def train(train_loader, model, criterion, optimizer, epoch):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
for i, (input, target) in enumerate(train_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute pred
end = time.time()
pred = model(input)
loss = criterion(pred, target)
optimizer.zero_grad()
loss.backward() # compute gradient and do SGD step
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
if (i + 1) % args.print_freq == 0:
print('=> output: {}'.format(output_directory))
print('Train Epoch: {0} [{1}/{2}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
epoch,
i + 1,
len(train_loader),
data_time=data_time,
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
with open(train_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'gpu_time': avg.gpu_time,
'data_time': avg.data_time
})
def train(train_loader, model, criterion, optimizer, epoch, logger):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
batch_num = len(train_loader)
for i, (input, target) in enumerate(train_loader):
# itr_count += 1
input, target = input.cuda(), target.cuda()
# print('input size = ', input.size())
# print('target size = ', target.size())
torch.cuda.synchronize()
data_time = time.time() - end
# compute pred
end = time.time()
pred = model(input) # @wx 注意输出
# print('pred size = ', pred.size())
# print('target size = ', target.size())
loss = criterion(pred, target)
optimizer.zero_grad()
loss.backward() # compute gradient and do SGD step
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
if (i + 1) % args.print_freq == 0:
print('=> output: {}'.format(output_directory))
print('Train Epoch: {0} [{1}/{2}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'Loss={Loss:.5f} '
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'RML={result.absrel:.2f}({average.absrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
epoch, i + 1, len(train_loader), data_time=data_time,
gpu_time=gpu_time, Loss=loss.item(), result=result, average=average_meter.average()))
current_step = epoch * batch_num + i
logger.add_scalar('Train/RMSE', result.rmse, current_step)
logger.add_scalar('Train/rml', result.absrel, current_step)
logger.add_scalar('Train/Log10', result.lg10, current_step)
logger.add_scalar('Train/Delta1', result.delta1, current_step)
logger.add_scalar('Train/Delta2', result.delta2, current_step)
logger.add_scalar('Train/Delta3', result.delta3, current_step)
avg = average_meter.average()
def validate(epoch, valData, model, logger):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
# skip = len(valData) // 9 # save images every skip iters
for i, (image, depth) in enumerate(valData):
image = image.cuda()
depth = depth.cuda()
# normal = normal.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
end = time.time()
with torch.no_grad():
pred = model(image)
torch.cuda.synchronize()
gpu_time = time.time() - end
result = Result()
result.evaluate(pred.data, depth.data)
average_meter.update(result, gpu_time, data_time, image.size(0))
end = time.time()
if (i + 1) % 10 == 0:
print('Test Epoch: [{0}/{1}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'RML={result.absrel:.2f}({average.absrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
i + 1, len(valData), data_time=data_time,
gpu_time=gpu_time, result=result, average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'Rel={average.absrel:.3f}\n'
'Log10={average.lg10:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'Delta2={average.delta2:.3f}\n'
'Delta3={average.delta3:.3f}\n'
't_GPU={time:.3f}\n'.format(
average=avg, time=avg.gpu_time))
logger.add_scalar('Test/rmse', avg.rmse, epoch)
logger.add_scalar('Test/Rel', avg.absrel, epoch)
logger.add_scalar('Test/log10', avg.lg10, epoch)
logger.add_scalar('Test/Delta1', avg.delta1, epoch)
logger.add_scalar('Test/Delta2', avg.delta2, epoch)
logger.add_scalar('Test/Delta3', avg.delta3, epoch)
return avg
def train(train_loader, model, criterion, optimizer, epoch):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
for i, (input, target) in enumerate(train_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute pred
end = time.time()
pred = model(input)
loss = criterion(pred, target)
optimizer.zero_grad()
loss.backward() # compute gradient and do SGD step
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
depth_in = np.hstack(input.data.cpu().numpy()[:4, 3] / 10.)
depth_in = cv2.applyColorMap((depth_in * 255).astype(np.uint8), cv2.COLORMAP_HOT)
tgt_out = np.hstack(np.squeeze(target[:4].data.cpu().numpy())) / 10.
tgt_out = cv2.applyColorMap((tgt_out * 255).astype(np.uint8), cv2.COLORMAP_HOT)
out = np.hstack(np.squeeze(pred[:4].data.cpu().numpy()))
out = np.clip(out / 10., 0., 1.)
out = cv2.applyColorMap((out * 255).astype(np.uint8), cv2.COLORMAP_HOT)
if i % 20 == 0:
cv2.imshow("Training Results", np.vstack([depth_in, tgt_out, out]))
cv2.waitKey(1)
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
if (i + 1) % args.print_freq == 0:
print('=> output: {}'.format(output_directory))
print('Train Epoch: {0} [{1}/{2}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
epoch, i+1, len(train_loader), data_time=data_time,
gpu_time=gpu_time, result=result, average=average_meter.average()))
avg = average_meter.average()
with open(train_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({'mse': avg.mse, 'rmse': avg.rmse, 'absrel': avg.absrel, 'lg10': avg.lg10,
'mae': avg.mae, 'delta1': avg.delta1, 'delta2': avg.delta2, 'delta3': avg.delta3,
'gpu_time': avg.gpu_time, 'data_time': avg.data_time})
def iterate(mode, args, loader, model, optimizer, logger, epoch):
block_average_meter = AverageMeter()
average_meter = AverageMeter()
meters = [block_average_meter, average_meter]
# switch to appropriate mode
assert mode in ["train", "val", "eval", "test_prediction", "test_completion"], \
"unsupported mode: {}".format(mode)
if mode == 'train':
model.train()
lr = helper.adjust_learning_rate(args.lr, optimizer, epoch)
else:
model.eval()
lr = 0
for i, batch_data in enumerate(loader):
start = time.time()
batch_data = {
key: val.to(device)
for key, val in batch_data.items() if val is not None
}
gt = batch_data[
'gt'] if mode != 'test_prediction' and mode != 'test_completion' else None
data_time = time.time() - start
start = time.time()
with torch.no_grad(): # 自己加的
pred = model(batch_data)
depth_loss, photometric_loss, smooth_loss, mask = 0, 0, 0, None
gpu_time = time.time() - start
# measure accuracy and record loss
with torch.no_grad():
mini_batch_size = next(iter(batch_data.values())).size(0)
result = Result()
if mode != 'test_prediction' and mode != 'test_completion':
#result.evaluate(pred.data, gt.data, photometric_loss)
result.evaluate(pred.data.cpu(), gt.data.cpu(),
photometric_loss)
[
m.update(result, gpu_time, data_time, mini_batch_size)
for m in meters
]
logger.conditional_print(mode, i, epoch, lr, len(loader),
block_average_meter, average_meter)
logger.conditional_save_img_comparison(mode, i, batch_data, pred,
epoch)
logger.conditional_save_pred(mode, i, pred, epoch)
avg = logger.conditional_save_info(mode, average_meter, epoch)
is_best = logger.rank_conditional_save_best(mode, avg, epoch)
if is_best and not (mode == "train"):
logger.save_img_comparison_as_best(mode, epoch)
logger.conditional_summarize(mode, avg, is_best)
return avg, is_best
def train(epoch, trainData, model, crite, optimizer, logger):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
for i, (image, depth) in enumerate(trainData):
image = image.cuda()
depth = depth.cuda()
# normal = normal.cuda()
# image = torch.autograd.Variable(image)
# depth = torch.autograd.Variable(depth)
torch.cuda.synchronize()
data_time = time.time() - end
end = time.time()
optimizer.zero_grad()
pred = model(image)
loss = crite(pred, depth)
loss.backward()
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
result = Result()
result.evaluate(pred.data, depth.data)
average_meter.update(result, gpu_time, data_time, image.size(0))
end = time.time()
if (i + 1) % 10 == 0:
print('=> output: {}'.format(opt.output_dir))
print('Train Epoch: {0} [{1}/{2}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'Loss={Loss:.5f} '
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'RML={result.absrel:.2f}({average.absrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
epoch, i + 1, len(trainData), data_time=data_time,
gpu_time=gpu_time, Loss=loss.item(), result=result, average=average_meter.average()))
current_step = epoch * len(trainData) + i
logger.add_scalar('Train/loss', loss, current_step)
logger.add_scalar('Train/RMSE', result.rmse, current_step)
logger.add_scalar('Train/rml', result.absrel, current_step)
logger.add_scalar('Train/Log10', result.lg10, current_step)
logger.add_scalar('Train/Delta1', result.delta1, current_step)
logger.add_scalar('Train/Delta2', result.delta2, current_step)
logger.add_scalar('Train/Delta3', result.delta3, current_step)
def compute_depth_metrics(self, verbose=True) -> Result:
"""Computes metrics on the difference between raw and fixed depth values"""
avg = AverageMeter()
for i, path in enumerate(self.paths):
_, depth_raw, depth_fix = self.load_images(path)
depth_raw = torch.tensor(depth_raw)
depth_fix = torch.tensor(depth_fix)
res = Result()
res.evaluate(depth_raw, depth_fix)
avg.update(res, 0, 0, 1)
if verbose:
stdout.write(f"=> computing img {i}/{len(self)}\r")
if verbose:
stdout.write("\n")
return avg.average()
def train_coarse(train_loader, model, criterion, optimizer, epoch):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
for i, (input, target) in enumerate(train_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute pred
end = time.time()
pred = model(input)
loss = criterion(pred, target)
optimizer.zero_grad()
loss.backward() # compute gradient and do SGD step
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
end = time.time()
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
eval_time = time.time() - end
if (i + 1) % args.print_freq == 0:
history_loss.append(loss.item())
print('=> output: {}'.format(output_directory))
print('Train Epoch: {0} [{1}/{2}]\t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
epoch,
i + 1,
len(train_loader),
data_time=data_time,
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
def evaluate(params, loader, model, experiment):
print("Testing...")
