Spaces:
Runtime error
Runtime error
Commit
·
eee3c6d
1
Parent(s):
6e64f14
from minGPT
Browse files
model.py
ADDED
|
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
GPT model:
|
| 3 |
+
- the initial stem consists of a combination of token encoding and a positional encoding
|
| 4 |
+
- the meat of it is a uniform sequence of Transformer blocks
|
| 5 |
+
- each Transformer is a sequential combination of a 1-hidden-layer MLP block and a self-attention block
|
| 6 |
+
- all blocks feed into a central residual pathway similar to resnets
|
| 7 |
+
- the final decoder is a linear projection into a vanilla Softmax classifier
|
| 8 |
+
"""
|
| 9 |
+
|
| 10 |
+
import math
|
| 11 |
+
import logging
|
| 12 |
+
|
| 13 |
+
import torch
|
| 14 |
+
import torch.nn as nn
|
| 15 |
+
from torch.nn import functional as F
|
| 16 |
+
|
| 17 |
+
logger = logging.getLogger(__name__)
|
| 18 |
+
|
| 19 |
+
class GPTConfig:
|
| 20 |
+
""" base GPT config, params common to all GPT versions """
|
| 21 |
+
embd_pdrop = 0.1
|
| 22 |
+
resid_pdrop = 0.1
|
| 23 |
+
attn_pdrop = 0.1
|
| 24 |
+
|
| 25 |
+
def __init__(self, vocab_size, block_size, **kwargs):
|
| 26 |
+
self.vocab_size = vocab_size
|
| 27 |
+
self.block_size = block_size
|
| 28 |
+
for k,v in kwargs.items():
|
| 29 |
+
setattr(self, k, v)
|
| 30 |
+
|
| 31 |
+
class GPT1Config(GPTConfig):
|
| 32 |
+
""" GPT-1 like network roughly 125M params """
|
| 33 |
+
n_layer = 12
|
| 34 |
+
n_head = 12
|
| 35 |
+
n_embd = 768
|
| 36 |
+
|
| 37 |
+
class CausalSelfAttention(nn.Module):
|
| 38 |
+
"""
|
| 39 |
+
A vanilla multi-head masked self-attention layer with a projection at the end.
|
| 40 |
+
It is possible to use torch.nn.MultiheadAttention here but I am including an
|
| 41 |
+
explicit implementation here to show that there is nothing too scary here.
|
| 42 |
+
"""
|
| 43 |
+
|
| 44 |
+
def __init__(self, config):
|
| 45 |
+
super().__init__()
|
| 46 |
+
assert config.n_embd % config.n_head == 0
|
| 47 |
+
# key, query, value projections for all heads
|
| 48 |
+
self.key = nn.Linear(config.n_embd, config.n_embd)
|
| 49 |
+
self.query = nn.Linear(config.n_embd, config.n_embd)
|
| 50 |
+
self.value = nn.Linear(config.n_embd, config.n_embd)
|
| 51 |
+
# regularization
|
| 52 |
+
self.attn_drop = nn.Dropout(config.attn_pdrop)
|
| 53 |
+
self.resid_drop = nn.Dropout(config.resid_pdrop)
|
| 54 |
+
# output projection
|
| 55 |
+
self.proj = nn.Linear(config.n_embd, config.n_embd)
|
| 56 |
+
# causal mask to ensure that attention is only applied to the left in the input sequence
|
| 57 |
+
self.register_buffer("mask", torch.tril(torch.ones(config.block_size, config.block_size))
|
| 58 |
+
.view(1, 1, config.block_size, config.block_size))
|
| 59 |
+
self.n_head = config.n_head
|
| 60 |
+
|
| 61 |
+
def forward(self, x):
|
| 62 |
+
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
|
| 63 |
+
|
| 64 |
+
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
|
| 65 |
+
k = self.key(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
|
| 66 |
+
q = self.query(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
|
| 67 |
+
v = self.value(x).view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
|
| 68 |
+
|
| 69 |
+
