class Visualization_demo():
def __init__(self, cfg, output_dir):
self.encoder = Encoder(cfg)
self.decoder = Decoder(cfg)
self.refiner = Refiner(cfg)
self.merger = Merger(cfg)
checkpoint = torch.load(cfg.CHECKPOINT)
encoder_state_dict = clean_state_dict(checkpoint['encoder_state_dict'])
self.encoder.load_state_dict(encoder_state_dict)
decoder_state_dict = clean_state_dict(checkpoint['decoder_state_dict'])
self.decoder.load_state_dict(decoder_state_dict)
if cfg.NETWORK.USE_REFINER:
refiner_state_dict = clean_state_dict(
checkpoint['refiner_state_dict'])
self.refiner.load_state_dict(refiner_state_dict)
if cfg.NETWORK.USE_MERGER:
merger_state_dict = clean_state_dict(
checkpoint['merger_state_dict'])
self.merger.load_state_dict(merger_state_dict)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
self.output_dir = output_dir
def run_on_images(self, imgs, sid, mid, iid, sampled_idx):
dir1 = os.path.join(output_dir, str(sid), str(mid))
if not os.path.exists(dir1):
os.makedirs(dir1)
deprocess = imagenet_deprocess(rescale_image=False)
image_features = self.encoder(imgs)
raw_features, generated_volume = self.decoder(image_features)
generated_volume = self.merger(raw_features, generated_volume)
generated_volume = self.refiner(generated_volume)
mesh = cubify(generated_volume, 0.3)
# mesh = voxel_to_world(meshes)
save_mesh = os.path.join(dir1, "%s_%s.obj" % (iid, sampled_idx))
verts, faces = mesh.get_mesh_verts_faces(0)
save_obj(save_mesh, verts, faces)
generated_volume = generated_volume.squeeze()
img = image_to_numpy(deprocess(imgs[0][0]))
save_img = os.path.join(dir1, "%02d.png" % (iid))
# cv2.imwrite(save_img, img[:, :, ::-1])
cv2.imwrite(save_img, img)
img1 = image_to_numpy(deprocess(imgs[0][1]))
save_img1 = os.path.join(dir1, "%02d.png" % (sampled_idx))
cv2.imwrite(save_img1, img1)
# cv2.imwrite(save_img1, img1[:, :, ::-1])
get_volume_views(generated_volume, dir1, iid, sampled_idx)
class Quantitative_analysis_demo():
def __init__(self, cfg, output_dir):
self.encoder = Encoder(cfg)
self.decoder = Decoder(cfg)
self.refiner = Refiner(cfg)
self.merger = Merger(cfg)
# self.thresh = cfg.VOXEL_THRESH
self.th = cfg.TEST.VOXEL_THRESH
checkpoint = torch.load(cfg.CHECKPOINT)
encoder_state_dict = clean_state_dict(checkpoint['encoder_state_dict'])
self.encoder.load_state_dict(encoder_state_dict)
decoder_state_dict = clean_state_dict(checkpoint['decoder_state_dict'])
self.decoder.load_state_dict(decoder_state_dict)
if cfg.NETWORK.USE_REFINER:
refiner_state_dict = clean_state_dict(
checkpoint['refiner_state_dict'])
self.refiner.load_state_dict(refiner_state_dict)
if cfg.NETWORK.USE_MERGER:
merger_state_dict = clean_state_dict(
checkpoint['merger_state_dict'])
self.merger.load_state_dict(merger_state_dict)
self.output_dir = output_dir
def calculate_iou(self, imgs, GT_voxels, sid, mid, iid):
dir1 = os.path.join(self.output_dir, str(sid), str(mid))
if not os.path.exists(dir1):
os.makedirs(dir1)
image_features = self.encoder(imgs)
raw_features, generated_volume = self.decoder(image_features)
generated_volume = self.merger(raw_features, generated_volume)
generated_volume = self.refiner(generated_volume)
generated_volume = generated_volume.squeeze()
sample_iou = []
for th in self.th:
_volume = torch.ge(generated_volume, th).float()
intersection = torch.sum(_volume.mul(GT_voxels)).float()
union = torch.sum(torch.ge(_volume.add(GT_voxels), 1)).float()
sample_iou.append((intersection / union).item())
return sample_iou
def train_net(cfg):
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
train_transforms = utils.data_transforms.Compose([
utils.data_transforms.RandomCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(
cfg.TRAIN.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.ColorJitter(cfg.TRAIN.BRIGHTNESS,
cfg.TRAIN.CONTRAST,
cfg.TRAIN.SATURATION),
utils.data_transforms.RandomNoise(cfg.TRAIN.NOISE_STD),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.RandomFlip(),
utils.data_transforms.RandomPermuteRGB(),
utils.data_transforms.ToTensor(),
])
val_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
# Set up data loader
train_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TRAIN_DATASET](cfg)
val_dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TEST_DATASET](cfg)
train_data_loader = torch.utils.data.DataLoader(
dataset=train_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.TRAIN, cfg.CONST.N_VIEWS_RENDERING,
train_transforms),
batch_size=cfg.CONST.BATCH_SIZE,
num_workers=cfg.TRAIN.NUM_WORKER,
pin_memory=True,
shuffle=True,
drop_last=True)
val_data_loader = torch.utils.data.DataLoader(
dataset=val_dataset_loader.get_dataset(
utils.data_loaders.DatasetType.VAL, cfg.CONST.N_VIEWS_RENDERING,
