.model)

#

Clear the gradients

158            self.optimizer.zero_grad()

#

Save the tracked metrics

161        tracker.save()

#

Initialize the auto-regressive model

164@option(Configs.model)
165def autoregressive_model(c: Configs):

#

169    m = AutoregressiveModel(c.n_tokens, c.d_model, c.transformer)
170    return m.to(c.device)

#

Initialize the switch transformer

173@option(Configs.transformer)
174def switch_transformer(c: Configs):

#

178    from labml_nn.transformers.switch import SwitchTransformer, SwitchTransformerLayer, SwitchFeedForward
179    from labml_nn.transformers import MultiHeadAttention
180    from labml_nn.transformers.feed_forward import FeedForward
181
182    return SwitchTransformer(
183        SwitchTransformerLayer(d_model=c.d_model,
184                               attn=MultiHeadAttention(c.heads, c.d_model, c.dropout),
185                               feed_forward=SwitchFeedForward(capacity_factor=c.capacity_factor,
186                                                              drop_tokens=c.drop_tokens,
187                                                              is_scale_prob=c.is_scale_prob,
188                                                              n_experts=c.n_experts,
189                                                              expert=FeedForward(c.d_model, c.d_ff, c.dropout),
190                                                              d_model=c.d_model),
191                               dropout_prob=c.dropout),
192        c.n_layers)

#

Run the experiment

195def main():

#

Create experiment

200    experiment.create(name="switch_transformer", comment='')

#

Create configs

202    conf = Configs()

#

Load configurations

204    experiment.configs(conf,

#

A dictionary of configurations to override

206                       {'tokenizer': 'character',
207                        'text': 'tiny_shakespeare',
208                        'optimizer.learning_rate': 1.,
209                        'optimizer.optimizer': 'Noam',
210                        'prompt': 'It is',
211                        'prompt_separator': '',
212
213                        'transformer': 'switch_transformer',
214                        'n_experts': 4,
215
216                        'drop_tokens': True,
217                        'capacity_factor': 1.2,
218
219                        'train_loader': 'shuffled_train_loader',
220                        'valid_loader': 'shuffled_valid_loader',
221
222                        'seq_len': 64,
223                        'epochs': 128,
224                        'batch_size': 32,
225                        'inner_iterations': 25,
226                        })

#

Set models for saving and loading

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

#

Start the experiment

232    with experiment.start():

#

TrainValidConfigs.run

234        conf.run()

#

238if __name__ == '__main__':
239    main()