Spaces:
Paused
Paused
File size: 16,431 Bytes
d6a9e66 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 |
"""
Simple training loop; Boilerplate that could apply to any arbitrary neural network,
so nothing in this file really has anything to do with GPT specifically.
"""
from typing import Optional, Tuple, List
import time
import os
from collections import defaultdict
from accelerate import Accelerator
import torch
from torch.nn import functional as F
from torch.utils.data.dataloader import DataLoader
from mingpt.utils import CfgNode as CN
from cube3d.training.utils import save_model_weights, mask_cross_entropy, normalize_bboxs, top_k_prob_mask
from cube3d.training.process_single_ldr import logits2ldr, logits2ldrot, logits2ldrp, logits2flatldrp, logits2flatldrpr
from cube3d.inference.utils import load_model_weights
from tqdm import tqdm
def generate_tokens(
engine,
prompt,
inputs_ids,
latent,
resolution_base=8.0,
disable_postprocess=False,
top_p=None,
bounding_box_xyz=None,
strategy=None
):
output_ids = engine.t2t(
#[prompt],
prompt,
#use_kv_cache=True,
inputs_ids=inputs_ids,
latent=latent,
use_kv_cache=False,
resolution_base=resolution_base,
top_p=top_p,
bounding_box_xyz=bounding_box_xyz,
strategy=strategy
)
return output_ids
class Trainer:
@staticmethod
def get_default_config():
C = CN()
# device to train on
C.device = 'auto'
# dataloder parameters
C.num_workers = 4
# optimizer parameters
C.max_iters = None
C.batch_size = 4
C.learning_rate = 3e-4
C.betas = (0.9, 0.95)
C.weight_decay = 0.1 # only applied on matmul weights
C.grad_norm_clip = 1.0
C.save_interval = None
return C
def __init__(
self,
config,
engine,
train_dataset,
accelerator,
tb,
prompt: str,
indices: Optional[List[int]] = None,
resolution_base: float = 8.0,
disable_postprocessing: bool = False,
top_p: float = None,
bounding_box_xyz: Optional[Tuple[float]] = None,
save_gpt_ckpt_path: str = None,
mode: str = 'train'
):
self.config = config
self.engine = engine
self.model = engine.gpt_model
self.optimizer = None
self.callbacks = defaultdict(list)
self.train_dataset = train_dataset
self.accelerator = accelerator
# Training parameters
self.prompt = prompt
self.targets = indices
self.resolution_base = resolution_base
self.disable_postprocessing = disable_postprocessing
self.top_p = top_p
self.bounding_box_xyz = bounding_box_xyz
self.save_gpt_ckpt_path = save_gpt_ckpt_path
# determine the device we'll train on
if config.device == 'auto':
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
else:
self.device = config.device
self.model = self.model.to(self.device)
print("running on device", self.device)
# variables that will be assigned to trainer class later for logging and etc
self.iter_num = 0
self.iter_time = 0.0
self.iter_dt = 0.0
self.tb_writer = tb
self.mode = mode
def add_callback(self, onevent: str, callback):
self.callbacks[onevent].append(callback)
def set_callback(self, onevent: str, callback):
self.callbacks[onevent] = [callback]
def trigger_callbacks(self, onevent: str):
for callback in self.callbacks.get(onevent, []):
callback(self)
def run(self):
model, config = self.model, self.config