val_transforms),
batch_size=1,
num_workers=1,
pin_memory=True,
shuffle=False)
# Set up networks
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
print('[DEBUG] %s Parameters in Encoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(encoder)))
print('[DEBUG] %s Parameters in Decoder: %d.' %
(dt.now(), utils.network_utils.count_parameters(decoder)))
print('[DEBUG] %s Parameters in Refiner: %d.' %
(dt.now(), utils.network_utils.count_parameters(refiner)))
print('[DEBUG] %s Parameters in Merger: %d.' %
(dt.now(), utils.network_utils.count_parameters(merger)))
# Initialize weights of networks
encoder.apply(utils.network_utils.init_weights)
decoder.apply(utils.network_utils.init_weights)
refiner.apply(utils.network_utils.init_weights)
merger.apply(utils.network_utils.init_weights)
# Set up solver
if cfg.TRAIN.POLICY == 'adam':
encoder_solver = torch.optim.Adam(filter(lambda p: p.requires_grad,
encoder.parameters()),
lr=cfg.TRAIN.ENCODER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
decoder_solver = torch.optim.Adam(decoder.parameters(),
lr=cfg.TRAIN.DECODER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
refiner_solver = torch.optim.Adam(refiner.parameters(),
lr=cfg.TRAIN.REFINER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
merger_solver = torch.optim.Adam(merger.parameters(),
lr=cfg.TRAIN.MERGER_LEARNING_RATE,
betas=cfg.TRAIN.BETAS)
elif cfg.TRAIN.POLICY == 'sgd':
encoder_solver = torch.optim.SGD(filter(lambda p: p.requires_grad,
encoder.parameters()),
lr=cfg.TRAIN.ENCODER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
decoder_solver = torch.optim.SGD(decoder.parameters(),
lr=cfg.TRAIN.DECODER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
refiner_solver = torch.optim.SGD(refiner.parameters(),
lr=cfg.TRAIN.REFINER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
merger_solver = torch.optim.SGD(merger.parameters(),
lr=cfg.TRAIN.MERGER_LEARNING_RATE,
momentum=cfg.TRAIN.MOMENTUM)
else:
raise Exception('[FATAL] %s Unknown optimizer %s.' %
(dt.now(), cfg.TRAIN.POLICY))
# Set up learning rate scheduler to decay learning rates dynamically
encoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
encoder_solver,
milestones=cfg.TRAIN.ENCODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
decoder_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
decoder_solver,
milestones=cfg.TRAIN.DECODER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
refiner_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
refiner_solver,
milestones=cfg.TRAIN.REFINER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
merger_lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
merger_solver,
milestones=cfg.TRAIN.MERGER_LR_MILESTONES,
gamma=cfg.TRAIN.GAMMA)
if torch.cuda.is_available():
encoder = torch.nn.DataParallel(encoder).cuda()
decoder = torch.nn.DataParallel(decoder).cuda()
refiner = torch.nn.DataParallel(refiner).cuda()
merger = torch.nn.DataParallel(merger).cuda()
# Set up loss functions
bce_loss = torch.nn.BCELoss()
# Load pretrained model if exists
init_epoch = 0
best_iou = -1
best_epoch = -1
if 'WEIGHTS' in cfg.CONST and cfg.TRAIN.RESUME_TRAIN:
print('[INFO] %s Recovering from %s ...' %
(dt.now(), cfg.CONST.WEIGHTS))
checkpoint = torch.load(cfg.CONST.WEIGHTS)
init_epoch = checkpoint['epoch_idx']
best_iou = checkpoint['best_iou']
best_epoch = checkpoint['best_epoch']
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
merger.load_state_dict(checkpoint['merger_state_dict'])
print('[INFO] %s Recover complete. Current epoch #%d, Best IoU = %.4f at epoch #%d.' \
% (dt.now(), init_epoch, best_iou, best_epoch))
# Summary writer for TensorBoard
output_dir = os.path.join(cfg.DIR.OUT_PATH, '%s', dt.now().isoformat())
log_dir = output_dir % 'logs'
ckpt_dir = output_dir % 'checkpoints'
train_writer = SummaryWriter(os.path.join(log_dir, 'train'))
val_writer = SummaryWriter(os.path.join(log_dir, 'test'))
# Training loop
for epoch_idx in range(init_epoch, cfg.TRAIN.NUM_EPOCHES):
# Tick / tock
epoch_start_time = time()
# Batch average meterics
batch_time = utils.network_utils.AverageMeter()
data_time = utils.network_utils.AverageMeter()
encoder_losses = utils.network_utils.AverageMeter()
refiner_losses = utils.network_utils.AverageMeter()
# Adjust learning rate
encoder_lr_scheduler.step()
decoder_lr_scheduler.step()
refiner_lr_scheduler.step()
merger_lr_scheduler.step()
# switch models to training mode
encoder.train()
decoder.train()
merger.train()
refiner.train()
batch_end_time = time()
n_batches = len(train_data_loader)
for batch_idx, (taxonomy_names, sample_names, rendering_images,
ground_truth_volumes) in enumerate(train_data_loader):
# Measure data time
data_time.update(time() - batch_end_time)