with experiment.test() and torch.no_grad():
average = AverageMeter()
end = time.time()
for i, (inputs, targets) in enumerate(loader):
inputs, targets = inputs.to(params["device"]), targets.to(
params["device"])
data_time = time.time() - end
# Predict
end = time.time()
outputs = model(inputs)
gpu_time = time.time() - end
# Clip prediction
outputs[outputs > params["depth_max"]] = params["depth_max"]
outputs[outputs < params["depth_min"]] = params["depth_min"]
result = Result()
result.evaluate(outputs.data, targets.data)
average.update(result, gpu_time, data_time, inputs.size(0))
# Log images to comet
img_merged = utils.log_image_to_comet(inputs[0],
targets[0],
outputs[0],
epoch=0,
id=i,
experiment=experiment,
result=result,
prefix="visual_test")
if params["save_test_images"]:
filename = os.path.join(
params["experiment_dir"],
"image_{}_epoch_{}.png".format(i,
str(params["start_epoch"])))
utils.save_image(img_merged, filename)
def validate(val_loader, model, epoch, logger):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
skip = len(val_loader) // 9 # save images every skip iters
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
pred = model(input)
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
# save 8 images for visualization
if args.dataset == 'kitti':
rgb = input[0]
pred = pred[0]
target = target[0]
else:
rgb = input
if i == 0:
img_merge = utils.merge_into_row(rgb, target, pred)
elif (i < 8 * skip) and (i % skip == 0):
row = utils.merge_into_row(rgb, target, pred)
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip:
filename = output_directory + '/comparison_' + str(epoch) + '.png'
utils.save_image(img_merge, filename)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'RML={result.absrel:.2f}({average.absrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'Rel={average.absrel:.3f}\n'
'Log10={average.lg10:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'Delta2={average.delta2:.3f}\n'
'Delta3={average.delta3:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
logger.add_scalar('Test/rmse', avg.rmse, epoch)
logger.add_scalar('Test/Rel', avg.absrel, epoch)
logger.add_scalar('Test/log10', avg.lg10, epoch)
logger.add_scalar('Test/Delta1', avg.delta1, epoch)
logger.add_scalar('Test/Delta2', avg.delta2, epoch)
logger.add_scalar('Test/Delta3', avg.delta3, epoch)
return avg, img_merge
def test(model):
global name
test_dir = config.test_dir
test_dataset = YT_dataset(test_dir, config, is_train_set=False)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=1,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
file_name = config.test_result_name + str('.csv')
test_csv_2 = os.path.join(config.test_output_dir, file_name)
#print(test_csv_2)
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
num = 0
for i, (Y, Y_1_2, Y_1_4, Y_1_8, LR, LR_8, HR,
name) in enumerate(test_loader):
Y = Y.cuda()
LR = LR.cuda()
LR_8 = LR_8.cuda()
HR = HR.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
end = time.time()
with torch.no_grad():
# print("I am for validation in main 342")
num = num + 1
if args.arch == 'VDSR_16':
pred_HR = model(LR)
elif args.arch == 'VDSR_16_2':
pred_HR = model(Y, LR)
elif args.arch == 'VDSR':
pred_HR, residule = model(LR_8, Y)
elif args.arch == 'ResNet_bicubic':
pred_HR, residule = model(LR_8, Y)
elif args.arch == 'resnet50_15_2' or 'resnet50_15_3':
pred_HR, residule = model(Y, LR, LR_8)
else:
pred_HR, thermal0 = model(Y, LR)
gpu_time = time.time() - end
image_pred_HR = trans_norm_tensor_into_T(pred_HR)
#image_pred_HR = trans_norm_tensor_into_T(thermal0)
image_LR = trans_norm_tensor_into_rgb(LR)
image_HR = trans_norm_tensor_into_rgb(HR)
#image_HR=trans_norm_tensor_into_T(HR)
name_HR = str(
config.test_output_dir_HR) + str(num) + str('_HR') + str(".png")
name_pred_HR = str(
config.test_output_dir_pred_HR) + str(num) + str(".png")
name_HR_bicubic_HR_res = str(
config.test_output_dir_res) + str(num) + str(".png")
#name_bicubic_HR = str(config.test_output_dir_bicubic_HR) + str(num) + str(".png")
#cv2.imwrite(name_HR,image_HR[...,::-1])
#print(name_pred_HR)
#cv2.imwrite(name_pred_HR,image_pred_HR[...,::-1])
#cv2.waitKey(0)
torch.cuda.synchronize()
# measure accuracy and record loss
result = Result()
result.evaluate(pred_HR, HR)
average_meter.update(result, gpu_time, data_time,
Y.size(0)) # Y.size(0) batch_size
end = time.time()
# save 8 images for visualization,对验证集合生产图片
dir_small = '/home1/sas/datasets/crop/smallthan7/'
dir_big = '/home1/sas/datasets/crop/bigthan/'
dir_out = '/home/zju231/sas/data/old_val_resnet50_15_1/'
#dir_out='/home1/sas/results_model/pytorch/YT-SD/resnet50_30_thermal0/'
#dir_out=test_dir
name1 = dir_out + str(name[0][:-4]) + str('.png')
cv2.imwrite(name1, image_pred_HR)
name2 = dir_out + str(name[0][:-4]) + str('_HR.png')
#cv2.imwrite(name2, image_HR[..., ::-1])
name3 = dir_out + str(name[0][:-4]) + str('_Y.png')
#cv2.imwrite(name3, np.squeeze(np.array(Y.cpu())) * 255)
name4 = dir_out + str(name[0][:-4]) + str('_LR.png')
#cv2.imwrite(name4, image_LR[..., ::-1])
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'PSNR={result.psnr:.5f}({average.psnr:.5f}) '
'RMSE={result.rmse:.5f}({average.rmse:.5f}) '
'MAE={result.mae:.5f}({average.mae:.5f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(test_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'PSNR={average.psnr:.5f}\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if 1:
with open(test_csv_2, 'w') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writeheader()
writer.writerow({
'psnr': 10 * math.log(1 / (avg.mse), 10),
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
def validate(val_loader, model, epoch, batch_epoch, write_to_file=True):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
for i, (Y, Y_1_2, Y_1_4, Y_1_8, LR, LR_8, HR,
name) in enumerate(val_loader):
Y = Y.cuda()
Y_1_2 = Y_1_2.cuda()
Y_1_4 = Y_1_4.cuda()
Y_1_8 = Y_1_8.cuda()
LR_8 = LR_8.cuda()
LR = LR.cuda()
HR = HR.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
# print("I am for validation in main 342")
if args.arch == 'VDSR_16':
pred_HR = model(LR)
elif args.arch == 'VDSR_16_2':
pred_HR = model(Y, LR)
elif args.arch == 'VDSR':
pred_HR, residule = model(LR_8, Y)
elif args.arch == 'ResNet_bicubic':
pred_HR, residule = model(LR_8, Y)
elif args.arch == 'resnet50_15_6' or 'resnet50_15_11' or 'resnet50_15_12':
pred_HR = model(Y_1_2, LR)
elif args.arch == 'resnet50_15_2' or 'resnet50_15_3' or 'resnet50_15_5' or 'resnet50_15_8' or 'resnet50_15_9':
pred_HR, residule = model(Y, LR, LR_8)
else:
if config.use_different_size_Y == 1:
pred_HR, pred_thermal0 = model(Y, Y_1_2, Y_1_4, Y_1_8, LR)
else:
pred_HR, pred_thermal = model(Y, LR)
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred_HR, HR)
average_meter.update(result, gpu_time, data_time,
Y.size(0)) #Y.size(0) batch_size
end = time.time()
# save 8 images for visualization,对验证集合生产图片
skip = config.skip
if i == 0:
img_merge = utils.merge_into_row_with_YT(Y, LR, pred_HR, HR)
elif (i < 8 * skip) and (i % skip == 0):
row = utils.merge_into_row_with_YT(Y, LR, pred_HR, HR)
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip: #储存最终的图片
filename = output_directory + '/Compair_data_epoch_' + str(
epoch) + '_batch_eopch_' + str(batch_epoch + 1) + '.png'
utils.save_image(img_merge, filename)
print("生成第" + str(batch_epoch + 1) + "图片")
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'dataset epoch': epoch,
'batch epoch': batch_num + 1,
'psnr': 10 * math.log(1 / (avg.mse), 10),
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def train(train_loader, val_loader, model, criterion, optimizer, epoch, lr):
average_meter = AverageMeter()
model.train() # switch to train mode
global batch_num, best_result
end = time.time()
#every batch
for i, (Y, Y_1_2, Y_1_4, Y_1_8, LR, LR_8, HR,
name) in enumerate(train_loader): #处理被train_loader进去的每一个数据
batch_num = batch_num + 1
Y = Y.cuda()
Y_1_2 = Y_1_2.cuda()
Y_1_4 = Y_1_4.cuda()
Y_1_8 = Y_1_8.cuda()
LR = LR.cuda()
LR_8 = LR_8.cuda()