# setup the optimizer
#self.optimizer = self.engine.configure_optimizers(config)
self.optimizer, self.scheduler = self.engine.configure_optimizers_scratch_linear(config) #self.engine.configure_optimizers_lora_linear(config)
# setup the dataloader
train_loader = DataLoader(
self.train_dataset,
shuffle=False if self.mode!='train' else True,
batch_size=config.batch_size,
)
model.train()
model, self.optimizer, train_loader = self.accelerator.prepare(model, self.optimizer, train_loader)
self.iter_num = 0
self.iter_time = time.time()
data_iter = iter(train_loader)
ema_loss_for_log = 0.0
ema_ploss_for_log = 0.0
ema_rloss_for_log = 0.0
ema_dloss_for_log = 0.0
ema_floss_for_log = 0.0
#loss
dat_num = 1217 #286
x_num = 251
y_num = 215
z_num = 525
rot_num = 24
shift = 0
stride = 5
attr_shift = stride-3 #with dat and rot，+1 for bert
bert_shift = 1
x = x_num
xy = x_num + y_num + rot_num
xyz = x_num + y_num + z_num + rot_num
progress_bar = tqdm(range(0, config.max_iters), desc="Training progress")
#while True:
for self.iter_num in range(0, config.max_iters+1):
# fetch the next batch (x, y) and re-init iterator if needed
try:
batch = next(data_iter)
except StopIteration:
data_iter = iter(train_loader)
batch = next(data_iter)
#batch = [t['latent'].to(self.device) for t in batch]
self.prompt, self.targets, self.box = batch['prompt'], batch['target'].to(self.device), batch['bbox']
#self.targets = batch['latent'].to(self.device)
targets = self.targets.clone()
logits, inputs_ids, strategy, mask, cut_idx = generate_tokens(
self.engine,
self.prompt,
targets,
None,
self.resolution_base,
self.disable_postprocessing,
self.top_p,
#self.bounding_box_xyz,
normalize_bboxs(self.box.float(), [x_num-1, y_num-1, z_num-1]), #batch_normalization(self.box)
None
)
# rotation_loss = F.cross_entropy(
# logits[:,:-1,:rot_num].permute(0, 2, 1),
# inputs_ids[:,shift:,:rot_num].argmax(-1),
# )
# px_loss = mask_cross_entropy(rot_num, x+rot_num, self.box[:, 0], logits, inputs_ids, shift)
# py_loss = mask_cross_entropy(x+rot_num, xy, self.box[:, 1], logits, inputs_ids, shift)
# pz_loss = mask_cross_entropy(xy, xyz, self.box[:, 2], logits, inputs_ids, shift)
px_loss = F.cross_entropy(
logits[:,1+attr_shift+bert_shift:-1:stride,rot_num+1:x+rot_num+1+1].permute(0, 2, 1),
inputs_ids[:,shift:,-5],
ignore_index=-1 #+1 for padding
)
py_loss = F.cross_entropy(
logits[:,0+attr_shift+bert_shift:-2:stride,x+rot_num+2:xy+3].permute(0, 2, 1),
inputs_ids[:,shift:,-4],
ignore_index=-1
)
pz_loss = F.cross_entropy(
logits[:,2+attr_shift+bert_shift::stride,xy+3:xyz+4].permute(0, 2, 1),
inputs_ids[:,shift:,-3],
ignore_index=-1
)
position_loss = px_loss + py_loss + pz_loss
# dat_loss = F.cross_entropy(
# logits[:,0:-4:stride,:dat_num+1].permute(0, 2, 1),
# inputs_ids[:,shift:,-6],
# ignore_index=-1
# )
rotation_loss = F.cross_entropy(
logits[:,1+bert_shift:-3:stride,:rot_num+1].permute(0, 2, 1),
inputs_ids[:,shift:,-7],
ignore_index=-1
)
# flag_loss = F.cross_entropy(
# logits[:,:-1,xyz+dat_num:xyz+dat_num+2].permute(0, 2, 1),
# inputs_ids[:,shift:,xyz+dat_num:xyz+dat_num+2].argmax(-1),
# )
# flag_loss = F.cross_entropy(
# logits[:,:-1,-2:].permute(0, 2, 1),
# inputs_ids[:,shift:,-2:].argmax(-1),
# )
lambda_posiition = 1.0