# Get data from data loader
rendering_images = utils.network_utils.var_or_cuda(
rendering_images)
ground_truth_volumes = utils.network_utils.var_or_cuda(
ground_truth_volumes)
# Train the encoder, decoder, refiner, and merger
image_features = encoder(rendering_images)
raw_features, generated_volumes = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volumes = merger(raw_features, generated_volumes)
else:
generated_volumes = torch.mean(generated_volumes, dim=1)
encoder_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volumes = refiner(generated_volumes)
refiner_loss = bce_loss(generated_volumes,
ground_truth_volumes) * 10
else:
refiner_loss = encoder_loss
# Gradient decent
encoder.zero_grad()
decoder.zero_grad()
refiner.zero_grad()
merger.zero_grad()
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
encoder_loss.backward(retain_graph=True)
refiner_loss.backward()
else:
encoder_loss.backward()
encoder_solver.step()
decoder_solver.step()
refiner_solver.step()
merger_solver.step()
# Append loss to average metrics
encoder_losses.update(encoder_loss.item())
refiner_losses.update(refiner_loss.item())
# Append loss to TensorBoard
n_itr = epoch_idx * n_batches + batch_idx
train_writer.add_scalar('EncoderDecoder/BatchLoss',
encoder_loss.item(), n_itr)
train_writer.add_scalar('Refiner/BatchLoss', refiner_loss.item(),
n_itr)
# Tick / tock
batch_time.update(time() - batch_end_time)
batch_end_time = time()
print('[INFO] %s [Epoch %d/%d][Batch %d/%d] BatchTime = %.3f (s) DataTime = %.3f (s) EDLoss = %.4f RLoss = %.4f' % \
(dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES, batch_idx + 1, n_batches, \
batch_time.val, data_time.val, encoder_loss.item(), refiner_loss.item()))
# Append epoch loss to TensorBoard
train_writer.add_scalar('EncoderDecoder/EpochLoss', encoder_losses.avg,
epoch_idx + 1)
train_writer.add_scalar('Refiner/EpochLoss', refiner_losses.avg,
epoch_idx + 1)
# Tick / tock
epoch_end_time = time()
print('[INFO] %s Epoch [%d/%d] EpochTime = %.3f (s) EDLoss = %.4f RLoss = %.4f' %
(dt.now(), epoch_idx + 1, cfg.TRAIN.NUM_EPOCHES, epoch_end_time - epoch_start_time, \
encoder_losses.avg, refiner_losses.avg))
# Update Rendering Views
if cfg.TRAIN.UPDATE_N_VIEWS_RENDERING:
n_views_rendering = random.randint(1, cfg.CONST.N_VIEWS_RENDERING)
train_data_loader.dataset.set_n_views_rendering(n_views_rendering)
print('[INFO] %s Epoch [%d/%d] Update #RenderingViews to %d' % \
(dt.now(), epoch_idx + 2, cfg.TRAIN.NUM_EPOCHES, n_views_rendering))
# Validate the training models
iou = test_net(cfg, epoch_idx + 1, output_dir, val_data_loader,
val_writer, encoder, decoder, refiner, merger)
# Save weights to file
if (epoch_idx + 1) % cfg.TRAIN.SAVE_FREQ == 0:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
utils.network_utils.save_checkpoints(cfg, \
os.path.join(ckpt_dir, 'ckpt-epoch-%04d.pth' % (epoch_idx + 1)), \
epoch_idx + 1, encoder, encoder_solver, decoder, decoder_solver, \
refiner, refiner_solver, merger, merger_solver, best_iou, best_epoch)
if iou > best_iou:
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
best_iou = iou
best_epoch = epoch_idx + 1
utils.network_utils.save_checkpoints(cfg, \
os.path.join(ckpt_dir, 'best-ckpt.pth'), \
epoch_idx + 1, encoder, encoder_solver, decoder, decoder_solver, \
refiner, refiner_solver, merger, merger_solver, best_iou, best_epoch)
# Close SummaryWriter for TensorBoard
train_writer.close()
val_writer.close()
def test_single_img(cfg):
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
cfg.CONST.WEIGHTS = 'D:/Pix2Vox/Pix2Vox/pretrained/Pix2Vox-A-ShapeNet.pth'
checkpoint = torch.load(cfg.CONST.WEIGHTS, map_location=torch.device('cpu'))
fix_checkpoint = {}
fix_checkpoint['encoder_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['encoder_state_dict'].items())
fix_checkpoint['decoder_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['decoder_state_dict'].items())
fix_checkpoint['refiner_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['refiner_state_dict'].items())
fix_checkpoint['merger_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['merger_state_dict'].items())
epoch_idx = checkpoint['epoch_idx']
encoder.load_state_dict(fix_checkpoint['encoder_state_dict'])
decoder.load_state_dict(fix_checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
print('Use refiner')
refiner.load_state_dict(fix_checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
print('Use merger')
merger.load_state_dict(fix_checkpoint['merger_state_dict'])
encoder.eval()
decoder.eval()
refiner.eval()
merger.eval()
img1_path = 'D:/Pix2Vox/Pix2Vox/rand/minecraft.png'
img1_np = cv2.imread(img1_path, cv2.IMREAD_UNCHANGED).astype(np.float32) / 255.