HR = HR.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
end = time.time()
if args.arch == 'VDSR_16':
pred_HR = model(LR)
loss = criterion(pred_HR, HR, Y)
elif args.arch == 'VDSR_16_2':
pred_HR = model(Y, LR)
loss = criterion(pred_HR, HR, Y)
elif args.arch == 'VDSR':
pred_HR, residule = model(LR_8, Y)
loss = criterion(pred_HR, HR, Y)
elif args.arch == 'ResNet_bicubic':
pred_HR, residule = model(LR_8, Y)
loss = criterion(pred_HR, HR, Y)
elif args.arch == 'resnet50_15_6' or 'resnet50_15_11' or 'resnet50_15_12':
pred_HR = model(Y_1_2, LR)
loss = criterion(pred_HR, HR, Y)
elif args.arch == 'resnet50_15_2' or 'resnet50_15_3' or 'resnet50_15_5' or 'resnet50_15_8' or 'resnet50_15_9':
pred_HR, residule = model(Y, LR, LR_8)
loss = criterion(pred_HR, HR, Y)
else:
if config.loss_num == 2:
if config.LOSS_1 == 'l2':
# 均方差
criterion1 = criteria.MaskedMSELoss().cuda()
elif config.LOSS_1 == 'l1':
criterion1 = criteria.MaskedL1Loss().cuda()
elif config.LOSS_1 == 'l1_canny':
# 均方差
criterion1 = criteria.MaskedL1_cannyLoss().cuda()
elif config.LOSS_1 == 'l1_from_rgb_sobel':
# 均方差
criterion1 = criteria.MaskedL1_from_rgb_sobel_Loss().cuda()
elif aconfig.LOSS_1 == 'l1_canny_from_GT_canny':
criterion1 = criteria.MaskedL1_canny_from_GT_Loss().cuda()
elif aconfig.LOSS_1 == 'l1_from_GT_sobel':
criterion1 = criteria.MaskedL1_from_GT_sobel_Loss().cuda()
elif config.LOSS_1 == 'l2_from_GT_sobel_Loss':
criterion1 = criteria.MaskedL2_from_GT_sobel_Loss().cuda()
if config.use_different_size_Y == 1:
pred_HR, pred_thermal0 = model(Y, Y_1_2, Y_1_4, Y_1_8, LR)
else:
pred_HR, pred_thermal0 = model(Y, LR)
#final loss
loss0 = criterion(pred_HR, HR, Y)
#therma upsample loss
loss1 = criterion1(pred_thermal0, HR, Y)
loss = config.LOSS_0_weight * loss0 + config.LOSS_1_weight * loss1
else:
if config.use_different_size_Y == 1:
pred_HR, pred_thermal0 = model(Y, Y_1_2, Y_1_4, Y_1_8, LR)
#writer = SummaryWriter(log_dir='logs')
#writer.add_graph(model, input_to_model=(Y,Y_1_2,Y_1_4,Y_1_8,LR,))
else:
pred_HR, pred_thermal0 = model(Y, LR)
loss = criterion(pred_HR, HR, Y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred_HR, HR, loss.cpu().detach().numpy())
average_meter.update(result, gpu_time, data_time, Y.size(0))
end = time.time()
if (i + 1) % args.print_freq == 0:
print('=> output: {}'.format(output_directory))
print('Dataset Epoch: {0} [{1}/{2}]\t'
'Batch Epoch: {3} \t'
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n'
'PSNR={result.psnr:.5f}({average.psnr:.5f}) '
'MSE={result.mse:.3f}({average.mse:.3f}) '
'RMSE={result.rmse:.3f}({average.rmse:.3f}) '
'MAE={result.mae:.3f}({average.mae:.3f}) '
'Delta1={result.delta1:.4f}({average.delta1:.4f}) '
'REL={result.absrel:.4f}({average.absrel:.4f}) '
'Lg10={result.lg10:.4f}({average.lg10:.4f}) '
'Loss={result.loss:}({average.loss:}) '.format(
epoch,
i + 1,
len(train_loader),
batch_num,
data_time=data_time,
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
else:
pass
if (batch_num + 1) % config.save_fc == 0:
print("==============Time to evaluate=================")
utils.adjust_learning_rate(optimizer, batch_num, lr)
print("==============SAVE_MODEL=================")
avg = average_meter.average()
average_meter = AverageMeter()
with open(train_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'dataset epoch': epoch,
'batch epoch': batch_num + 1,
'psnr': 10 * math.log(1 / (avg.mse), 10),
'mse': result.mse,
'rmse': result.rmse,
'absrel': result.absrel,
'lg10': result.lg10,
'mae': result.mae,
'delta1': result.delta1,
'delta2': result.delta2,
'delta3': result.delta3,
'gpu_time': result.gpu_time,
'data_time': result.data_time,
'loss': result.loss
})
#------------------#
# VALIDATION #
#------------------#
result_val, img_merge = validate(
val_loader, model, epoch,
batch_num) # evaluate on validation set，每次训练完以后都要测试一下
#------------------#
# SAVE BEST MODEL #
#------------------#
is_best = result_val.rmse < best_result.rmse
if is_best:
best_result = result_val
with open(best_txt, 'w') as txtfile:
txtfile.write(
"dataset epoch={}\nbatch epoch={}\npsnr={:.5f}\nmse={:.3f}\nrmse={:.3f}\nabsrel={:.3f}\nlg10={:.3f}\nmae={:.3f}\ndelta1={:.3f}\nt_gpu={:.4f}\n"
.format(epoch, batch_num + 1,
10 * math.log(1 / (best_result.mse), 10),
best_result.mse, best_result.rmse,
best_result.absrel, best_result.lg10,
best_result.mae, best_result.delta1,
best_result.gpu_time))
if img_merge is not None:
img_filename = output_directory + '/comparison_best.png'
utils.save_image(img_merge, img_filename)
utils.save_checkpoint(
{
'args': args,
'epoch': epoch,
'batch_epoch': batch_num,
'arch': args.arch,
'model': model,
'best_result': best_result,
'optimizer': optimizer,
}, is_best, epoch, batch_num, output_directory)
def validate(val_loader, model, write_to_file=True):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
print_freq = 10
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
pred = model(input)
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
# save 8 images for visualization
skip = 10
rgb = input
if i == 0:
import matplotlib.pyplot as plt
plt.imsave('pred.png', np.squeeze(pred.cpu().numpy()))
img_merge = utils.merge_into_row(rgb, target, pred)
elif (i < 8 * skip) and (i % skip == 0):
row = utils.merge_into_row(rgb, target, pred)
img_merge = utils.add_row(img_merge, row)
if (i + 1) % print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def val_rand(epoch):
avg_loss = 0
avg_psnr = 0
frame_count = 0
epoch_start = time.time()
end = time.time()
average_meter = AverageMeter()
modelME.module.eval()
modelME.module.netM.eval()
modelME.module.netE.eval()
for batch in val_loader:
image_target, depth_target, depth_mask = Variable(batch[0]), Variable(
batch[1]), Variable(batch[2])
batch_size = depth_target.size(0)
depth_input = depth_target.clone()
image_height = image_target.size(2)
image_width = image_target.size(3)
# Corrupt the target image
for i in range(0, batch_size):
n_depth_mask = depth_mask[i, 0, :, :].sum().item()
# Adjust the sampling rate based on the depth_mask
sample_rate = opt.sample_rate / n_depth_mask * (image_height *
image_width)
corrupt_mask = np.random.binomial(1, (1 - sample_rate),
(image_height, image_width))
corrupt_mask.astype(np.bool)
corrupt_mask = torch.BoolTensor(corrupt_mask)
depth_input[i, 0, :, :].masked_fill_(corrupt_mask, (0.0))
if torch.cuda.is_available():
image_target = image_target.cuda()
depth_input = depth_input.cuda()
depth_target = depth_target.cuda()
depth_mask = depth_mask.cuda()
rgb_sparse_d_input = torch.cat((image_target, depth_input), 1)
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
depth_prediction = modelME.module.netE(rgb_sparse_d_input)
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(depth_prediction.data, depth_target.data)
average_meter.update(result, gpu_time, data_time, depth_input.size(0))
end = time.time()
for i in range(0, depth_input.shape[0]):
loss_depth = criterion_depth(depth_prediction[[i]],
depth_target[[i]], depth_mask[[i]])
loss_mse = criterion_mse(depth_prediction[[i]], depth_target[[i]],
depth_mask[[i]])
psnr = 10 * log10(1 / loss_mse.data.item())
avg_loss += loss_depth.data.item()
avg_psnr += psnr
frame_count += 1
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
with open(val_rand_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
epoch_end = time.time()
print(
"===> Epoch {} Random Validation: Avg. Loss: {:.4f}, Avg.PSNR: {:.4f} dB, Time: {:.4f}"
.format(epoch, avg_loss / frame_count, avg_psnr / frame_count,
(epoch_end - epoch_start)))
log_value('val_loss_depth_rand', avg_loss / frame_count, epoch)
log_value('val_psnr_rand', avg_psnr / frame_count, epoch)
log_images('depth_input_rand',
reshape_4D_array(depth_input[[0], :, :, :].data.cpu().numpy(),
1),
step=1)
log_images('depth_prediction_rand',
reshape_4D_array(
depth_prediction[[0], :, :, :].data.cpu().numpy(), 1),
step=1)
log_images('depth_target_rand',
reshape_4D_array(depth_target[[0], :, :, :].data.cpu().numpy(),
1),
step=1)
log_images('image_target_rand',
