Auto-regressive NLP model trainer

11from typing import Callable
12
13import torch
14import torch.nn as nn
15from torch.utils.data import DataLoader
16
17from labml import lab, monit, logger, tracker
18from labml.configs import option
19from labml.logger import Text
20from labml_helpers.datasets.text import TextDataset, SequentialDataLoader, SequentialUnBatchedDataset, TextFileDataset
21from labml_helpers.device import DeviceConfigs
22from labml_helpers.metrics.accuracy import Accuracy
23from labml_helpers.module import Module
24from labml_helpers.train_valid import TrainValidConfigs, hook_model_outputs, BatchIndex
25from labml_nn.optimizers.configs import OptimizerConfigs

Cross entropy loss

28class CrossEntropyLoss(Module):
33    def __init__(self):
34        super().__init__()
35        self.loss = nn.CrossEntropyLoss()
37    def __call__(self, outputs, targets):
38        return self.loss(outputs.view(-1, outputs.shape[-1]), targets.view(-1))

Trainer configurations

This has the basic configurations for NLP auto-regressive task training. All the properties are configurable.

41class NLPAutoRegressionConfigs(TrainValidConfigs):

Optimizer

52    optimizer: torch.optim.Adam

Training device

54    device: torch.device = DeviceConfigs()

Autoregressive model

57    model: Module

Text dataset

59    text: TextDataset

Batch size

61    batch_size: int = 16

Length of the sequence, or context size

63    seq_len: int = 512

Number of token in vocabulary

65    n_tokens: int

Tokenizer

67    tokenizer: Callable = 'character'

Text prompt to start sampling (for illustration)

70    prompt: str

The token separator when sampling (blank for character level tokenization)

72    prompt_separator: str

Whether to periodically save models

75    is_save_models = True

Loss function

78    loss_func = CrossEntropyLoss()

Accuracy function

80    accuracy = Accuracy()

Model embedding size

82    d_model: int = 512

Gradient clipping

84    grad_norm_clip: float = 1.0

Training data loader

87    train_loader: DataLoader = 'shuffled_train_loader'

Validation data loader

89    valid_loader: DataLoader = 'shuffled_valid_loader'

Initialization

91    def init(self):

Set tracker configurations

96        tracker.set_scalar("accuracy.*", True)
97        tracker.set_scalar("loss.*", True)

Add a hook to log module outputs

99        hook_model_outputs(self.mode, self.model, 'model')

Add accuracy as a state module. The name is probably confusing, since it’s meant to store states between training and validation for RNNs. This will keep the accuracy metric stats separate for training and validation.

104        self.state_modules = [self.accuracy]

Training or validation step

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

Move data to the device

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

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

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

Whether to capture model outputs

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

Get model outputs. It’s returning a tuple for states when using RNNs. This is not implemented yet. 😜

123            output, *_ = self.model(data)

Calculate and log loss

126        loss = self.loss_func(output, target)
127        tracker.add("loss.", loss)

Calculate and log accuracy

130        self.accuracy(output, target)
131        self.accuracy.track()

Train the model

134        if self.mode.is_train:

Calculate gradients

136            loss.backward()

Clip gradients

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

Take optimizer step

140            self.optimizer.step()

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

142            if batch_idx.is_last:
143                tracker.add('model', self.model)

Clear the gradients

145            self.optimizer.zero_grad()

Save the tracked metrics

148        tracker.save()

Sampling function to generate samples periodically while training

150    def sample(self):

Starting prompt

156        prompt = self.prompt

Collect output for printing

158        log = [(prompt, Text.subtle)]

Sample 25 tokens

160        for i in monit.iterate('Sample', 25):

Tokenize the prompt

162            data = self.text.text_to_i(prompt).unsqueeze(-1)
163            data = data.to(self.device)

Get the model output

165            output, *_ = self.model(data)

Get the model prediction (greedy)

167            output = output.argmax(dim=-1).squeeze()

Add the prediction to prompt

169            prompt += self.prompt_separator + self.text.itos[output[-1]]

Add the prediction for logging

171            log += [(self.prompt_separator + self.text.itos[output[-1]], Text.value)]

Print the sampled output

174        logger.log(log)
177@option(NLPAutoRegressionConfigs.optimizer)
178def _optimizer(c: NLPAutoRegressionConfigs):
183    optimizer = OptimizerConfigs()
184    optimizer.parameters = c.model.parameters()
185    optimizer.optimizer = 'Adam'
186    optimizer.d_model = c.d_model
187
188    return optimizer

Get number of tokens

191@option(NLPAutoRegressionConfigs.n_tokens)
192def _n_tokens(c: NLPAutoRegressionConfigs):
196    return c.text.n_tokens

Basic english tokenizer

We use character level tokenizer in this experiment. You can switch by setting,

    'tokenizer': 'basic_english',

as the configurations dictionary when starting the experiment.

199@option(NLPAutoRegressionConfigs.tokenizer)
200def basic_english():
214    from torchtext.data import get_tokenizer
215    return get_tokenizer('basic_english')

Character level tokenizer

218def character_tokenizer(x: str):
222    return list(x)

Character level tokenizer configuration

225@option(NLPAutoRegressionConfigs.tokenizer)
226def character():
230    return character_tokenizer

Tiny Shakespeare dataset

It will download from the url if not present

233@option(NLPAutoRegressionConfigs.text)
234def tiny_shakespeare(c: NLPAutoRegressionConfigs):
240    return TextFileDataset(
241        lab.get_data_path() / 'tiny_shakespeare.txt',
242        c.tokenizer,
243        url='https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt')

Sequential training data loader

246@option(NLPAutoRegressionConfigs.train_loader)
247def sequential_train_loader(c: NLPAutoRegressionConfigs):
251    return SequentialDataLoader(text=c.text.train,
252                                dataset=c.text,
253                                batch_size=c.batch_size,
254                                seq_len=c.seq_len)

Sequential validation data loader

257@option(NLPAutoRegressionConfigs.valid_loader)
258def sequential_valid_loader(c: NLPAutoRegressionConfigs):
262    return SequentialDataLoader(text=c.text.valid,
263                                dataset=c.text,
264                                batch_size=c.batch_size,
265                                seq_len=c.seq_len)

Transpose batch

DataLoader collects the batches on the first dimension. We need to transpose it to be sequence first.

268def transpose_batch(batch):
276    transposed_data = list(zip(*batch))

Stack the batch along the second dimension dim=1

278    src = torch.stack(transposed_data[0], dim=1)
279    tgt = torch.stack(transposed_data[1], dim=1)
280
281    return src, tgt

Shuffled training data loader

284@option(NLPAutoRegressionConfigs.train_loader)
285def shuffled_train_loader(c: NLPAutoRegressionConfigs):
289    return DataLoader(SequentialUnBatchedDataset(text=c.text.train,
290                                                 dataset=c.text,
291                                                 seq_len=c.seq_len),
292                      batch_size=c.batch_size,
293                      collate_fn=transpose_batch,
294                      shuffle=True)

Shuffled validation data loader

297@option(NLPAutoRegressionConfigs.valid_loader)
298def shuffled_valid_loader(c: NLPAutoRegressionConfigs):
302    return DataLoader(SequentialUnBatchedDataset(text=c.text.valid,
303                                                 dataset=c.text,
304                                                 seq_len=c.seq_len),
305                      batch_size=c.batch_size,
306                      collate_fn=transpose_batch,
307                      shuffle=True)