sample = np.array([img1_np])
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN, std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
rendering_images = test_transforms(rendering_images=sample)
rendering_images = rendering_images.unsqueeze(0)
with torch.no_grad():
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volume = refiner(generated_volume)
generated_volume = generated_volume.squeeze(0)
img_dir = 'D:/Pix2Vox/Pix2Vox/output'
gv = generated_volume.cpu().numpy()
gv_new = np.swapaxes(gv, 2, 1)
print(gv_new)
rendering_views = utils.binvox_visualization.get_volume_views(gv_new, os.path.join(img_dir),
epoch_idx)
def test_net(cfg, epoch_idx=-1, output_dir=None, test_data_loader=None, \
test_writer=None, encoder=None, decoder=None, refiner=None, merger=None):
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Load taxonomies of dataset
taxonomies = []
with open(cfg.DATASETS[cfg.DATASET.TEST_DATASET.upper()].TAXONOMY_FILE_PATH, encoding='utf-8') as file:
taxonomies = json.loads(file.read())
taxonomies = {t['taxonomy_id']: t for t in taxonomies}
# Set up data loader
if test_data_loader is None:
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN, std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[cfg.DATASET.TEST_DATASET](cfg)
test_data_loader = torch.utils.data.DataLoader(
dataset=dataset_loader.get_dataset(utils.data_loaders.DatasetType.TEST,
cfg.CONST.N_VIEWS_RENDERING, test_transforms),
batch_size=1,
num_workers=1,
pin_memory=True,
shuffle=False)
# Set up networks
if decoder is None or encoder is None:
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
if torch.cuda.is_available():
encoder = torch.nn.DataParallel(encoder).cuda()
decoder = torch.nn.DataParallel(decoder).cuda()
refiner = torch.nn.DataParallel(refiner).cuda()
merger = torch.nn.DataParallel(merger).cuda()
print('[INFO] %s Loading weights from %s ...' % (dt.now(), cfg.CONST.WEIGHTS))
checkpoint = torch.load(cfg.CONST.WEIGHTS)
epoch_idx = checkpoint['epoch_idx']
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
merger.load_state_dict(checkpoint['merger_state_dict'])
# Set up loss functions
bce_loss = torch.nn.BCELoss()
# Testing loop
n_samples = len(test_data_loader)
test_iou = dict()
encoder_losses = utils.network_utils.AverageMeter()
refiner_losses = utils.network_utils.AverageMeter()
# Switch models to evaluation mode
encoder.eval()
decoder.eval()
refiner.eval()
merger.eval()
for sample_idx, (taxonomy_id, sample_name, rendering_images, ground_truth_volume) in enumerate(test_data_loader):
taxonomy_id = taxonomy_id[0] if isinstance(taxonomy_id[0], str) else taxonomy_id[0].item()
sample_name = sample_name[0]
with torch.no_grad():
# Get data from data loader
rendering_images = utils.network_utils.var_or_cuda(rendering_images)
ground_truth_volume = utils.network_utils.var_or_cuda(ground_truth_volume)
# Test the encoder, decoder, refiner and merger
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
encoder_loss = bce_loss(generated_volume, ground_truth_volume) * 10
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volume = refiner(generated_volume)
refiner_loss = bce_loss(generated_volume, ground_truth_volume) * 10
else:
refiner_loss = encoder_loss
print("vox shape {}".format(generated_volume.shape))
# Append loss and accuracy to average metrics
encoder_losses.update(encoder_loss.item())
refiner_losses.update(refiner_loss.item())
# IoU per sample
sample_iou = []
for th in cfg.TEST.VOXEL_THRESH:
_volume = torch.ge(generated_volume, th).float()
intersection = torch.sum(_volume.mul(ground_truth_volume)).float()
union = torch.sum(torch.ge(_volume.add(ground_truth_volume), 1)).float()
sample_iou.append((intersection / union).item())
# IoU per taxonomy
if not taxonomy_id in test_iou:
test_iou[taxonomy_id] = {'n_samples': 0, 'iou': []}
test_iou[taxonomy_id]['n_samples'] += 1
