File size: 10,971 Bytes
e290a7d |
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 |
import os
import time
from tqdm import tqdm
import cv2
import numpy as np
import torch
import pdb
import torch.cuda.amp as amp
import torch.distributed as dist
import torch.nn.functional as F
import wandb
from loguru import logger
from utils.dataset_verbonly import tokenize
from utils.misc import (AverageMeter, ProgressMeter, concat_all_gather,
trainMetricGPU)
## todo : add oIoU metric
def train(train_loader, model, optimizer, scheduler, scaler, epoch, args):
# torch.autograd.set_detect_anomaly(True)
batch_time = AverageMeter('Batch', ':2.2f')
data_time = AverageMeter('Data', ':2.2f')
lr = AverageMeter('Lr', ':1.6f')
loss_meter = AverageMeter('Loss', ':2.4f')
iou_meter = AverageMeter('IoU', ':2.2f')
pr_meter = AverageMeter('Prec@50', ':2.2f')
progress = ProgressMeter(
len(train_loader),
[batch_time, data_time, lr, loss_meter, iou_meter, pr_meter],
prefix="Training: Epoch=[{}/{}] ".format(epoch, args.epochs))
model.train()
time.sleep(2)
end = time.time()
# size_list = [320, 352, 384, 416, 448, 480, 512]
# idx = np.random.choice(len(size_list))
# new_size = size_list[idx]
for i, (image, text, target, hardpos, params) in enumerate(train_loader):
data_time.update(time.time() - end)
# data
image = image.cuda(non_blocking=True)
text = text.cuda(non_blocking=True)
target = target.cuda(non_blocking=True).unsqueeze(1)
hardpos = hardpos.cuda(non_blocking=True)
hp_emb = params['hardpos_emb'].cuda(non_blocking=True)
with amp.autocast():
pred, target, loss = model(image, text, target, hardpos, hp_emb) # , fq, vis, word, state
# backward
optimizer.zero_grad()
# scaler.scale(loss).backward()
scaler.scale(loss).backward()
# loss.backward()
# for name, param in model.named_parameters():
# if param.grad is not None:
# if torch.isinf(param.grad).any() or torch.isnan(param.grad).any():
# print(f"Inf/NaN in gradients: {name}")
# for name, param in model.named_parameters():
# if param.grad is not None:
# grad_norm = param.grad.norm()
# if torch.isnan(grad_norm):
# print(f"NaN gradient detected in {name}")
if args.max_norm:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_norm)
# optimizer.step()
# scheduler.step()
scaler.step(optimizer)
scaler.update()
# metric
iou, pr5 = trainMetricGPU(pred, target, 0.35, 0.5)
dist.all_reduce(loss.detach())
dist.all_reduce(iou)
dist.all_reduce(pr5)
loss = loss / dist.get_world_size()
iou = iou / dist.get_world_size()
pr5 = pr5 / dist.get_world_size()
loss_meter.update(loss.item(), image.size(0))
iou_meter.update(iou.item(), image.size(0))
pr_meter.update(pr5.item(), image.size(0))
lr.update(scheduler.get_last_lr()[-1])
batch_time.update(time.time() - end)
end = time.time()
# if (i + 1) % args.print_freq == 0:
# progress.display(i + 1)
# if dist.get_rank() in [-1, 0]:
# wandb.log(
# {
# "time/batch": batch_time.val,
# "time/data": data_time.val,
# "training/lr": lr.val,
# "training/loss": loss_meter.val,
# "training/iou": iou_meter.val,
# "training/prec@50": pr_meter.val,
# },
# step=epoch * len(train_loader) + (i + 1))
@torch.no_grad()
def validate(val_loader, model, epoch, args):
iou_list = []
I_list = []
U_list = []
model.eval()
time.sleep(2)
for imgs, texts, masks, param in val_loader:
# data
imgs = imgs.cuda(non_blocking=True)
texts = texts.cuda(non_blocking=True)
# inference
preds = model(imgs, texts)
preds = torch.sigmoid(preds)
if preds.shape[-2:] != imgs.shape[-2:]:
preds = F.interpolate(preds,
size=imgs.shape[-2:],
mode='bicubic',
align_corners=True).squeeze(1)