reshape_4D_array(image_target[[0], :, :, :].data.cpu().numpy(),
1),
step=1)
def iterate(mode, args, loader, model, optimizer, logger, epoch):
block_average_meter = AverageMeter()
average_meter = AverageMeter()
meters = [block_average_meter, average_meter]
# switch to appropriate mode
assert mode in ["train", "val", "eval", "test_prediction", "test_completion"], \
"unsupported mode: {}".format(mode)
if mode == 'train':
model.train()
lr = helper.adjust_learning_rate(args.lr, optimizer, epoch)
else:
model.eval(
) # batchnorm or dropout layers will work in eval mode instead of training mode
lr = 0
for i, batch_data in enumerate(
loader
): # batch_data keys: 'd' (depth), 'gt' (ground truth), 'g' (gray)
start = time.time()
batch_data = {
key: val.to(device)
for key, val in batch_data.items() if val is not None
}
gt = batch_data[
'gt'] if mode != 'test_prediction' and mode != 'test_completion' else None
data_time = time.time() - start
start = time.time()
pred = model(batch_data)
if args.save_images: # save depth predictions
pred_out_dir = max(glob.glob('../outputs/var_final_NN/var.test*'),
key=os.path.getmtime) + '/dense_depth_images'
pred1 = pred.cpu().detach().numpy()[:, 0, :, :]
for im_idx, pred_im in enumerate(pred1):
pred_out_dir1 = os.path.abspath(pred_out_dir)
cur_path = os.path.abspath((loader.dataset.paths)['d'][i])
basename = os.path.basename(cur_path)
cur_dir = os.path.abspath(os.path.dirname(cur_path))
cur_dir = cur_dir.split('var_final_NN/')[1]
new_dir = os.path.abspath(pred_out_dir1 + '/' + cur_dir)
new_path = os.path.abspath(new_dir + '/' + basename)
if os.path.isdir(new_dir) == False:
os.makedirs(new_dir)
depth_write(new_path, pred_im)
depth_loss, photometric_loss, smooth_loss, mask = 0, 0, 0, None
if mode == 'train':
# Loss 1: the direct depth supervision from ground truth label
# mask=1 indicates that a pixel does not ground truth labels
if 'sparse' in args.train_mode:
depth_loss = depth_criterion(pred, batch_data['d'])
mask = (batch_data['d'] < 1e-3).float()
elif 'dense' in args.train_mode:
depth_loss = depth_criterion(pred, gt)
mask = (gt < 1e-3).float()
# Loss 2: the self-supervised photometric loss
if args.use_pose:
# create multi-scale pyramids
pred_array = helper.multiscale(pred)
rgb_curr_array = helper.multiscale(batch_data['rgb'])
rgb_near_array = helper.multiscale(batch_data['rgb_near'])
if mask is not None:
mask_array = helper.multiscale(mask)
num_scales = len(pred_array)
# compute photometric loss at multiple scales
for scale in range(len(pred_array)):
pred_ = pred_array[scale]
rgb_curr_ = rgb_curr_array[scale]
rgb_near_ = rgb_near_array[scale]
mask_ = None
if mask is not None:
mask_ = mask_array[scale]
# compute the corresponding intrinsic parameters
height_, width_ = pred_.size(2), pred_.size(3)
intrinsics_ = kitti_intrinsics.scale(height_, width_)
# inverse warp from a nearby frame to the current frame
warped_ = homography_from(rgb_near_, pred_,
batch_data['r_mat'],
batch_data['t_vec'], intrinsics_)
photometric_loss += photometric_criterion(
rgb_curr_, warped_, mask_) * (2**(scale - num_scales))
# Loss 3: the depth smoothness loss
smooth_loss = smoothness_criterion(pred) if args.w2 > 0 else 0
# backprop
loss = depth_loss + args.w1 * photometric_loss + args.w2 * smooth_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
gpu_time = time.time() - start
# measure accuracy and record loss of each batch
with torch.no_grad(
): # impacts the autograd engine and deactivate it (will reduce memory usage and speed up computations)
mini_batch_size = next(iter(batch_data.values())).size(0)
result = Result() # metrics
if mode != 'test_prediction' and mode != 'test_completion':
result.evaluate(pred.data, gt.data, photometric_loss)
[
m.update(result, gpu_time, data_time, mini_batch_size)
for m in meters
]
logger.conditional_print(mode, i, epoch, args.epochs, lr,
len(loader), block_average_meter,
average_meter)
logger.conditional_save_img_comparison(mode, i, batch_data, pred,
epoch)
logger.conditional_save_pred(mode, i, pred, epoch)
del pred
avg = logger.conditional_save_info(
mode, average_meter,
epoch) # take the avg of all the batches, to get the epoch metrics
is_best = logger.rank_conditional_save_best(mode, avg, epoch, args.epochs)
if is_best and not (mode == "train"):
logger.save_img_comparison_as_best(mode, epoch)
logger.conditional_summarize(mode, avg, is_best)
return avg, is_best
def validate(val_loader, model, epoch, write_to_file=True):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
##############################################################
## Start of PnP-Depth modification ##
##############################################################
# Original inference
with torch.no_grad():
ori_pred = model.pnp_forward_front(model.pnp_forward_rear(
input)) # equivalent to `ori_pred = model(input)`
# Inference with PnP
sparse_target = input[:, -1:] # NOTE: written for rgbd input
criterion = criteria.MaskedL1Loss().cuda(
) # NOTE: criterion function defined here only for clarity
pnp_iters = 5 # number of iterations
pnp_alpha = 0.01 # update/learning rate
pnp_z = model.pnp_forward_front(input)
for pnp_i in range(pnp_iters):
if pnp_i != 0:
pnp_z = pnp_z - pnp_alpha * torch.sign(pnp_z_grad) # iFGM
pnp_z = Variable(pnp_z, requires_grad=True)
pred = model.pnp_forward_rear(pnp_z)
if pnp_i < pnp_iters - 1:
pnp_loss = criterion(pred, sparse_target)
pnp_z_grad = Grad([pnp_loss], [pnp_z], create_graph=True)[0]
##############################################################
## End of PnP-Depth modification ##
##############################################################
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
# save 8 images for visualization
skip = 50
if args.modality == 'd':
img_merge = None
else:
if args.modality == 'rgb':
rgb = input
elif args.modality == 'rgbd':
rgb = input[:, :3, :, :]
depth = input[:, 3:, :, :]
if i == 0:
if args.modality == 'rgbd':
img_merge = utils.merge_into_row_with_gt(
rgb, depth, target, pred)
else:
img_merge = utils.merge_into_row(rgb, target, pred)
elif (i < 8 * skip) and (i % skip == 0):
if args.modality == 'rgbd':
row = utils.merge_into_row_with_gt(rgb, depth, target,
pred)
else:
row = utils.merge_into_row(rgb, target, pred)
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip:
filename = output_directory + '/comparison_' + str(
epoch) + '.png'
utils.save_image(img_merge, filename)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def iterate(mode, args, loader, model, optimizer, logger, epoch):
block_average_meter = AverageMeter()
average_meter = AverageMeter()
meters = [block_average_meter, average_meter]
# switch to appropriate mode
assert mode in ["train", "val", "eval", "test_prediction", "test_completion"], \
"unsupported mode: {}".format(mode)
if mode == 'train':
model.train()
lr = helper.adjust_learning_rate(args.lr, optimizer, epoch)
else:
model.eval()
lr = 0
torch.set_printoptions(profile="full")
table_is = np.zeros(400)
for i, batch_data in enumerate(loader):
sparse_depth_pathname = batch_data['d_path'][0]
print(sparse_depth_pathname)
del batch_data['d_path']
print("i: ", i)
start = time.time()
batch_data = {
key: val.to(device)
for key, val in batch_data.items() if val is not None
}
gt = batch_data[
'gt'] if mode != 'test_prediction' and mode != 'test_completion' else None
# adjust depth for features
depth_adjust = args.depth_adjust
adjust_features = False
if depth_adjust and args.use_d:
if args.type_feature == "sq":
if args.use_rgb:
depth_new, alg_mode, feat_mode, features, shape = depth_adjustment(
batch_data['d'], args.test_mode, args.feature_mode,
args.feature_num, args.rank_file_global_sq,
adjust_features, i, model_orig, args.seed,
batch_data['rgb'])
else:
depth_new, alg_mode, feat_mode, features, shape = depth_adjustment(
batch_data['d'], args.test_mode, args.feature_mode,
args.feature_num, args.rank_file_global_sq,
adjust_features, i, model_orig, args.seed)
elif args.type_feature == "lines":
depth_new, alg_mode, feat_mode, features = depth_adjustment_lines(