test_iou[taxonomy_id]['iou'].append(sample_iou)
# Append generated volumes to TensorBoard
if output_dir and sample_idx < 3:
img_dir = output_dir % 'images'
# Volume Visualization
gv = generated_volume.cpu().numpy()
rendering_views = utils.binvox_visualization.get_volume_views(gv, os.path.join(img_dir, 'test'),
epoch_idx)
if not test_writer is None:
test_writer.add_image('Test Sample#%02d/Volume Reconstructed' % sample_idx, rendering_views, epoch_idx)
gtv = ground_truth_volume.cpu().numpy()
rendering_views = utils.binvox_visualization.get_volume_views(gtv, os.path.join(img_dir, 'test'),
epoch_idx)
if not test_writer is None:
test_writer.add_image('Test Sample#%02d/Volume GroundTruth' % sample_idx, rendering_views, epoch_idx)
# Print sample loss and IoU
print('[INFO] %s Test[%d/%d] Taxonomy = %s Sample = %s EDLoss = %.4f RLoss = %.4f IoU = %s' % \
(dt.now(), sample_idx + 1, n_samples, taxonomy_id, sample_name, encoder_loss.item(), \
refiner_loss.item(), ['%.4f' % si for si in sample_iou]))
# Output testing results
mean_iou = []
for taxonomy_id in test_iou:
test_iou[taxonomy_id]['iou'] = np.mean(test_iou[taxonomy_id]['iou'], axis=0)
mean_iou.append(test_iou[taxonomy_id]['iou'] * test_iou[taxonomy_id]['n_samples'])
mean_iou = np.sum(mean_iou, axis=0) / n_samples
# Print header
print('============================ TEST RESULTS ============================')
print('Taxonomy', end='\t')
print('#Sample', end='\t')
print('Baseline', end='\t')
for th in cfg.TEST.VOXEL_THRESH:
print('t=%.2f' % th, end='\t')
print()
# Print body
for taxonomy_id in test_iou:
print('%s' % taxonomies[taxonomy_id]['taxonomy_name'].ljust(8), end='\t')
print('%d' % test_iou[taxonomy_id]['n_samples'], end='\t')
if 'baseline' in taxonomies[taxonomy_id]:
print('%.4f' % taxonomies[taxonomy_id]['baseline']['%d-view' % cfg.CONST.N_VIEWS_RENDERING], end='\t\t')
else:
print('N/a', end='\t\t')
for ti in test_iou[taxonomy_id]['iou']:
print('%.4f' % ti, end='\t')
print()
# Print mean IoU for each threshold
print('Overall ', end='\t\t\t\t')
for mi in mean_iou:
print('%.4f' % mi, end='\t')
print('\n')
# Add testing results to TensorBoard
max_iou = np.max(mean_iou)
if not test_writer is None:
test_writer.add_scalar('EncoderDecoder/EpochLoss', encoder_losses.avg, epoch_idx)
test_writer.add_scalar('Refiner/EpochLoss', refiner_losses.avg, epoch_idx)
test_writer.add_scalar('Refiner/IoU', max_iou, epoch_idx)
return max_iou
def test_single_img_net(cfg):
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
print('[INFO] %s Loading weights from %s ...' %
(dt.now(), cfg.CONST.WEIGHTS))
checkpoint = torch.load(cfg.CONST.WEIGHTS,
map_location=torch.device('cpu'))
fix_checkpoint = {}
fix_checkpoint['encoder_state_dict'] = OrderedDict(
(k.split('module.')[1:][0], v)
for k, v in checkpoint['encoder_state_dict'].items())
fix_checkpoint['decoder_state_dict'] = OrderedDict(
(k.split('module.')[1:][0], v)
for k, v in checkpoint['decoder_state_dict'].items())
fix_checkpoint['refiner_state_dict'] = OrderedDict(
(k.split('module.')[1:][0], v)
for k, v in checkpoint['refiner_state_dict'].items())
fix_checkpoint['merger_state_dict'] = OrderedDict(
(k.split('module.')[1:][0], v)
for k, v in checkpoint['merger_state_dict'].items())
epoch_idx = checkpoint['epoch_idx']
encoder.load_state_dict(fix_checkpoint['encoder_state_dict'])
decoder.load_state_dict(fix_checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
print('Use refiner')
refiner.load_state_dict(fix_checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
print('Use merger')
merger.load_state_dict(fix_checkpoint['merger_state_dict'])
encoder.eval()
decoder.eval()
refiner.eval()
merger.eval()
img1_path = '/media/caig/FECA2C89CA2C406F/dataset/ShapeNetRendering_copy/03001627/1a74a83fa6d24b3cacd67ce2c72c02e/rendering/00.png'
img1_np = cv2.imread(img1_path, cv2.IMREAD_UNCHANGED).astype(
np.float32) / 255.
