File size: 8,892 Bytes
897fe06 8db8077 d1ea2f2 897fe06 0c2ae95 897fe06 5a3962c 897fe06 94752a5 897fe06 f550456 3f792e3 f550456 897fe06 d1ea2f2 0c2ae95 d1ea2f2 5a3962c d1ea2f2 8b94669 d1ea2f2 8b94669 8db8077 8b94669 8db8077 8b94669 8db8077 8b94669 897fe06 8db8077 897fe06 f550456 897fe06 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 | from typing import Any, Dict, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from lightning import LightningModule
from torchmetrics import MaxMetric, MeanMetric
from torchmetrics.classification.accuracy import Accuracy
from transformers import BertModel
class MiniAgentModule(LightningModule):
def __init__(
self,
bert_model: str,
inst_proj_model: nn.Module,
tool_proj_model: nn.Module,
pred_model: nn.Module,
lr: float,
) -> None:
super().__init__()
self.save_hyperparameters(
logger=False, ignore=["inst_proj_model", "tool_proj_model", "pred_model"]
)
self.bert_model = BertModel.from_pretrained(bert_model)
self.inst_proj_model = inst_proj_model
self.tool_proj_model = tool_proj_model
self.pred_model = pred_model
self.val_1_acc = Accuracy(task="binary")
self.val_1_precision = MeanMetric()
self.val_1_recall = MeanMetric()
self.val_2_acc = Accuracy(task="binary")
self.val_2_precision = MeanMetric()
self.val_2_recall = MeanMetric()
self.val_other_acc = Accuracy(task="binary")
self.val_other_precision = MeanMetric()
self.val_other_recall = MeanMetric()
self.lr = lr
def forward(self, x: torch.Tensor) -> torch.Tensor:
pass
def on_train_start(self) -> None:
pass
def training_step(
self, batch: Tuple[torch.Tensor, torch.Tensor], batch_idx: int
) -> torch.Tensor:
B = batch["inst_ids"].shape[0]
inst_ids = batch["inst_ids"]
inst_mask = batch["inst_mask"]
tool_ids = batch["tool_ids"]
tool_mask = batch["tool_mask"]
inst_z = self.bert_model(inst_ids, inst_mask, return_dict=False)[0]
tool_z = self.bert_model(tool_ids, tool_mask, return_dict=False)[0]
inst_emb = self.inst_proj_model(inst_z)
tool_emb = self.tool_proj_model(tool_z)
inst_emb_r = inst_emb.unsqueeze(1).repeat(1, B, 1).view(B * B, -1)
tool_emb_r = tool_emb.unsqueeze(0).repeat(B, 1, 1).view(B * B, -1)
pred = self.pred_model(inst_emb_r, tool_emb_r) # [BxB, 1]
pred = pred.view(B, B) # [B, B]
target = torch.eye(B, device=pred.device).float()
pos_weight = torch.tensor([B - 1], device=pred.device)
# pos_weight = torch.tensor([1], device=pred.device)
loss = F.binary_cross_entropy_with_logits(pred, target, pos_weight=pos_weight)
self.log("train/loss", loss, on_step=True, sync_dist=True, prog_bar=True)
return loss
def on_train_epoch_end(self) -> None:
pass
def validation_step(
self, batch: Tuple[torch.Tensor, torch.Tensor], batch_idx: int
) -> None:
inst_ids = batch["inst_ids"]
inst_mask = batch["inst_mask"]
tool_ids = batch["tool_ids"]
tool_mask = batch["tool_mask"]
correct_tool_mask = batch["correct_tool_mask"]
B = inst_ids.shape[0] # batch size
C = correct_tool_mask.shape[1] # tool capacity
tool_ids = tool_ids.view(-1, tool_ids.shape[-1]) # [B*C, L]
tool_mask = tool_mask.view(-1, tool_mask.shape[-1]) # [B*C, L]
inst_z = self.bert_model(inst_ids, inst_mask, return_dict=False)[0]
tool_z = self.bert_model(tool_ids, tool_mask, return_dict=False)[0]
inst_emb = self.inst_proj_model(inst_z) # [B, D]
tool_emb = self.tool_proj_model(tool_z) # [B*C, D]
inst_emb_r = inst_emb.unsqueeze(1).repeat(1, C, 1).view(B * C, -1)
tool_emb_r = tool_emb.view(B * C, -1)