# process one batch
# for pred, mask_dir, mat, ori_size in zip(preds, param['mask_dir'],
# param['inverse'],
# param['ori_size']):
# h, w = np.array(ori_size)
# mat = np.array(mat)
# pred = pred.cpu().numpy()
# pred = cv2.warpAffine(pred, mat, (w, h),
# flags=cv2.INTER_CUBIC,
# borderValue=0.)
# pred = np.array(pred > 0.35)
# mask = cv2.imread(mask_dir, flags=cv2.IMREAD_GRAYSCALE)
# mask = mask / 255.
# # iou
# inter = np.logical_and(pred, mask)
# union = np.logical_or(pred, mask)
# iou = np.sum(inter) / (np.sum(union) + 1e-6)
# iou_list.append(iou)
# I_list.append(inter)
# U_list.append(union)
for pred, mask in zip(preds, masks):
# h, w = np.array(ori_size)
# mat = np.array(mat)
pred = pred.cpu().numpy()
# pred = cv2.warpAffine(pred, mat, (w, h),
# flags=cv2.INTER_CUBIC,
# borderValue=0.)
pred = np.array(pred > 0.35)
# mask = cv2.imread(mask_dir, flags=cv2.IMREAD_GRAYSCALE)
# mask = mask / 255.
mask = mask.numpy()
# iou
inter = np.logical_and(pred, mask)
union = np.logical_or(pred, mask)
iou = np.sum(inter) / (np.sum(union) + 1e-6)
I_list.append(inter)
U_list.append(union)
iou_list.append(iou)
iou_list = np.stack(iou_list)
iou_list = torch.from_numpy(iou_list).to(imgs.device)
iou_list = concat_all_gather(iou_list)
I_list = np.stack(I_list)
I_list = torch.from_numpy(I_list).to(imgs.device)
I_list = concat_all_gather(I_list)
U_list = np.stack(U_list)
U_list = torch.from_numpy(U_list).to(imgs.device)
U_list = concat_all_gather(U_list)
overall_I = I_list.sum().item()
overall_U = U_list.sum().item()
overall_IoU = overall_I / (overall_U + 1e-6) # to avoid division by zero
prec_list = []
for thres in torch.arange(0.5, 1.0, 0.1):
tmp = (iou_list > thres).float().mean()
prec_list.append(tmp)
iou = iou_list.mean()
prec = {}
temp = ' '
for i, thres in enumerate(range(5, 10)):
key = 'Pr@{}'.format(thres * 10)
value = prec_list[i].item()
prec[key] = value
temp += "{}: {:.2f} ".format(key, 100. * value)
head = 'Evaluation: Epoch=[{}/{}] IoU={:.2f} OIoU={:.4f}'.format(
epoch, args.epochs, 100. * iou.item(), 100. * overall_IoU)
logger.info(head + temp)
# print(head)
# return three results : mIoU, oIoU and prec results
return iou.item(), overall_IoU, prec
@torch.no_grad()
def inference(test_loader, model, args):
iou_list = []
I_list = []
U_list = []
tbar = tqdm(test_loader, desc='Inference:', ncols=100)
model.eval()
time.sleep(2)
for img, mask, param in tbar:
# data
# img = img.cuda(non_blocking=True)
# mask = cv2.imread(param['mask_dir'][0], flags=cv2.IMREAD_GRAYSCALE)
img = img.cuda(non_blocking=True)
mask = mask[0].cpu().numpy()
# dump image & mask
if args.visualize:
seg_id = param['seg_id'][0].cpu().numpy()
img_name = '{}-img.jpg'.format(seg_id)
mask_name = '{}-mask.png'.format(seg_id)
cv2.imwrite(filename=os.path.join(args.vis_dir, img_name),
img=param['ori_img'][0].cpu().numpy())
cv2.imwrite(filename=os.path.join(args.vis_dir, mask_name),
img=mask)
# multiple sentences
for sent in param['sents']:
# mask = mask / 255.
text = tokenize(sent, args.word_len, True)
text = text.cuda(non_blocking=True)
# inference
pred = model(img, text)
pred = torch.sigmoid(pred)
if pred.shape[-2:] != img.shape[-2:]:
pred = F.interpolate(pred,
size=img.shape[-2:],
mode='bicubic',
align_corners=True).squeeze()
# process one sentence
# h, w = param['ori_size'].numpy()[0]
# mat = param['inverse'].numpy()[0]
pred = pred.cpu().numpy()
# pred = cv2.warpAffine(pred, mat, (w, h),
# flags=cv2.INTER_CUBIC,
# borderValue=0.)
pred = np.array(pred > 0.35)
# iou
inter = np.logical_and(pred, mask)
union = np.logical_or(pred, mask)
iou = np.sum(inter) / (np.sum(union) + 1e-6)
iou_list.append(iou)
I_list.append(inter)
U_list.append(union)
# dump prediction
if args.visualize:
pred = np.array(pred*255, dtype=np.uint8)
sent = "_".join(sent[0].split(" "))
pred_name = '{}-iou={:.2f}-{}.png'.format(seg_id, iou*100, sent)
cv2.imwrite(filename=os.path.join(args.vis_dir, pred_name),
img=pred)
logger.info('=> Metric Calculation <=')
iou_list = np.stack(iou_list)
iou_list = torch.from_numpy(iou_list).to(img.device)
I_list = np.stack(I_list)
I_list = torch.from_numpy(I_list).to(img.device)
U_list = np.stack(U_list)
U_list = torch.from_numpy(U_list).to(img.device)
overall_I = I_list.sum().item()
overall_U = U_list.sum().item()
overall_IoU = overall_I / (overall_U + 1e-6) # to avoid division by zero
prec_list = []
for thres in torch.arange(0.5, 1.0, 0.1):
tmp = (iou_list > thres).float().mean()
prec_list.append(tmp)
iou = iou_list.mean()
prec = {}
for i, thres in enumerate(range(5, 10)):
key = 'Pr@{}'.format(thres*10)
value = prec_list[i].item()
prec[key] = value
logger.info('IoU={:.2f} OIoU={:.4f}'.format(100.*iou.item(), 100. * overall_IoU))
print('IoU={:.2f} OIoU={:.4f}'.format(100.*iou.item(), 100. * overall_IoU))
for k, v in prec.items():
logger.info('{}: {:.2f}.'.format(k, 100.*v))
return iou.item(), overall_IoU, prec
|