sample = np.array([img1_np])
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
rendering_images = test_transforms(rendering_images=sample)
rendering_images = rendering_images.unsqueeze(0)
with torch.no_grad():
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volume = refiner(generated_volume)
generated_volume = generated_volume.squeeze(0)
img_dir = '/media/caig/FECA2C89CA2C406F/sketch3D/sketch3D/test_output'
gv = generated_volume.cpu().numpy()
gv_new = np.swapaxes(gv, 2, 1)
rendering_views = utils.binvox_visualization.get_volume_views(
gv_new, os.path.join(img_dir), epoch_idx)
def test_img(cfg):
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
cfg.CONST.WEIGHTS = '/Users/pranavpomalapally/Downloads/new-Pix2Vox-A-ShapeNet.pth'
checkpoint = torch.load(cfg.CONST.WEIGHTS,
map_location=torch.device('cpu'))
print()
# fix_checkpoint = {}
# fix_checkpoint['encoder_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['encoder_state_dict'].items())
# fix_checkpoint['decoder_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['decoder_state_dict'].items())
# fix_checkpoint['refiner_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['refiner_state_dict'].items())
# fix_checkpoint['merger_state_dict'] = OrderedDict((k.split('module.')[1:][0], v) for k, v in checkpoint['merger_state_dict'].items())
# fix_checkpoint['encoder_state_dict'] = OrderedDict((k.split('module.')[0], v) for k, v in checkpoint['encoder_state_dict'].items())
# fix_checkpoint['decoder_state_dict'] = OrderedDict((k.split('module.')[0], v) for k, v in checkpoint['decoder_state_dict'].items())
# fix_checkpoint['refiner_state_dict'] = OrderedDict((k.split('module.')[0], v) for k, v in checkpoint['refiner_state_dict'].items())
# fix_checkpoint['merger_state_dict'] = OrderedDict((k.split('module.')[0], v) for k, v in checkpoint['merger_state_dict'].items())
epoch_idx = checkpoint['epoch_idx']
# encoder.load_state_dict(fix_checkpoint['encoder_state_dict'])
# decoder.load_state_dict(fix_checkpoint['decoder_state_dict'])
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
# if cfg.NETWORK.USE_REFINER:
# print('Use refiner')
# refiner.load_state_dict(fix_checkpoint['refiner_state_dict'])
print('Use refiner')
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
print('Use merger')
# merger.load_state_dict(fix_checkpoint['merger_state_dict'])
merger.load_state_dict(checkpoint['merger_state_dict'])
encoder.eval()
decoder.eval()
refiner.eval()
merger.eval()
#img1_path = '/Users/pranavpomalapally/Downloads/ShapeNetRendering/02691156/1a04e3eab45ca15dd86060f189eb133/rendering/00.png'
img1_path = '/Users/pranavpomalapally/Downloads/09 copy.png'
img1_np = cv2.imread(img1_path, cv2.IMREAD_UNCHANGED).astype(
np.float32) / 255.
sample = np.array([img1_np])
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
rendering_images = test_transforms(rendering_images=sample)
rendering_images = rendering_images.unsqueeze(0)
with torch.no_grad():
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
# if cfg.NETWORK.USE_REFINER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
# generated_volume = refiner(generated_volume)
generated_volume = refiner(generated_volume)
generated_volume = generated_volume.squeeze(0)
img_dir = '/Users/pranavpomalapally/Downloads/outputs'
# gv = generated_volume.cpu().numpy()
gv = generated_volume.cpu().detach().numpy()
gv_new = np.swapaxes(gv, 2, 1)
os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
rendering_views = utils.binvox_visualization.get_volume_views(
gv_new, img_dir, epoch_idx)
def test_net(cfg, epoch_idx=-1, output_dir=None, test_data_loader=None, \
test_writer=None, encoder=None, decoder=None, refiner=None, merger=None):
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Load taxonomies of dataset
taxonomies = []
with open(
cfg.DATASETS[cfg.DATASET.TEST_DATASET.upper()].TAXONOMY_FILE_PATH,
encoding='utf-8') as file:
taxonomies = json.loads(file.read())
taxonomies = {t['taxonomy_id']: t for t in taxonomies}
# Set up data loader
if test_data_loader is None:
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(
cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TEST_DATASET](cfg)
test_data_loader = torch.utils.data.DataLoader(
dataset=dataset_loader.get_dataset(
utils.data_loaders.DatasetType.TEST,
cfg.CONST.N_VIEWS_RENDERING, test_transforms),
batch_size=1,
num_workers=1,
pin_memory=True,
shuffle=False)
# Set up networks
if decoder is None or encoder is None:
encoder = Encoder(cfg)
decoder = Decoder(cfg)
refiner = Refiner(cfg)
merger = Merger(cfg)
if torch.cuda.is_available():
encoder = torch.nn.DataParallel(encoder).cuda()
decoder = torch.nn.DataParallel(decoder).cuda()
refiner = torch.nn.DataParallel(refiner).cuda()
merger = torch.nn.DataParallel(merger).cuda()
print('[INFO] %s Loading weights from %s ...' %
(dt.now(), cfg.CONST.WEIGHTS))
if torch.cuda.is_available():
checkpoint = torch.load(cfg.CONST.WEIGHTS)
else:
map_location = torch.device('cpu')
checkpoint = torch.load(cfg.CONST.WEIGHTS,
map_location=map_location)
epoch_idx = checkpoint['epoch_idx']
print('Epoch ID of the current model is {}'.format(epoch_idx))
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
if cfg.NETWORK.USE_REFINER:
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
merger.load_state_dict(checkpoint['merger_state_dict'])
# Set up loss functions
bce_loss = torch.nn.BCELoss()
# Testing loop
n_samples = len(test_data_loader)
test_iou = dict()
encoder_losses = utils.network_utils.AverageMeter()
refiner_losses = utils.network_utils.AverageMeter()
# Switch models to evaluation mode
encoder.eval()
decoder.eval()
refiner.eval()
merger.eval()
print("test data loader type is {}".format(type(test_data_loader)))
for sample_idx, (taxonomy_id, sample_name,
rendering_images) in enumerate(test_data_loader):
taxonomy_id = taxonomy_id[0] if isinstance(
taxonomy_id[0], str) else taxonomy_id[0].item()
sample_name = sample_name[0]
print("sample IDx {}".format(sample_idx))
print("taxonomy id {}".format(taxonomy_id))
with torch.no_grad():
# Get data from data loader
rendering_images = utils.network_utils.var_or_cuda(
rendering_images)
print("Shape of the loaded images {}".format(
rendering_images.shape))
# Test the encoder, decoder, refiner and merger
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
if cfg.NETWORK.USE_REFINER:
generated_volume = refiner(generated_volume)
print("vox shape {}".format(generated_volume.shape))
gv = generated_volume.cpu().numpy()
rendering_views = utils.binvox_visualization.get_volume_views(
gv,
os.path.join('./LargeDatasets/inference_images/', 'inference'),
sample_idx)
print("gv shape is {}".format(gv.shape))
return gv, rendering_images
def test_net(cfg,
model_type,
dataset_type,
results_file_name,
epoch_idx=-1,
test_data_loader=None,
test_writer=None,
encoder=None,
decoder=None,
refiner=None,
merger=None,
save_results_to_file=False,
show_voxels=False,
path_to_times_csv=None):
if model_type == Pix2VoxTypes.Pix2Vox_A or model_type == Pix2VoxTypes.Pix2Vox_Plus_Plus_A:
use_refiner = True
else:
use_refiner = False
# Enable the inbuilt cudnn auto-tuner to find the best algorithm to use
torch.backends.cudnn.benchmark = True
# Set up data loader
if test_data_loader is None:
# Set up data augmentation
IMG_SIZE = cfg.CONST.IMG_H, cfg.CONST.IMG_W
CROP_SIZE = cfg.CONST.CROP_IMG_H, cfg.CONST.CROP_IMG_W
test_transforms = utils.data_transforms.Compose([
utils.data_transforms.CenterCrop(IMG_SIZE, CROP_SIZE),
utils.data_transforms.RandomBackground(
cfg.TEST.RANDOM_BG_COLOR_RANGE),
utils.data_transforms.Normalize(mean=cfg.DATASET.MEAN,
std=cfg.DATASET.STD),
utils.data_transforms.ToTensor(),
])
dataset_loader = utils.data_loaders.DATASET_LOADER_MAPPING[
cfg.DATASET.TEST_DATASET](cfg)
test_data_loader = torch.utils.data.DataLoader(
dataset=dataset_loader.get_dataset(dataset_type,
cfg.CONST.N_VIEWS_RENDERING,
test_transforms),
batch_size=1,
num_workers=cfg.CONST.NUM_WORKER,
pin_memory=True,
shuffle=False)
# Set up networks
if decoder is None or encoder is None:
encoder = Encoder(cfg, model_type)
decoder = Decoder(cfg, model_type)
if use_refiner:
refiner = Refiner(cfg)
merger = Merger(cfg, model_type)
if torch.cuda.is_available():
encoder = torch.nn.DataParallel(encoder).cuda()
decoder = torch.nn.DataParallel(decoder).cuda()
if use_refiner:
refiner = torch.nn.DataParallel(refiner).cuda()
merger = torch.nn.DataParallel(merger).cuda()
logging.info('Loading weights from %s ...' % (cfg.CONST.WEIGHTS))
checkpoint = torch.load(cfg.CONST.WEIGHTS)
epoch_idx = checkpoint['epoch_idx']
encoder.load_state_dict(checkpoint['encoder_state_dict'])
decoder.load_state_dict(checkpoint['decoder_state_dict'])
if use_refiner:
refiner.load_state_dict(checkpoint['refiner_state_dict'])
if cfg.NETWORK.USE_MERGER:
merger.load_state_dict(checkpoint['merger_state_dict'])
# Set up loss functions
bce_loss = torch.nn.BCELoss()
# Testing loop
n_samples = len(test_data_loader)
