This is a tutorial/implementation of OpenAI GPT architecture in PyTorch. We got a bunch of implementation details from minGPT by @karpathy. This implementation also uses character tiny shakespeare dataset.
GPT model is essentially a standard transformer with a few tweaks. GPT-2 and especially GPT-3 models are quite large and won't fit on a single GPU and will need model parallelism. This implementation doesn't even use data parallelism and is intended to be more of a tutorial.
Main differences of this compared to a simple autoregressive transformer are the parameter initialization, weight decay, and learning rate schedule. For the transformer we reuse the existing labml/nn transformer implementation.
Here's a notebook for training a GPT model on Tiny Shakespeare dataset.
34import torch
35from torch import nn
36
37from labml import experiment
38from labml.configs import option
39from labml_helpers.module import Module
40from labml_nn.experiments.nlp_autoregression import NLPAutoRegressionConfigs
41from labml_nn.optimizers.configs import OptimizerConfigs
42from labml_nn.transformers import TransformerConfigs, Encoder
43from labml_nn.transformers.utils import subsequent_mask
This consists of a token embedding layer, transformer encoder, and a final linear layer that gives token logits.
46class GPT(Module):
encoder
is the transformer Encoder src_embed
is the token embedding module (with positional encodings) generator
is the final fully connected layer that gives the logits.54 def __init__(self, encoder: Encoder, src_embed: Module, generator: Module):
61 super().__init__()
62 self.src_embed = src_embed
63 self.encoder = encoder
64 self.generator = generator
The mask will be initialized on the first call
67 self.mask = None
69 def forward(self, x: torch.Tensor):
Create subsequent mask if mask is not initialized or if the size of the mask is different
72 if self.mask is None or self.mask.size(0) != len(x):
Subsequent mask, will mask out tokens from seeing future tokens
74 self.mask = subsequent_mask(len(x)).to(x.device)
Get the token embeddings with positional encodings
76 x = self.src_embed(x)
Transformer encoder
78 x = self.encoder(x, self.mask)
Get logits
80 x = self.generator(x)
Return results (second value is for state, since our trainer is used with RNNs also)
84 return x, None
87class Configs(NLPAutoRegressionConfigs):
GPT model
96 model: GPT
Transformer
98 transformer: TransformerConfigs
Weight decay
100 weight_decay: float = 0.1
Number of tokens for wamup
102 warmup_steps: int = 128 * 128 * 20
Custom optimizer
105 optimizer = 'transformer_optimizer'
108@option(Configs.transformer, 'GPT')
109def _transformer_configs(c: Configs):
We use our configurable transformer implementation
116 conf = TransformerConfigs()
Set the vocabulary sizes for embeddings and generating logits
118 conf.n_src_vocab = c.n_tokens
119 conf.n_tgt_vocab = c.n_tokens
GPT uses GELU activation for position wise feedforward
121 conf.ffn.activation = 'GELU'
124 return conf
Weights of linear layers and embedding layers are initialized to instead of the default Xavier initialzation.
127def _init_weights(module):
136 if not isinstance(module, (nn.Linear, nn.Embedding)):
137 return
138
139 module.weight.data.normal_(mean=0.0, std=0.02)
Initialize biases to
142 if isinstance(module, nn.Linear) and module.bias is not None:
143 module.bias.data.zero_()
Create GPT model and initialize weights
146@option(Configs.model)
147def _model(c: Configs):
151 m = GPT(c.transformer.encoder,
152 c.transformer.src_embed,
153 c.transformer.generator).to(c.device)
Apply custom weight initialization
156 m.apply(_init_weights)
157
158 return m
This code is taken from minGPT. This applies weight decay only to weights of linear layers.
161@option(NLPAutoRegressionConfigs.optimizer)
162def transformer_optimizer(c: NLPAutoRegressionConfigs):
Collect names of parameters to apply weight decay
170 decay = set()
171 for mn, m in c.model.named_modules():
172 for pn, p in m.named_parameters():
173 fpn = f'{mn}.{pn}' if mn else pn # full param name
174
175 if fpn.endswith('weight') and isinstance(m, nn.Linear):
176 decay.add(fpn)
Get all the parameters
179 param_dict = {pn: p for pn, p in c.model.named_parameters()}
Parameters that are not decayed
181 no_decay = set(param_dict.keys()) - decay
create the pytorch optimizer object
184 opt_groups = [
185 {"params": [param_dict[pn] for pn in sorted(list(decay))], "weight_decay": c.weight_decay},
186 {"params": [param_dict[pn] for pn in sorted(list(no_decay))], "weight_decay": 0.0},
187 ]
Create a configurable optimizer, so that we can change these simply by passing a config dictionary.
192 optimizer = OptimizerConfigs()
Set parameter groups for optimization.
195 optimizer.parameters = opt_groups
Use cosine decay optimizer. This is what GPT uses.
198 optimizer.optimizer = 'AdamWarmupCosineDecay'
Set model embedding size, required if we use Noam optimizer which has an exponential decay.
201 optimizer.d_model = c.d_model
Set default weight decay. This is not required since we set the weight decay in the parameter groups.
204 optimizer.weight_decay = c.weight_decay
GPT uses a maximum learning rate of .
206 optimizer.learning_rate = 6e-4
208 optimizer.betas = (0.9, 0.95)
210 optimizer.eps = 1e-8
Weight decay is decoupled from gradients
212 optimizer.weight_decouple = True
Total number of optimization steps for learning rate cosine decay
214 optimizer.total_steps = c.epochs * len(c.text.train) // (c.batch_size * c.seq_len)
Number of warmup optimization steps
216 optimizer.warmup = c.warmup_steps // (c.batch_size * c.seq_len)
217
218 return optimizer
221def main():
Create experiment
223 experiment.create(name="gpt")
Create configs
225 conf = Configs()
Override configurations
227 experiment.configs(conf, {
Use character level tokenizer
229 'tokenizer': 'character',
Prompt separator is blank
231 'prompt_separator': '',
Starting prompt for sampling
233 'prompt': 'It is ',
Use Tiny Shakespeare dataset
235 'text': 'tiny_shakespeare',
Use a context size of
238 'seq_len': 128,
Train for epochs
240 'epochs': 32,
Batch size
242 'batch_size': 128,
Switch between training and validation for times per epoch
245 'inner_iterations': 10,
Transformer configurations
248 'transformer.d_model': 512,
249 'transformer.ffn.d_ff': 2048,
250 'transformer.n_heads': 8,
251 'transformer.n_layers': 6
252 })
Set models for saving and loading
255 experiment.add_pytorch_models({'model': conf.model})
Start the experiment
258 with experiment.start():
Run training
260 conf.run()
264if __name__ == '__main__':
265 main()