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
|
| 70 |
+
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
|
| 71 |
+
att = att.masked_fill(self.mask[:,:,:T,:T] == 0, float('-inf'))
|
| 72 |
+
att = F.softmax(att, dim=-1)
|
| 73 |
+
att = self.attn_drop(att)
|
| 74 |
+
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
|
| 75 |
+
y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
|
| 76 |
+
|
| 77 |
+
# output projection
|
| 78 |
+
y = self.resid_drop(self.proj(y))
|
| 79 |
+
return y
|
| 80 |
+
|
| 81 |
+
class Block(nn.Module):
|
| 82 |
+
""" an unassuming Transformer block """
|
| 83 |
+
|
| 84 |
+
def __init__(self, config):
|
| 85 |
+
super().__init__()
|
| 86 |
+
self.ln1 = nn.LayerNorm(config.n_embd)
|
| 87 |
+
self.ln2 = nn.LayerNorm(config.n_embd)
|
| 88 |
+
self.attn = CausalSelfAttention(config)
|
| 89 |
+
self.mlp = nn.Sequential(
|
| 90 |
+
nn.Linear(config.n_embd, 4 * config.n_embd),
|
| 91 |
+
nn.GELU(),
|
| 92 |
+
nn.Linear(4 * config.n_embd, config.n_embd),
|
| 93 |
+
nn.Dropout(config.resid_pdrop),
|
| 94 |
+
)
|
| 95 |
+
|
| 96 |
+
def forward(self, x):
|
| 97 |
+
x = x + self.attn(self.ln1(x))
|
| 98 |
+
x = x + self.mlp(self.ln2(x))
|
| 99 |
+
return x
|
| 100 |
+
|
| 101 |
+
class GPT(nn.Module):
|
| 102 |
+
""" the full GPT language model, with a context size of block_size """
|
| 103 |
+
|
| 104 |
+
def __init__(self, config):
|
| 105 |
+
super().__init__()
|
| 106 |
+
|
| 107 |
+
# input embedding stem
|
| 108 |
+
self.tok_emb = nn.Embedding(config.vocab_size, config.n_embd)
|
| 109 |
+
self.pos_emb = nn.Parameter(torch.zeros(1, config.block_size, config.n_embd))
|
| 110 |
+
self.drop = nn.Dropout(config.embd_pdrop)
|
| 111 |
+
# transformer
|
| 112 |
+
self.blocks = nn.Sequential(*[Block(config) for _ in range(config.n_layer)])
|
| 113 |
+
# decoder head
|
| 114 |
+
self.ln_f = nn.LayerNorm(config.n_embd)
|
| 115 |
+
self.head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
|
| 116 |
+
|
| 117 |
+
self.block_size = config.block_size
|
| 118 |
+
self.apply(self._init_weights)
|
| 119 |
+
|
| 120 |
+
logger.info("number of parameters: %e", sum(p.numel() for p in self.parameters()))
|
| 121 |
+
|
| 122 |
+
def get_block_size(self):
|
| 123 |
+
return self.block_size
|
| 124 |
+
|
| 125 |
+
def _init_weights(self, module):
|
| 126 |
+
if isinstance(module, (nn.Linear, nn.Embedding)):
|
| 127 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
|
| 128 |
+
if isinstance(module, nn.Linear) and module.bias is not None:
|
| 129 |
+
torch.nn.init.zeros_(module.bias)
|
| 130 |
+
elif isinstance(module, nn.LayerNorm):
|
| 131 |
+
torch.nn.init.zeros_(module.bias)
|
| 132 |
+
torch.nn.init.ones_(module.weight)
|
| 133 |
+
elif isinstance(module, GPT):
|
| 134 |
+
torch.nn.init.normal_(module.pos_emb, mean=0.0, std=0.02)
|
| 135 |
+
|
| 136 |
+
def configure_optimizers(self, train_config):
|
| 137 |
+
"""
|
| 138 |
+
This long function is unfortunately doing something very simple and is being very defensive:
|
| 139 |
+
We are separating out all parameters of the model into two buckets: those that will experience
|
| 140 |
+
weight decay for regularization and those that won't (biases, and layernorm/embedding weights).