lambda_rotation = 1.0
lambda_dat = 1.0
lambda_flag = 50.0
self.loss = lambda_posiition * position_loss #+ \
#lambda_rotation * rotation_loss #+ \
#lambda_flag * flag_loss
#lambda_dat * dat_loss + \
if strategy==1 or strategy==2:
self.loss+=lambda_rotation * rotation_loss
# targets = self.targets.clone()
# # mask_topk, mask_inv = top_k_prob_mask(F.softmax(logits[:, 1:-3:stride, :rot_num+1], dim=2), cut_idx, top_percent=0.5)
# # targets[:,shift:,-7][mask_topk] = logits[:,1:-3:stride,:rot_num+1].permute(0, 2, 1).argmax(dim=1)[mask_topk]
# # targets[:,shift:,-7][mask_inv] = self.engine.gpt_model.rot_num+1
# targets[:,shift:,-7] = logits[:,1:-3:stride,:rot_num+1].permute(0, 2, 1).argmax(dim=1)
# #targets[:,shift:,-4] = logits_y[:,0+attr_shift:-2:stride,x+rot_num+2:xy+3].permute(0, 2, 1).argmax(dim=1)
# logits_x, inputs_ids, strategy, mask, cut_idx = generate_tokens(
# self.engine,
# self.prompt,
# targets,
# None,
# self.resolution_base,
# self.disable_postprocessing,
# self.top_p,
# #self.bounding_box_xyz,
# normalize_bboxs(self.box.float(), [x_num-1, y_num-1, z_num-1]), #batch_normalization(self.box)
# 0
# )
# targets = self.targets.clone()
# targets[:,shift:,-7] = logits_x[:,1+bert_shift:-3:stride,:rot_num+1].permute(0, 2, 1).argmax(dim=1)
# mask_x, mask_x_inv = top_k_prob_mask(F.softmax(logits[:,1+attr_shift+bert_shift:-1:stride,rot_num+1:x+rot_num+1+1], dim=2), cut_idx, top_percent=0.5)
# mask_y, mask_y_inv = top_k_prob_mask(F.softmax(logits[:,0+attr_shift+bert_shift:-2:stride,x+rot_num+2:xy+3], dim=2), cut_idx, top_percent=0.5)
# mask_z, mask_z_inv = top_k_prob_mask(F.softmax(logits[:,2+attr_shift+bert_shift::stride,xy+3:xyz+4], dim=2), cut_idx, top_percent=0.5)
# targets[:,shift:,-5][mask_x] = logits_x[:,1+attr_shift+bert_shift:-1:stride,rot_num+1:x+rot_num+1+1].permute(0, 2, 1).argmax(dim=1)[mask_x]
# targets[:,shift:,-5][mask_x_inv] = self.engine.gpt_model.x_num+1
# targets[:,shift:,-4][mask_y] = logits_x[:,0+attr_shift+bert_shift:-2:stride,x+rot_num+2:xy+3].permute(0, 2, 1).argmax(dim=1)[mask_y]
# targets[:,shift:,-4][mask_y_inv] = self.engine.gpt_model.y_num+1
# targets[:,shift:,-3][mask_z] = logits_x[:,2+attr_shift+bert_shift::stride,xy+3:xyz+4].permute(0, 2, 1).argmax(dim=1)[mask_z]
# targets[:,shift:,-3][mask_z_inv] = self.engine.gpt_model.z_num+1
# logits_p, inputs_ids, strategy, mask, cut_idx = generate_tokens(
# self.engine,
# self.prompt,
# targets,
# None,
# self.resolution_base,
# self.disable_postprocessing,
# self.top_p,
# #self.bounding_box_xyz,
# normalize_bboxs(self.box.float(), [x_num-1, y_num-1, z_num-1]), #batch_normalization(self.box)
# None
# )
# logits_p[:,1+bert_shift:-3:stride,:rot_num+1] = logits[:,1+bert_shift:-3:stride,:rot_num+1]
# logits2flatldrpr(logits_p[0].cpu().detach().numpy(), inputs_ids[0].cpu().detach().numpy(), stride, 0, output_file=f"test_rightd2r2p2p_{self.iter_num}_scratch_0p5_bert.ldr")
# targets = self.targets.clone()
# targets[:,shift:,-7] = logits[:,1:-3:stride,:rot_num+1].permute(0, 