batch_data['d'], args.test_mode, args.feature_mode,
args.feature_num, args.rank_file_global_sq, i, model_orig,
args.seed)
batch_data['d'] = torch.Tensor(depth_new).unsqueeze(0).unsqueeze(
1).to(device)
data_time = time.time() - start
start = time.time()
if mode == "train":
pred = model(batch_data)
else:
with torch.no_grad():
pred = model(batch_data)
# im = batch_data['d'].detach().cpu().numpy()
# im_sq = im.squeeze()
# plt.figure()
# plt.imshow(im_sq)
# plt.show()
# for i in range(im_sq.shape[0]):
# print(f"{i} - {np.sum(im_sq[i])}")
# pred = pred +0.155
# gt = gt+0.155
# compute loss
depth_loss, photometric_loss, smooth_loss, mask = 0, 0, 0, None
if mode == 'train':
# Loss 1: the direct depth supervision from ground truth label
# mask=1 indicates that a pixel does not ground truth labels
if 'sparse' in args.train_mode:
depth_loss = depth_criterion(pred, batch_data['d'])
mask = (batch_data['d'] < 1e-3).float()
elif 'dense' in args.train_mode:
depth_loss = depth_criterion(pred, gt)
mask = (gt < 1e-3).float()
# Loss 2: the self-supervised photometric loss
if args.use_pose:
# create multi-scale pyramids
pred_array = helper.multiscale(pred)
rgb_curr_array = helper.multiscale(batch_data['rgb'])
rgb_near_array = helper.multiscale(batch_data['rgb_near'])
if mask is not None:
mask_array = helper.multiscale(mask)
num_scales = len(pred_array)
# compute photometric loss at multiple scales
for scale in range(len(pred_array)):
pred_ = pred_array[scale]
rgb_curr_ = rgb_curr_array[scale]
rgb_near_ = rgb_near_array[scale]
mask_ = None
if mask is not None:
mask_ = mask_array[scale]
# compute the corresponding intrinsic parameters
height_, width_ = pred_.size(2), pred_.size(3)
intrinsics_ = kitti_intrinsics.scale(height_, width_)
# inverse warp from a nearby frame to the current frame
warped_ = homography_from(rgb_near_, pred_,
batch_data['r_mat'],
batch_data['t_vec'], intrinsics_)
photometric_loss += photometric_criterion(
rgb_curr_, warped_, mask_) * (2**(scale - num_scales))
# Loss 3: the depth smoothness loss
smooth_loss = smoothness_criterion(pred) if args.w2 > 0 else 0
# backprop
loss = depth_loss + args.w1 * photometric_loss + args.w2 * smooth_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
gpu_time = time.time() - start
# measure accuracy and record loss
with torch.no_grad():
mini_batch_size = next(iter(batch_data.values())).size(0)
result = Result()
if mode != 'test_prediction' and mode != 'test_completion':
result.evaluate(pred.data, gt.data, photometric_loss)
[
m.update(result, gpu_time, data_time, mini_batch_size)
for m in meters
]
print(f"rmse: {result.rmse:,}")
if result.rmse < 6000:
print("good rmse")
elif result.rmse > 12000:
print("bad rmse")
logger.conditional_print(mode, i, epoch, lr, len(loader),
block_average_meter, average_meter)
logger.conditional_save_img_comparison(mode, i, batch_data, pred,
epoch)
logger.conditional_save_pred(mode, i, pred, epoch)
# save log and checkpoint
every = 999 if mode == "val" else 200
if i % every == 0 and i != 0:
print(
f"test settings (main_orig eval): {args.type_feature} {args.test_mode} {args.feature_mode} {args.feature_num}"
)
avg = logger.conditional_save_info(mode, average_meter, epoch)
is_best = logger.rank_conditional_save_best(mode, avg, epoch)
if is_best and not (mode == "train"):
logger.save_img_comparison_as_best(mode, epoch)
logger.conditional_summarize(mode, avg, is_best)
if mode != "val":
#if 1:
helper.save_checkpoint({ # save checkpoint
'epoch': epoch,
'model': model.module.state_dict(),
'best_result': logger.best_result,
'optimizer': optimizer.state_dict(),
'args': args,
}, is_best, epoch, logger.output_directory, args.type_feature, args.test_mode, args.feature_num, args.feature_mode, args.depth_adjust, i, every, "scratch")
# draw features
# run_info = [args.type_feature, alg_mode, feat_mode, model_orig]
# if batch_data['rgb'] != None and 1 and (i % 1) == 0:
# draw("sq", batch_data['rgb'], batch_data['d'], features, shape[1], run_info, i, result)
return avg, is_best
def train(params, train_loader, val_loader, model, criterion, optimizer,
scheduler, experiment):
phase_list = ["phase_1"]
current_epoch = 1
mean_val_loss = -1
try:
train_step = int(
np.ceil(params["num_training_examples"] / params["batch_size"]) *
params["start_epoch"])
val_step = int(
np.ceil(params["num_validation_examples"] / params["batch_size"] *
params["start_epoch"]))
for epoch in range(params["num_epochs"] - params["start_epoch"]):
current_epoch = params["start_epoch"] + epoch + 1
epoch_loss = 0.0
running_loss = 0.0
average = AverageMeter()
img_idxs = np.random.randint(0, len(train_loader), size=5)
model.train()
with experiment.train():
for i, (inputs, target) in enumerate(train_loader):
# Send data to GPU
inputs, target = inputs.to(params["device"]), target.to(
params["device"])
# Zero the parameter gradients
optimizer.zero_grad()
prediction = model(inputs)
loss = criterion(prediction, target, phase_list=phase_list)
# Compute loss
loss.backward()
optimizer.step()
# Calculate metrics
result = Result()
result.evaluate(prediction.data, target.data)
average.update(result, 0, 0, inputs.size(0))
epoch_loss += loss.item()
# Log to Comet
utils.log_comet_metrics(experiment,
result,
loss.item(),
step=train_step,
epoch=current_epoch)
train_step += 1
# Log images to Comet
if i in img_idxs:
utils.log_image_to_comet(inputs[0], target[0],
prediction[0], current_epoch,
i, experiment, result,
"train", train_step)
# Print statistics
running_loss += loss.item()
if (i + 1) % params["stats_frequency"] == 0 and i != 0:
print('[%d, %5d] loss: %.3f' %
(current_epoch, i + 1,
running_loss / params["stats_frequency"]))
running_loss = 0.0
# Log epoch metrics to Comet
mean_train_loss = epoch_loss / len(train_loader)
utils.log_comet_metrics(experiment,
average.average(),
mean_train_loss,
prefix="epoch",
step=train_step,
epoch=current_epoch)
# Validation each epoch
epoch_loss = 0.0
average = AverageMeter()
with experiment.validate() and torch.no_grad():
img_idxs = np.random.randint(0, len(val_loader), size=5)
model.eval()
for i, (inputs, target) in enumerate(val_loader):
inputs, target = inputs.to(params["device"]), target.to(
params["device"])
prediction = model(inputs)
loss = criterion(prediction, target, phase_list=phase_list)
# Calculate metrics
result = Result()
result.evaluate(prediction.data, target.data)
average.update(result, 0, 0, inputs.size(0))
epoch_loss += loss.item()
# Log to Comet
utils.log_comet_metrics(experiment,
result,
loss.item(),
step=val_step,
epoch=current_epoch)
val_step += 1
# Log images to Comet
if i in img_idxs:
utils.log_image_to_comet(inputs[0], target[0],
prediction[0], current_epoch,
i, experiment, result, "val",
val_step)
# Log epoch metrics to Comet
mean_val_loss = epoch_loss / len(val_loader)
utils.log_comet_metrics(experiment,
average.average(),
mean_val_loss,
prefix="epoch",
step=val_step,
epoch=current_epoch)
print("Validation Loss [%d]: %.3f" %
(current_epoch, mean_val_loss))
# Update training phase
try:
if current_epoch + 1 == params["loss"]["phase_2"]["start"]:
phase_list.append("phase_2")
if current_epoch + 1 == params["loss"]["phase_3"]["start"]:
phase_list.append("phase_3")
except KeyError:
pass
# Save periodically
if (current_epoch + 1) % params["save_frequency"] == 0:
save_path = utils.get_save_path(current_epoch,
params["experiment_dir"])
utils.save_model(model, optimizer, save_path, current_epoch,
mean_val_loss, params["max_checkpoints"])
experiment.log_model(save_path.split("/")[-1], save_path)
print("Saving new checkpoint")
experiment.log_epoch_end(current_epoch)
scheduler.step()
# Save the final model
save_path = utils.get_save_path(params["num_epochs"],
params["experiment_dir"])
utils.save_model(model, optimizer, save_path, current_epoch,
mean_val_loss, params["max_checkpoints"])
experiment.log_model(save_path.split("/")[-1], save_path)
print("Model saved to ", os.path.abspath(save_path))
except KeyboardInterrupt:
print("Saving model and quitting...")