test_iou = dict()
encoder_losses = AverageMeter()
if use_refiner:
refiner_losses = AverageMeter()
# Switch models to evaluation mode
encoder.eval()
decoder.eval()
if use_refiner:
refiner.eval()
merger.eval()
samples_names = []
edlosses = []
rlosses = []
ious_dict = {}
for iou_threshold in cfg.TEST.VOXEL_THRESH:
ious_dict[iou_threshold] = []
if path_to_times_csv is not None:
n_view_list = []
times_list = []
for sample_idx, (taxonomy_id, sample_name, rendering_images,
ground_truth_volume) in enumerate(test_data_loader):
taxonomy_id = taxonomy_id[0] if isinstance(
taxonomy_id[0], str) else taxonomy_id[0].item()
sample_name = sample_name[0]
with torch.no_grad():
# Get data from data loader
rendering_images = utils.helpers.var_or_cuda(rendering_images)
ground_truth_volume = utils.helpers.var_or_cuda(
ground_truth_volume)
if path_to_times_csv is not None:
start_time = time.time()
# Test the encoder, decoder, refiner and merger
image_features = encoder(rendering_images)
raw_features, generated_volume = decoder(image_features)
if cfg.NETWORK.USE_MERGER and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_MERGER:
generated_volume = merger(raw_features, generated_volume)
else:
generated_volume = torch.mean(generated_volume, dim=1)
encoder_loss = bce_loss(generated_volume, ground_truth_volume) * 10
if use_refiner and epoch_idx >= cfg.TRAIN.EPOCH_START_USE_REFINER:
generated_volume = refiner(generated_volume)
refiner_loss = bce_loss(generated_volume,
ground_truth_volume) * 10
else:
refiner_loss = encoder_loss
if path_to_times_csv is not None:
end_time = time.time()
n_view_list.append(rendering_images.size()[1])
times_list.append(end_time - start_time)
# Append loss and accuracy to average metrics
encoder_losses.update(encoder_loss.item())
if use_refiner:
refiner_losses.update(refiner_loss.item())
# IoU per sample
sample_iou = []
for th in cfg.TEST.VOXEL_THRESH:
_volume = torch.ge(generated_volume, th).float()
intersection = torch.sum(
_volume.mul(ground_truth_volume)).float()
union = torch.sum(torch.ge(_volume.add(ground_truth_volume),
1)).float()
sample_iou.append((intersection / union).item())
ious_dict[th].append((intersection / union).item())
# IoU per taxonomy
if taxonomy_id not in test_iou:
test_iou[taxonomy_id] = {'n_samples': 0, 'iou': []}
test_iou[taxonomy_id]['n_samples'] += 1
test_iou[taxonomy_id]['iou'].append(sample_iou)
# Append generated volumes to TensorBoard
if show_voxels:
with open("model.binvox", "wb") as f:
v = br.Voxels(
torch.ge(generated_volume,
0.2).float().cpu().numpy()[0], (32, 32, 32),
(0, 0, 0), 1, "xyz")
v.write(f)
subprocess.run([VIEWVOX_EXE, "model.binvox"])
with open("model.binvox", "wb") as f:
v = br.Voxels(ground_truth_volume.cpu().numpy()[0],
(32, 32, 32), (0, 0, 0), 1, "xyz")
v.write(f)
subprocess.run([VIEWVOX_EXE, "model.binvox"])
# Print sample loss and IoU
logging.info(
'Test[%d/%d] Taxonomy = %s Sample = %s EDLoss = %.4f RLoss = %.4f IoU = %s'
% (sample_idx + 1, n_samples, taxonomy_id, sample_name,
encoder_loss.item(), refiner_loss.item(),
['%.4f' % si for si in sample_iou]))
samples_names.append(sample_name)
edlosses.append(encoder_loss.item())
if use_refiner:
rlosses.append(refiner_loss.item())
if save_results_to_file:
save_test_results_to_csv(samples_names,
edlosses,
rlosses,
ious_dict,
path_to_csv=results_file_name)
if path_to_times_csv is not None:
save_times_to_csv(times_list,
n_view_list,
path_to_csv=path_to_times_csv)
# Output testing results
mean_iou = []
for taxonomy_id in test_iou:
test_iou[taxonomy_id]['iou'] = np.mean(test_iou[taxonomy_id]['iou'],
axis=0)
mean_iou.append(test_iou[taxonomy_id]['iou'] *
test_iou[taxonomy_id]['n_samples'])
mean_iou = np.sum(mean_iou, axis=0) / n_samples
# Print header
print(
'============================ TEST RESULTS ============================'
)
print('Taxonomy', end='\t')
print('#Sample', end='\t')
print('Baseline', end='\t')
for th in cfg.TEST.VOXEL_THRESH:
print('t=%.2f' % th, end='\t')
print()
# Print mean IoU for each threshold
print('Overall ', end='\t\t\t\t')
for mi in mean_iou:
print('%.4f' % mi, end='\t')
print('\n')
# Add testing results to TensorBoard
max_iou = np.max(mean_iou)
if test_writer is not None:
test_writer.add_scalar('EncoderDecoder/EpochLoss', encoder_losses.avg,
epoch_idx)
if use_refiner:
test_writer.add_scalar('Refiner/EpochLoss', refiner_losses.avg,
epoch_idx)
test_writer.add_scalar('Refiner/IoU', max_iou, epoch_idx)
return max_iou