2, 1).argmax(dim=1)
# targets[:,shift:,-4] = logits_y[:,0+attr_shift:-2:stride,x+rot_num+2:xy+3].permute(0, 2, 1).argmax(dim=1)
# targets[:,shift:,-5] = logits_x[:,1+attr_shift:-1:stride,rot_num+1:x+rot_num+1+1].permute(0, 2, 1).argmax(dim=1)
# logits_z, inputs_ids, strategy = generate_tokens(
# self.engine,
# self.prompt,
# targets,
# None,
# self.resolution_base,
# self.disable_postprocessing,
# self.top_p,
# #self.bounding_box_xyz,
# normalize_bboxs(self.box.float(), [x_num-1, y_num-1, z_num-1]), #batch_normalization(self.box)
# 3
# )
# backprop and update the parameters
model.zero_grad(set_to_none=True)
# #if self.mode!='train':
# logits_z[:,1:-3:stride,:rot_num+1] = logits[:,1:-3:stride,:rot_num+1]
# logits_z[:,0+attr_shift:-2:stride,x+rot_num+2:xy+3] = logits_y[:,0+attr_shift:-2:stride,x+rot_num+2:xy+3]
# logits_z[:,1+attr_shift:-1:stride,rot_num+1:x+rot_num+1+1] = logits_x[:,1+attr_shift:-1:stride,rot_num+1:x+rot_num+1+1]
# if self.iter_num>4:
# break
self.accelerator.backward(self.loss)
torch.nn.utils.clip_grad_norm_(model.parameters(), config.grad_norm_clip)
self.optimizer.step()
self.scheduler.step()
with torch.no_grad():
# Progress bar
ema_loss_for_log = 0.4 * self.loss.item() + 0.6 * ema_loss_for_log
ema_ploss_for_log = 0.4 * position_loss.item() + 0.6 * ema_ploss_for_log
ema_rloss_for_log = 0.4 * rotation_loss.item() + 0.6 * ema_rloss_for_log
#ema_dloss_for_log = 0.4 * dat_loss.item() + 0.6 * ema_dloss_for_log
#ema_floss_for_log = 0.4 * flag_loss.item() + 0.6 * ema_floss_for_log
if self.iter_num % 10 == 0:
progress_bar.set_postfix({"Loss": f"{ema_loss_for_log:.{7}f}",
"Positon_Loss": f"{ema_ploss_for_log:.{7}f}",
"Rotation_Loss": f"{ema_rloss_for_log:.{7}f}",
#"Dat_Loss": f"{ema_dloss_for_log:.{7}f}",
#"Flag_Loss": f"{ema_floss_for_log:.{7}f}",
})
progress_bar.update(10)
#logits2ldr(logits[0].cpu().detach().numpy())
if (self.iter_num % config.save_interval == 0 and self.iter_num != 0):
if self.accelerator.is_main_process:
save_model_weights(
self.engine.gpt_model,
self.save_gpt_ckpt_path,
)
# self.engine.gpt_model.save_pretrained(self.save_gpt_ckpt_path)
# torch.save({
# "ldr_proj": self.engine.gpt_model.ldr_proj.state_dict(),
# "ldr_head": self.engine.gpt_model.ldr_head.state_dict(),
# "rte": self.engine.gpt_model.rte.state_dict(),
# "dte": self.engine.gpt_model.dte.state_dict(),
# "xte": self.engine.gpt_model.xte.state_dict(),
# "yte": self.engine.gpt_model.yte.state_dict(),
# "zte": self.engine.gpt_model.zte.state_dict(),
# }, f"{self.save_gpt_ckpt_path}/unfrozen_weights.pth")
if self.tb_writer: #and self.accelerator.is_main_process:
self.tb_writer.add_scalar(f'train_loss/position_loss', position_loss.item(), self.iter_num)
self.tb_writer.add_scalar(f'train_loss/rotation_loss', rotation_loss.item(), self.iter_num)
#self.tb_writer.add_scalar(f'train_loss/dat_loss', dat_loss.item(), self.iter_num)
#self.tb_writer.add_scalar(f'train_loss/flag_loss', flag_loss.item(), self.iter_num)
self.tb_writer.add_scalar(f'train_loss/total_loss', self.loss.item(), self.iter_num)
if self.iter_num == config.max_iters:
progress_bar.close() |