save_path = utils.get_save_path(current_epoch,
params["experiment_dir"])
utils.save_model(model, optimizer, save_path, current_epoch,
mean_val_loss, params["max_checkpoints"])
experiment.log_model(save_path.split("/")[-1], save_path)
print("Model saved to ", os.path.abspath(save_path))
def validation(device,
data_loader,
model,
ord_loss,
output_dir,
epoch,
logger,
PRINT_FREQ,
BETA,
GAMMA,
ORD_NUM=80.0):
avg80 = AverageMeter()
avg50 = AverageMeter()
model.eval()
end = time.time()
skip = 1
img_list = []
evalbar = tqdm(total=len(data_loader))
for i, (_input, _sparse_depth, _dense_depth) in enumerate(data_loader):
_input, _sparse_depth, _dense_depth = _input.to(
device), _sparse_depth.to(device), _dense_depth.to(device)
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
_pred_prob, _pred_label = model(_input)
loss = ord_loss(_pred_prob, _dense_depth)
torch.cuda.synchronize()
gpu_time = time.time() - end
pred_depth = utils.label2depth_sid(_pred_label,
K=ORD_NUM,
alpha=1.0,
beta=BETA,
gamma=GAMMA)
abs_rel, sq_rel, rmse, rmse_log, a1, a2, a3 = compute_errors(
_sparse_depth, pred_depth.to(device))
# measure accuracy and record loss
result80 = Result()
result80.evaluate(pred_depth, _sparse_depth.data, cap=80)
result50 = Result()
result50.evaluate(pred_depth, _sparse_depth.data, cap=50)
avg80.update(result80, gpu_time, data_time, _input.size(0))
avg50.update(result50, gpu_time, data_time, _input.size(0))
end = time.time()
# save images for visualization
if i == 0:
img_merge = utils.merge_into_row(_input, _dense_depth, pred_depth)
elif (i < 8 * skip) and (i % skip == 0):
row = utils.merge_into_row(_input, _dense_depth, pred_depth)
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip:
filename = os.path.join(output_dir,
'eval_{}.png'.format(int(epoch)))
print('save validation figures at {}'.format(filename))
utils.save_image(img_merge, filename)
if (i + 1) % PRINT_FREQ == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'AbsRel={result.absrel:.2f}({average.absrel:.2f}) '
'SqRel={result.sqrel:.2f}({average.sqrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'RMSE_log={result.rmse_log:.3f}({average.rmse_log:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result80,
average=avg80.average()))
# update progress bar and show loss
evalbar.set_postfix(
ORD_LOSS=
'{:.2f},RMSE/log= {:.2f}/{:.2f},delta={:.2f}/{:.2f}/{:.2f},AbsRel/SqRe;={:.2f}/{:.2f}'
.format(loss, rmse, rmse_log, a1, a2, a3, abs_rel, sq_rel))
evalbar.update(1)
i = i + 1
print('\n**** CAP=80 ****\n'
'RMSE={average.rmse:.3f}\n'
'RMSE_log={average.rmse_log:.3f}\n'
'AbsRel={average.absrel:.3f}\n'
'SqRel={average.sqrel:.3f}\n'
'Log10={average.lg10:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'Delta2={average.delta2:.3f}\n'
'Delta3={average.delta3:.3f}\n'
'iRMSE={average.irmse:.3f}\n'
'iMAE={average.imae:.3f}\n'
't_GPU={average.gpu_time:.3f}\n'.format(average=avg80.average()))
print('\n**** CAP=50 ****\n'
'RMSE={average.rmse:.3f}\n'
'RMSE_log={average.rmse_log:.3f}\n'
'AbsRel={average.absrel:.3f}\n'
'SqRel={average.sqrel:.3f}\n'
'Log10={average.lg10:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'Delta2={average.delta2:.3f}\n'
'Delta3={average.delta3:.3f}\n'
'iRMSE={average.irmse:.3f}\n'
'iMAE={average.imae:.3f}\n'
't_GPU={average.gpu_time:.3f}\n'.format(average=avg50.average()))
logger.add_scalar('VAL_CAP80/RMSE', avg80.average().rmse, epoch)
logger.add_scalar('VAL_CAP80/RMSE_log', avg80.average().rmse_log, epoch)
logger.add_scalar('VAL_CAP80/AbsRel', avg80.average().absrel, epoch)
logger.add_scalar('VAL_CAP80/SqRel', avg80.average().sqrel, epoch)
logger.add_scalar('VAL_CAP80/Delta1', avg80.average().delta1, epoch)
logger.add_scalar('VAL_CAP80/Delta2', avg80.average().delta2, epoch)
logger.add_scalar('VAL_CAP80/Delta3', avg80.average().delta3, epoch)
logger.add_scalar('VAL_CAP50/RMSE', avg50.average().rmse, epoch)
logger.add_scalar('VAL_CAP50/RMSE_log', avg50.average().rmse_log, epoch)
logger.add_scalar('VAL_CAP50/AbsRel', avg50.average().absrel, epoch)
logger.add_scalar('VAL_CAP50/SqRel', avg50.average().sqrel, epoch)
logger.add_scalar('VAL_CAP50/Delta1', avg50.average().delta1, epoch)
logger.add_scalar('VAL_CAP50/Delta2', avg50.average().delta2, epoch)
logger.add_scalar('VAL_CAP50/Delta3', avg50.average().delta3, epoch)
def train_one_epoch(device,
train_loader,
model,
output_dir,
ord_loss,
optimizer,
epoch,
logger,
PRINT_FREQ,
BETA,
GAMMA,
ORD_NUM=80.0):
avg80 = AverageMeter()
model.train() # switch to train mode
iter_per_epoch = len(train_loader)
trainbar = tqdm(total=iter_per_epoch)
end = time.time()
for i, (_input, _sparse_depth, _dense_depth) in enumerate(train_loader):
_input, _sparse_depth, _dense_depth = _input.to(
device), _sparse_depth.to(device), _dense_depth.to(device)
torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.autograd.detect_anomaly():
_pred_prob, _pred_label = model(_input)
loss = ord_loss(
_pred_prob,
_dense_depth) # calculate ord loss with dense_depth
optimizer.zero_grad()
loss.backward()
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
pred_depth = utils.label2depth_sid(_pred_label,
K=ORD_NUM,
alpha=1.0,
beta=BETA,
gamma=GAMMA)
# calculate metrices with ground truth sparse depth
abs_rel, sq_rel, rmse, rmse_log, a1, a2, a3 = compute_errors(
_sparse_depth, pred_depth.to(device))
result80 = Result()
result80.evaluate(pred_depth, _sparse_depth.data, cap=80)
avg80.update(result80, gpu_time, data_time, _input.size(0))
end = time.time()
# update progress bar and show loss
trainbar.set_postfix(
ORD_LOSS=
'{:.2f},RMSE/log= {:.2f}/{:.2f},delta={:.2f}/{:.2f}/{:.2f},AbsRel/SqRe;={:.2f}/{:.2f}'
.format(loss, rmse, rmse_log, a1, a2, a3, abs_rel, sq_rel))
trainbar.update(1)
if (i + 1) % PRINT_FREQ == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'AbsRel={result.absrel:.2f}({average.absrel:.2f}) '
'SqRel={result.sqrel:.2f}({average.sqrel:.2f}) '
'Log10={result.lg10:.3f}({average.lg10:.3f}) '
'RMSE_log={result.rmse_log:.3f}({average.rmse_log:.3f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'Delta2={result.delta2:.3f}({average.delta2:.3f}) '
'Delta3={result.delta3:.3f}({average.delta3:.3f})'.format(
i + 1,
len(train_loader),
gpu_time=gpu_time,
result=result80,
average=avg80.average()))
current_step = int(epoch * iter_per_epoch + i + 1)
img_merge = utils.batch_merge_into_row(_input, _dense_depth.data,
pred_depth)
filename = os.path.join(output_dir,
'step_{}.png'.format(current_step))
utils.save_image(img_merge, filename)
logger.add_scalar('TRAIN/RMSE', avg80.average().rmse, current_step)
logger.add_scalar('TRAIN/RMSE_log',
avg80.average().rmse_log, current_step)
logger.add_scalar('TRAIN/AbsRel',
avg80.average().absrel, current_step)
logger.add_scalar('TRAIN/SqRel',
avg80.average().sqrel, current_step)
logger.add_scalar('TRAIN/Delta1',
avg80.average().delta1, current_step)
logger.add_scalar('TRAIN/Delta2',
avg80.average().delta2, current_step)
logger.add_scalar('TRAIN/Delta3',
avg80.average().delta3, current_step)
# reset average meter
result80 = Result()
avg80 = AverageMeter()
def depth_estimate(model, rgb_frame, depth, save=False,
switch=False):
to_tensor = transforms.ToTensor()
cmap_depth = plt.cm.viridis
cmap_error = plt.cm.inferno
rgb_np, depth_np = image_transform(rgb_frame, depth)
##creat sparse depth
mask_keep = sampler(depth_np, args.sample_spacing)
sample_number = np.count_nonzero(
mask_keep.astype('int'))
if args.modality == 'rgb':
sample_number = 0
if switch:
print('Total samples = ' + str(sample_number))
