Spaces:
Sleeping
Sleeping
File size: 8,228 Bytes
fe7e262 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 | from enum import Enum
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from transformers import GPTNeoXConfig, GPTNeoXModel
from . import assets
from .utils import load_checkpoint, load_config, top_p
def _get_device(module):
return next(module.parameters()).device
class ConditionEncoder(nn.Module):
def __init__(self, hp):
super().__init__()
self.l1_encoder = nn.TransformerEncoder(
nn.TransformerEncoderLayer(
d_model=hp.d_model,
nhead=hp.num_heads,
dim_feedforward=hp.d_model * 4,
dropout=hp.dropout,
activation=hp.activation,
batch_first=True,
),
hp.num_layers_encoder,
)
self.pos_emb = nn.Embedding(hp.condition_class, hp.d_model)
self.bottlenect = nn.Sequential(
nn.Linear(hp.d_model, hp.d_bottleneck),
nn.ReLU(),
nn.Linear(hp.d_bottleneck, hp.d_model),
)
def forward(self, input_embs):
B, L, N, D = input_embs.shape
pos = torch.arange(N).to(input_embs.device)
pos = self.pos_emb(pos)[None, None, :, :].expand(B, L, N, D)
input_embs = input_embs + pos
out = self.l1_encoder(input_embs.view(B * L, N, D)).view(B, L, N, D)
out = out[:, :, 0, :]
assert out.shape == (B, L, D)
out = self.bottlenect(out)
return out
class PiCoGenDecoder(nn.Module):
class InputClass(Enum):
TARGET = 0
CONDITION = 1
def __init__(self, hp):
super().__init__()
self.hp = hp
config = GPTNeoXConfig(
vocab_size=hp.vocab_size,
hidden_size=hp.d_model,
num_hidden_layers=hp.num_layers,
num_attention_heads=hp.num_heads,
intermediate_size=hp.d_model * 4,
hidden_act=hp.activation,
hidden_dropout=hp.dropout,
max_position_embeddings=hp.max_position_embeddings,
)
self.model = GPTNeoXModel(config)
self.word_emb = nn.Embedding(hp.vocab_size, hp.d_model, padding_idx=0)
self.cond_encoder = ConditionEncoder(hp)
self.cls_emb = nn.Embedding(
hp.token_class, hp.d_model, padding_idx=0
) # 0: target, 1: condition
self.lm_head = nn.Linear(hp.d_model, hp.vocab_size)
@staticmethod
def from_pretrained(
ckpt_file=None,
config_file=None,
device="cpu",
):
ckpt_file = ckpt_file if ckpt_file is not None else assets.checkpoint_file()
config_file = config_file if config_file is not None else assets.config_file()
hp = load_config(config_file)
model = PiCoGenDecoder(hp)
state_dict = load_checkpoint(ckpt_file, device)
model.load_state_dict(state_dict["model"])
model.to(device)
model.eval()
return model
def generate(
self, input_seg, input_cls_ids, need_encode, kv_cache=None, temperature=1.0, thres=0.9
):
B, L = input_cls_ids.shape
if kv_cache is None:
input_ids = torch.zeros(B, L, device=_get_device(self.word_emb)).long()
input_cond_embs = torch.zeros(
B,
L,
self.hp.condition_class,
self.hp.d_model,
device=_get_device(self.cond_encoder),
).float()
for b in range(B):
for ll in range(L):
if need_encode[b, ll]:
emb = torch.FloatTensor(np.array(input_seg[b][ll])).to(
_get_device(self.cond_encoder)
)
input_cond_embs[b, ll] = emb
else:
input_ids[b, ll] = input_seg[b][ll]
else: # NOTE: only use the last token as input
input_ids = torch.zeros(B, 1, device=_get_device(self.word_emb)).long()
input_cond_embs = torch.zeros(
B,
1,
self.hp.condition_class,
self.hp.d_model,
device=_get_device(self.cond_encoder),
).float()
for b in range(B):
if need_encode[b, -1]:
emb = torch.FloatTensor(np.array(input_seg[b][-1])).to(
_get_device(self.cond_encoder)
)
input_cond_embs[b, -1] = emb
else:
input_ids[b, -1] = input_seg[b][-1]
input_cls_ids = input_cls_ids[:, -1:]
assert input_ids.shape == input_cls_ids.shape
input_embs = self.word_emb(input_ids)
input_cond_embs = self.cond_encoder(input_cond_embs)
input_cls_embs = self.cls_emb(input_cls_ids)
if kv_cache is None:
mask = (input_embs.sum(dim=-1, keepdim=True) != 0).expand(B, L, self.hp.d_model)
else:
mask = (input_embs.sum(dim=-1, keepdim=True) != 0).expand(B, 1, self.hp.d_model)
input_cond_embs[mask] = 0 # NOTE: where input_embs is not zero
input_embs = input_embs + input_cond_embs + input_cls_embs
model_out = self.model(
inputs_embeds=input_embs,
past_key_values=kv_cache,
)
logits = self.lm_head(model_out.last_hidden_state)[:, -1, :]
assert logits.shape == (B, self.hp.vocab_size)
probs = F.softmax(top_p(logits, thres=thres, temperature=temperature), dim=-1)
output_ids = torch.multinomial(probs, num_samples=1)
assert output_ids.shape == (B, 1)
return output_ids, model_out.past_key_values
def forward(
self,
input_seqs,
input_cls_ids,
need_encode,
input_ids=None,
input_cond_embs=None,
labels=None,
kv_cache=None,
):
B, L = input_cls_ids.shape
input_cls_ids = input_cls_ids.to(_get_device(self.cls_emb))
if input_seqs is not None:
assert input_ids is None and input_cond_embs is None
input_ids = torch.zeros(B, L, device=_get_device(self.word_emb)).long()
input_cond_embs = torch.zeros(
B,
L,
self.hp.condition_class,
self.hp.d_model,
device=_get_device(self.cond_encoder),
).float()
for b in range(B):
for ll in range(L):
if need_encode[b, ll]:
emb = torch.FloatTensor(np.array(input_seqs[b][ll])).to(
_get_device(self.cond_encoder)
)
input_cond_embs[b, ll] = emb
else:
input_ids[b, ll] = input_seqs[b][ll]
else:
assert input_ids is not None and input_cond_embs is not None
input_ids = input_ids.to(_get_device(self.word_emb))
input_cond_embs = input_cond_embs.to(_get_device(self.cond_encoder))
input_embs = self.word_emb(input_ids)
input_cond_embs = self.cond_encoder(input_cond_embs)
input_cls_embs = self.cls_emb(input_cls_ids)
mask = (input_embs.sum(dim=-1, keepdim=True) != 0).expand(B, L, self.hp.d_model)
input_cond_embs[mask] = 0 # NOTE: where input_embs is not zero
input_embs = input_embs + input_cond_embs + input_cls_embs
model_out = self.model(
inputs_embeds=input_embs,
past_key_values=kv_cache,
)
logits = self.lm_head(model_out.last_hidden_state)
assert logits.shape == (B, L, self.hp.vocab_size)
lm_loss = None
if labels is not None:
assert labels.shape == (B, L)
labels = labels.to(logits.device)
loss_fct = F.cross_entropy
lm_loss = loss_fct(logits.view(-1, self.hp.vocab_size), labels.view(-1))
out = {
"loss": lm_loss,
"logits": logits,
"past_key_values": model_out.past_key_values,
"hidden_states": model_out.hidden_states,
"attentions": model_out.attentions,
}
return out
|