pred = self.pred_model(inst_emb_r, tool_emb_r) # [B*C, 1]
pred = pred.view(B, C)
pred = torch.sigmoid(pred)
pred_tool_mask = pred > 0.5
true_pos_mask = pred_tool_mask & correct_tool_mask
one_tool_mask = correct_tool_mask.sum(dim=1) == 1
two_tool_mask = correct_tool_mask.sum(dim=1) == 2
other_mask = ~(one_tool_mask | two_tool_mask)
# one tool
one_tool_pos_sample = (
(pred_tool_mask[one_tool_mask] == correct_tool_mask[one_tool_mask])
.all(dim=1)
.long()
)
one_tool_precision = true_pos_mask[one_tool_mask].sum(dim=1) / torch.clamp(
pred_tool_mask[one_tool_mask].sum(dim=1), min=1
)
one_tool_recall = true_pos_mask[one_tool_mask].sum(dim=1) / torch.clamp(
correct_tool_mask[one_tool_mask].sum(dim=1), min=1
)
# two tool
two_tool_pos_sample = (
(pred_tool_mask[two_tool_mask] == correct_tool_mask[two_tool_mask])
.all(dim=1)
.long()
)
two_tool_precision = true_pos_mask[two_tool_mask].sum(dim=1) / torch.clamp(
pred_tool_mask[two_tool_mask].sum(dim=1), min=1
)
two_tool_recall = true_pos_mask[two_tool_mask].sum(dim=1) / torch.clamp(
correct_tool_mask[two_tool_mask].sum(dim=1), min=1
)
# other
other_pos_sample = (
(pred_tool_mask[other_mask] == correct_tool_mask[other_mask])
.all(dim=1)
.long()
)
other_precision = true_pos_mask[other_mask].sum(dim=1) / torch.clamp(
pred_tool_mask[other_mask].sum(dim=1), min=1
)
other_recall = true_pos_mask[other_mask].sum(dim=1) / torch.clamp(
correct_tool_mask[other_mask].sum(dim=1), min=1
)
if one_tool_pos_sample.sum().item() > 0:
self.val_1_acc.update(
one_tool_pos_sample, torch.ones_like(one_tool_pos_sample)
)
self.val_1_precision.update(one_tool_precision)
self.val_1_recall.update(one_tool_recall)
self.log(
"val/1_acc",
self.val_1_acc,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/1_precision",
self.val_1_precision,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/1_recall",
self.val_1_recall,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
if two_tool_pos_sample.sum().item() > 0:
self.val_2_acc.update(
two_tool_pos_sample, torch.ones_like(two_tool_pos_sample)
)
self.val_2_precision.update(two_tool_precision)
self.val_2_recall.update(two_tool_recall)
self.log(
"val/2_acc",
self.val_2_acc,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/2_precision",
self.val_2_precision,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/2_recall",
self.val_2_recall,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
if other_pos_sample.sum().item() > 0:
self.val_other_acc.update(
other_pos_sample, torch.ones_like(other_pos_sample)
)
self.val_other_precision.update(other_precision)
self.val_other_recall.update(other_recall)
self.log(
"val/other_acc",
self.val_other_acc,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/other_precision",
self.val_other_precision,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
self.log(
"val/other_recall",
self.val_other_recall,
on_epoch=True,
sync_dist=True,
prog_bar=True,
)
def on_validation_epoch_end(self) -> None:
pass
def test_step(
self, batch: Tuple[torch.Tensor, torch.Tensor], batch_idx: int
) -> None:
pass
def on_test_epoch_end(self) -> None:
pass
def configure_optimizers(self):
opt = torch.optim.AdamW(
[
{"params": self.bert_model.parameters(), "lr": 1e-5},
{
"params": list(self.inst_proj_model.parameters())
+ list(self.tool_proj_model.parameters())
+ list(self.pred_model.parameters()),
"lr": self.lr,
},
],
weight_decay=1e-4,
)
return opt
|