sparse_depth = np.zeros(depth_np.shape)
sparse_depth[mask_keep] = depth_np[mask_keep]
sparse_depth_np = sparse_depth
##choose input
if args.modality == 'rgb':
input_np = rgb_np
elif args.modality == 'rgbd':
rgbd = np.append(rgb_np,
np.expand_dims(sparse_depth_np,
axis=2), axis=2)
input_np = np.asfarray(rgbd, dtype='float')
elif args.modality == 'd':
input_np = sparse_depth_np
input_tensor = to_tensor(input_np)
while input_tensor.dim() < 4:
input_tensor = input_tensor.unsqueeze(0)
input_tensor = input_tensor.cuda()
torch.cuda.synchronize()
##get prediction
end = time.time()
with torch.no_grad():
pred = model(input_tensor)
torch.cuda.synchronize()
gpu_time = time.time() - end
##get result
target_tensor = to_tensor(depth_np)
while target_tensor.dim() < 4:
target_tensor = target_tensor.unsqueeze(0)
target_tensor = target_tensor.cuda()
torch.cuda.synchronize()
result = Result()
result.evaluate(pred.data, target_tensor.data)
pred_depth = np.squeeze(pred.data.cpu().numpy())
##convert depth to colour map
d_min = min(np.min(pred_depth), np.min(depth_np))
d_max = max(np.max(pred_depth), np.max(depth_np))
# # d_min = float(0.5)
# # d_max = float(6)
pred_depth_rgb_map = color_map(pred_depth, d_min, d_max,
cmap_depth)
input_depth_color_map = color_map(depth_np, d_min,
d_max, cmap_depth)
sparse_depth_color_map = color_map(sparse_depth, d_min,
d_max, cmap_depth)
##get error map
mask = depth_np <= 0
combined_map = depth_np
combined_map[mask] = pred_depth[mask]
abs_error_map = np.absolute(combined_map - pred_depth)
# error_min = min(np.min(combined_map), np.min(
# pred_depth))
# error_max = max(np.max(combined_map), np.max(
# pred_depth))
error_min = np.min(abs_error_map)
error_max = np.max(abs_error_map)
error_map_color = color_map(abs_error_map, error_min,
error_max, cmap_error)
##show images
rgb_image = rgb_np * 255
merged_image = np.hstack([rgb_image, pred_depth_rgb_map,
input_depth_color_map,
sparse_depth_color_map,
error_map_color])
##save image
if args.write and save:
images_save(sample_number, pred_depth_rgb_map,
sparse_depth_color_map, error_map_color,
rgb_image, input_depth_color_map,
result)
return merged_image, result.rmse
def validate(val_loader, model, epoch, write_to_file=True):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
eval_file = output_directory + '/evaluation.csv'
f = open(eval_file, "w+")
f.write("Max_Error,Depth,RMSE,GPU_TIME,Number_Of_Frame\r\n")
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
# torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
pred = model(input)
# torch.cuda.synchronize()
gpu_time = time.time() - end
abs_err = (target.data - pred.data).abs().cpu()
max_err_ind = np.unravel_index(np.argmax(abs_err, axis=None),
abs_err.shape)
max_err_depth = target.data[max_err_ind]
max_err = abs_err[max_err_ind]
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
f.write(
f'{max_err},{max_err_depth},{result.rmse:.2f},{gpu_time},{i+1}\r\n'
)
# save 8 images for visualization
skip = 50
if args.modality == 'rgb':
rgb = input
if i == 0:
img_merge = utils.merge_into_row_with_gt(rgb, target, pred,
(target - pred).abs())
elif (i < 8 * skip) and (i % skip == 0):
row = utils.merge_into_row_with_gt(rgb, target, pred,
(target - pred).abs())
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip:
filename = output_directory + '/comparison_' + str(epoch) + '.png'
utils.save_image(img_merge, filename)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
f.close()
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def demo(val_loader, model, epoch, write_to_file=True):
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
end = time.time()
#viz.line([[0.,0.]],[0],win='demo2',opts=dict(title='resluts2',legend=['RMSE','MAE']))
viz.line([[0., 0., 0., 0., 0., 0.]], [0],
win='demo1',
opts=dict(
title='resluts1',
legend=['t_GPU', 'Delta1', 'REL', 'lG10', 'RMSE', 'MAE']))
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
# torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
pred = model(input)
# torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
# step 50 images for visualization
skip = 10
if args.modality == 'rgb':
rgb = input
img_merge = utils.merge_into_row(rgb, target, pred)
#row = utils.merge_into_row(rgb, target, pred)
#img_merge = utils.add_row(img_merge, row)
#filename = output_directory + '/comparison_' + str(epoch) + '.png'
#utils.save_image(img_merge, filename)
#print(img_merge)
if i % skip == 0:
img = np.transpose(img_merge, (2, 0, 1))
img = torch.from_numpy(img)
viz.images(img, win='depth_estimation')
t_GPU = float('{gpu_time:.3f}'.format(gpu_time=gpu_time))
RMSE = float('{result.rmse:.2f}'.format(result=result))
MAE = float('{result.mae:.2f}'.format(result=result))
Delta1 = float('{result.delta1:.3f}'.format(result=result))
REL = float('{result.absrel:.3f}'.format(result=result))
Lg10 = float('{result.lg10:.3f}'.format(result=result))
#print(t_GPU)
#viz.line([[RMSE,MAE]],[i],win='demo2',update='append')
viz.line([[t_GPU, Delta1, REL, Lg10, RMSE, MAE]], [i],
win='demo1',
update='append')
time.sleep(0.2)
avg = average_meter.average()
viz.text('\n*\n'
'RMSE={result.rmse:.2f}({average.rmse:.3f})\n\t'
'MAE={result.mae:.2f}({average.mae:.3f})\n\t'
'Delta1={result.delta1:.3f}({average.delta1:.3f})\n\t'
'REL={result.absrel:.3f}({average.absrel:.3f})\n\t'
'Lg10={result.lg10:.3f}({average.lg10:.3f})\n\t'
't_GPU={gpu_time:.3f}{time:.3f}\n'.format(gpu_time=gpu_time,
average=avg,
time=avg.gpu_time,
result=result))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def iterate(mode, args, loader, model, optimizer, logger, epoch):
block_average_meter = AverageMeter()
average_meter = AverageMeter()
meters = [block_average_meter, average_meter]
# switch to appropriate mode
assert mode in ["train", "val", "eval", "test_prediction", "test_completion"], \
"unsupported mode: {}".format(mode)
if mode == 'train':
model.train()
lr = helper.adjust_learning_rate(args.lr, optimizer, epoch)
else:
model.eval()
lr = 0
for i, batch_data in enumerate(loader):
print("The batch data keys are {}".format(batch_data.keys()))
start = time.time()
batch_data = {
key: val.to(device)
for key, val in batch_data.items() if val is not None
}
gt = batch_data[
'gt'] if mode != 'test_prediction' and mode != 'test_completion' else None
data_time = time.time() - start
start = time.time()
temp_d = batch_data['d']
temp_gt = batch_data['gt']
temp_g = batch_data['g']
print("The depth min:{}, max:{}, shape:{}, dtype:{}".format(
torch.min(temp_d), torch.max(temp_d), temp_d.shape, temp_d.dtype))
print("The groundtruth min:{}, max:{}, shape:{}, dtype:{}".format(
torch.min(temp_gt), torch.max(temp_gt), temp_gt.shape,
temp_gt.dtype))
print("The greyscale min:{}, max:{}, shape:{}, dtype:{}".format(
torch.min(temp_g), torch.max(temp_g), temp_g.shape, temp_g.dtype))
pred = model(batch_data)
temp_out = pred.detach().cpu()
print("The output min:{}, max:{}, shape:{}, dtype:{}".format(
torch.min(temp_out), torch.max(temp_out), temp_out.shape,
temp_out.dtype))
depth_loss, photometric_loss, smooth_loss, mask = 0, 0, 0, None
if mode == 'train':
# Loss 1: the direct depth supervision from ground truth label
# mask=1 indicates that a pixel does not ground truth labels
if 'sparse' in args.train_mode:
depth_loss = depth_criterion(pred, batch_data['d'])
mask = (batch_data['d'] < 1e-3).float()
elif 'dense' in args.train_mode:
depth_loss = depth_criterion(pred, gt)
mask = (gt < 1e-3).float()
# Loss 2: the self-supervised photometric loss
if args.use_pose:
