def validation(self):
tbar = tqdm(self.valloader)
self.model.eval()
metric = IOUandSek(num_classes=len(ChangeDetection.CLASSES))
if self.args.save_mask:
cmap = color_map()
with torch.no_grad():
for img1, img2, mask1, mask2, id in tbar:
img1, img2 = img1.cuda(), img2.cuda()
out1, out2, out_bin = self.model(img1, img2, self.args.tta)
out1 = torch.argmax(out1, dim=1).cpu().numpy() + 1
out2 = torch.argmax(out2, dim=1).cpu().numpy() + 1
out_bin = (out_bin > 0.5).cpu().numpy().astype(np.uint8)
out1[out_bin == 1] = 0
out2[out_bin == 1] = 0
if self.args.save_mask:
for i in range(out1.shape[0]):
mask = Image.fromarray(out1[i].astype(np.uint8),
mode="P")
mask.putpalette(cmap)
mask.save("outdir/masks/val/im1/" + id[i])
mask = Image.fromarray(out2[i].astype(np.uint8),
mode="P")
mask.putpalette(cmap)
mask.save("outdir/masks/val/im2/" + id[i])
metric.add_batch(out1, mask1.numpy())
metric.add_batch(out2, mask2.numpy())
score, miou, sek = metric.evaluate()
tbar.set_description(
"Score: %.2f, IOU: %.2f, SeK: %.2f" %
(score * 100.0, miou * 100.0, sek * 100.0))
if self.args.load_from:
exit(0)
score *= 100.0
if score >= self.previous_best:
if self.previous_best != 0:
model_path = "outdir/models/%s_%s_%.2f.pth" % \
(self.args.model, self.args.backbone, self.previous_best)
if os.path.exists(model_path):
os.remove(model_path)
torch.save(
self.model.module.state_dict(),
"outdir/models/%s_%s_%.2f.pth" %
(self.args.model, self.args.backbone, score))
self.previous_best = score
len(dataset.CLASSES) - 1, True)
model4.load_state_dict(
torch.load('outdir/models/pspnet_hrnet_w48_37.63.pth'), strict=True)
model5 = get_model('pspnet', 'hrnet_w64', False,
len(dataset.CLASSES) - 1, True)
model5.load_state_dict(
torch.load('outdir/models/pspnet_hrnet_w64_37.89.pth'), strict=True)
models = [model1, model2, model3, model4, model5]
for i in range(len(models)):
models[i] = DataParallel(models[i]).cuda()
models[i].eval()
tbar = tqdm(dataloader)
cmap = color_map()
for img1, img2, mask1, mask2, id in tbar:
img1, img2 = img1.cuda(), img2.cuda()
pseudo_mask1_list, pseudo_mask2_list = [], []
mask1 = mask1.numpy()
mask2 = mask2.numpy()
for model in models:
with torch.no_grad():
out1, out2, out_bin = model(img1, img2, True)
out1 = torch.argmax(out1, dim=1).cpu().numpy() + 1
out2 = torch.argmax(out2, dim=1).cpu().numpy() + 1
pseudo_mask1 = np.zeros_like(out1)
pseudo_mask2 = np.zeros_like(out2)