#

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_nn.helpers.trainer import BatchIndex
20from labml_nn.experiments.nlp_autoregression import NLPAutoRegressionConfigs

#

Auto regressive model

23class AutoregressiveModel(nn.Module):

#

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

#

Token embedding module

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

#

Transformer

33        self.transformer = transformer

#

Final layer

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

#

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

#

Initialize the subsequent mask

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

#

Token embeddings

44        x = self.src_embed(x)

#

Run it through the transformer

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

#

Generate logits of the next token

48        res = self.generator(res)

#

50        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

53class Configs(NLPAutoRegressionConfigs):

#

62    model: AutoregressiveModel
63    transformer: nn.Module

#

Token embedding size

66    d_model: int = 128

#

Number of attention heads

68    heads: int = 4

#

Dropout probability

70    dropout: float = 0.0

#

Number of features in FFN hidden layer

72    d_ff: int = 256

#

Number of transformer layers

74    n_layers: int = 6

#

Number of experts

76    n_experts: int = 4

#

Load balancing coefficient

78    load_balancing_loss_ceof = 0.01

#

Whether to scale the chosen expert outputs by the routing probability

80    is_scale_prob: bool = True

#

Whether to drop tokens

82    drop_tokens: bool = False

#

Capacity factor to determine capacity of each model

84    capacity_factor: float = 1.0

#

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

#

Initialize tracking indicators

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

#

Training or validation step

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

#

Move data to the device

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

#

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

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

#

Get model outputs.

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

#

Calculate and cross entropy loss

109        cross_entropy_loss = self.loss_func(output, target)

#

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

111        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 $i$ .

115        route_frac = counts / total

#

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

118        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.

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

#

Track stats

126        tracker.add('dropped.', total.new_tensor(n_dropped) / total)
127        tracker.add('route.min.', route_frac.min())
128        tracker.add('route.max.', route_frac.max())
129        tracker.add('route.std.', route_frac.std())
130        tracker.add('route.max_prob.', route_prob_max)
131        tracker.add("loss.", cross_entropy_loss)
132        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 $α = 0.01$ .

137        loss = cross_entropy_loss + self.load_balancing_loss_ceof * load_balancing_loss

#

Calculate and log accuracy

140        self.accuracy(output, target)
141        self.accuracy.track()

#

Train the model

144        if self.mode.is_train:

#

Calculate gradients

146            loss.backward()

#

Clip gradients

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

#

Take optimizer step

150            self.optimizer.step()

#

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

152            if batch_idx.is_last:
153                tracker.add('model', self.model)

#

Clear the gradients

155            self.optimizer.zero_grad()

#

Save the tracked metrics

158        tracker.save()

#

Initialize the auto-regressive model

161@option(Configs.model)
162def autoregressive_model(c: Configs):

#

166    m = AutoregressiveModel(c.n_tokens, c.d_model, c.transformer)
167    return m.to(c.device)

#

Initialize the switch transformer

170@option(Configs.transformer)
171def switch_transformer(c: Configs):

#

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

#

Run the experiment

192def main():

#

Create experiment

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

#

Create configs

199    conf = Configs()

#

Load configurations

201    experiment.configs(conf,

#

A dictionary of configurations to override

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

#

Set models for saving and loading

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

#

Start the experiment

229    with experiment.start():

#

TrainValidConfigs.run

231        conf.run()

#

235if __name__ == '__main__':
236    main()