|
| 141 |
+
We are then returning the PyTorch optimizer object.
|
| 142 |
+
"""
|
| 143 |
+
|
| 144 |
+
# separate out all parameters to those that will and won't experience regularizing weight decay
|
| 145 |
+
decay = set()
|
| 146 |
+
no_decay = set()
|
| 147 |
+
whitelist_weight_modules = (torch.nn.Linear, )
|
| 148 |
+
blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
|
| 149 |
+
for mn, m in self.named_modules():
|
| 150 |
+
for pn, p in m.named_parameters():
|
| 151 |
+
fpn = '%s.%s' % (mn, pn) if mn else pn # full param name
|
| 152 |
+
|
| 153 |
+
if pn.endswith('bias'):
|
| 154 |
+
# all biases will not be decayed
|
| 155 |
+
no_decay.add(fpn)
|
| 156 |
+
elif pn.endswith('weight') and isinstance(m, whitelist_weight_modules):
|
| 157 |
+
# weights of whitelist modules will be weight decayed
|
| 158 |
+
decay.add(fpn)
|
| 159 |
+
elif pn.endswith('weight') and isinstance(m, blacklist_weight_modules):
|
| 160 |
+
# weights of blacklist modules will NOT be weight decayed
|
| 161 |
+
no_decay.add(fpn)
|
| 162 |
+
|
| 163 |
+
# special case the position embedding parameter in the root GPT module as not decayed
|
| 164 |
+
no_decay.add('pos_emb')
|
| 165 |
+
|
| 166 |
+
# validate that we considered every parameter
|
| 167 |
+
param_dict = {pn: p for pn, p in self.named_parameters()}
|
| 168 |
+
inter_params = decay & no_decay
|
| 169 |
+
union_params = decay | no_decay
|
| 170 |
+
assert len(inter_params) == 0, "parameters %s made it into both decay/no_decay sets!" % (str(inter_params), )
|
| 171 |
+
assert len(param_dict.keys() - union_params) == 0, "parameters %s were not separated into either decay/no_decay set!" \
|
| 172 |
+
% (str(param_dict.keys() - union_params), )
|
| 173 |
+
|
| 174 |
+
# create the pytorch optimizer object
|
| 175 |
+
optim_groups = [
|
| 176 |
+
{"params": [param_dict[pn] for pn in sorted(list(decay))], "weight_decay": train_config.weight_decay},
|
| 177 |
+
{"params": [param_dict[pn] for pn in sorted(list(no_decay))], "weight_decay": 0.0},
|
| 178 |
+
]
|
| 179 |
+
optimizer = torch.optim.AdamW(optim_groups, lr=train_config.learning_rate, betas=train_config.betas)
|
| 180 |
+
return optimizer
|
| 181 |
+
|
| 182 |
+
def forward(self, idx, targets=None):
|
| 183 |
+
b, t = idx.size()
|
| 184 |
+
assert t <= self.block_size, "Cannot forward, model block size is exhausted."
|
| 185 |
+
|
| 186 |
+
# forward the GPT model
|
| 187 |
+
token_embeddings = self.tok_emb(idx) # each index maps to a (learnable) vector
|
| 188 |
+
position_embeddings = self.pos_emb[:, :t, :] # each position maps to a (learnable) vector
|
| 189 |
+
x = self.drop(token_embeddings + position_embeddings)
|
| 190 |
+
x = self.blocks(x)
|
| 191 |
+
x = self.ln_f(x)
|
| 192 |
+
logits = self.head(x)
|
| 193 |
+
|
| 194 |
+
# if we are given some desired targets also calculate the loss
|
| 195 |
+
loss = None
|
| 196 |
+
if targets is not None:
|
| 197 |
+
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
|
| 198 |
+
|
| 199 |
+
return logits, loss
|