# create multi-scale pyramids
pred_array = helper.multiscale(pred)
rgb_curr_array = helper.multiscale(batch_data['rgb'])
rgb_near_array = helper.multiscale(batch_data['rgb_near'])
if mask is not None:
mask_array = helper.multiscale(mask)
num_scales = len(pred_array)
# compute photometric loss at multiple scales
for scale in range(len(pred_array)):
pred_ = pred_array[scale]
rgb_curr_ = rgb_curr_array[scale]
rgb_near_ = rgb_near_array[scale]
mask_ = None
if mask is not None:
mask_ = mask_array[scale]
# compute the corresponding intrinsic parameters
height_, width_ = pred_.size(2), pred_.size(3)
intrinsics_ = kitti_intrinsics.scale(height_, width_)
# inverse warp from a nearby frame to the current frame
warped_ = homography_from(rgb_near_, pred_,
batch_data['r_mat'],
batch_data['t_vec'], intrinsics_)
photometric_loss += photometric_criterion(
rgb_curr_, warped_, mask_) * (2**(scale - num_scales))
# Loss 3: the depth smoothness loss
smooth_loss = smoothness_criterion(pred) if args.w2 > 0 else 0
# backprop
loss = depth_loss + args.w1 * photometric_loss + args.w2 * smooth_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
gpu_time = time.time() - start
# measure accuracy and record loss
with torch.no_grad():
mini_batch_size = next(iter(batch_data.values())).size(0)
result = Result()
if mode != 'test_prediction' and mode != 'test_completion':
result.evaluate(pred.data, gt.data, photometric_loss)
[
m.update(result, gpu_time, data_time, mini_batch_size)
for m in meters
]
logger.conditional_print(mode, i, epoch, lr, len(loader),
block_average_meter, average_meter)
logger.conditional_save_img_comparison(mode, i, batch_data, pred,
epoch)
logger.conditional_save_pred(mode, i, pred, epoch)
avg = logger.conditional_save_info(mode, average_meter, epoch)
is_best = logger.rank_conditional_save_best(mode, avg, epoch)
if is_best and not (mode == "train"):
logger.save_img_comparison_as_best(mode, epoch)
logger.conditional_summarize(mode, avg, is_best)
return avg, is_best
def train(train_loader, model, criterion,smoothloss,photometric_loss,optimizer, epoch):
average_meter = AverageMeter()
model.train() # switch to train mode
end = time.time()
iheight, iwidth = 480, 640 # raw image size
fx_rgb = 5.1885790117450188e+02;
fy_rgb = 5.1946961112127485e+02;
cx_rgb = 3.2558244941119034e+02;
cy_rgb = 2.5373616633400465e+02;
new_intrinsics = Intrinsics(304,228,fx_rgb*(250.0/iheight),fy_rgb*(250.0/iheight),155.1, 121.0).cuda()
for i, (batch_data, intrinsics) in enumerate(train_loader):
#input, target = input.cuda(), target.cuda()
batch_data = {key:val.cuda() for key,val in batch_data.items() if val is not None}
oheight, owidth = intrinsics["output_size"]
#new_intrinsics = Intrinsics(owidth,oheight,intrinsics["fx"],intrinsics["fy"],intrinsics["cx"],intrinsics["cy"]).cuda()
target = batch_data['gt']
torch.cuda.synchronize()
data_time = time.time() - end
# compute pred
end = time.time()
#pred = model(input[:,3:,:,:],input[:,:3,:,:]) # (depth,image)
#candidates = {"rgb":input[:,:3,:,:], "d":input[:,3:,:,:], "gt":input[:,3:,:,:]}
#batch_data = {key:val.cuda() for key,val in candidates.items() if val is not None}
pred = model(batch_data) # (depth,image)
#loss = criterion(pred, input[:,3:,:,:]) + 0.01*smoothloss(pred)
photoloss = 0.0
if args.use_pose:
# warp near frame to current frame
#hh, ww = pred.size(2), pred.size(3)
#new_intrinsics = new_intrinsics.scale(hh,ww)
mask = (batch_data['d'] < 1e-3).float()
pred_array = multiscale(pred)
rgb_curr_array = multiscale(batch_data['rgb'])
rgb_near_array = multiscale(batch_data['rgb_near'])
mask_array = multiscale(mask)
num_scales = len(pred_array)
# compute photometric loss at multiple scales
for scale in range(len(pred_array)):
pred_ = pred_array[scale]
rgb_curr_ = rgb_curr_array[scale]
rgb_near_ = rgb_near_array[scale]
mask_ = None
if mask is not None:
mask_ = mask_array[scale]
# compute the corresponding intrinsic parameters
height_, width_ = pred_.size(2), pred_.size(3)
intrinsics_ = new_intrinsics.scale(height_, width_)
warped_ = homography_from(rgb_near_,pred_,batch_data["r_mat"],batch_data["t_vec"],intrinsics_)
#warped = homography_from(batch_data["rgb_near"],pred,batch_data["r_mat"],batch_data["t_vec"],new_intrinsics)
#photoloss = photometric_loss(batch_data["rgb"],warped,mask)
photoloss += photometric_loss(rgb_curr_,warped_,mask_)*(2**(scale-num_scales))
loss = criterion(pred, target) + 0.01*smoothloss(pred) + 0.1*photoloss
optimizer.zero_grad()
loss.backward() # compute gradient and do SGD step
optimizer.step()
torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, batch_data['rgb'].size(0))
end = time.time()
if (i + 1) % args.print_freq == 0:
print('=> output: {}'.format(output_directory))
print('Train Epoch: {0} [{1}/{2}]\t'
'LOSS={loss:.3f} '
't_Data={data_time:.3f}({average.data_time:.3f}) '
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
epoch, i+1, len(train_loader),loss=loss.item(), data_time=data_time,
gpu_time=gpu_time, result=result, average=average_meter.average()))
avg = average_meter.average()
with open(train_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({'mse': avg.mse, 'rmse': avg.rmse, 'absrel': avg.absrel, 'lg10': avg.lg10,
'mae': avg.mae, 'delta1': avg.delta1, 'delta2': avg.delta2, 'delta3': avg.delta3,
'gpu_time': avg.gpu_time, 'data_time': avg.data_time})
def iterate(mode, args, loader, model, optimizer, logger, epoch):
block_average_meter = AverageMeter()
average_meter = AverageMeter()
meters = [block_average_meter, average_meter]
# switch to appropriate mode
assert mode in ["train", "val", "eval", "test_prediction", "test_completion"], \
"unsupported mode: {}".format(mode)
if mode == 'train':
model.train()
lr = helper.adjust_learning_rate(args.lr, optimizer, epoch)
else:
model.eval()
lr = 0
torch.set_printoptions(profile="full")
for i, batch_data in enumerate(loader):
name = batch_data['name'][0]
print(name)
del batch_data['name']
print("i: ", i)
# each batch data is 1 and has three keys d, gt, g and dim [1, 352, 1216]
start = time.time()
batch_data = {
key: val.to(device)
for key, val in batch_data.items() if val is not None
}
gt = batch_data[
'gt'] if mode != 'test_prediction' and mode != 'test_completion' else None
# if args.type_feature=="sq":
# depth_adjustment(gt, False)
data_time = time.time() - start
start = time.time()
if mode == "train":
pred = model(batch_data)
else:
with torch.no_grad():
pred = model(batch_data)
vis=False
if vis:
im = batch_data['gt'].detach().cpu().numpy()
im_sq = im.squeeze()
plt.figure()
plt.imshow(im_sq)
plt.show()
# for i in range(im_sq.shape[0]):
# print(f"{i} - {np.sum(im_sq[i])}")
depth_loss, photometric_loss, smooth_loss, mask = 0, 0, 0, None
if mode == 'train':
# Loss 1: the direct depth supervision from ground truth label
# mask=1 indicates that a pixel does not ground truth labels
if 'sparse' in args.train_mode:
depth_loss = depth_criterion(pred, batch_data['d'])
print("d pts: ", len(torch.where(batch_data['d']>0)[0]))
mask = (batch_data['d'] < 1e-3).float()
elif 'dense' in args.train_mode:
depth_loss = depth_criterion(pred, gt)
mask = (gt < 1e-3).float()
# Loss 2: the self-supervised photometric loss
if args.use_pose:
# create multi-scale pyramids
pred_array = helper.multiscale(pred)
rgb_curr_array = helper.multiscale(batch_data['rgb'])
rgb_near_array = helper.multiscale(batch_data['rgb_near'])
if mask is not None:
mask_array = helper.multiscale(mask)
num_scales = len(pred_array)
# compute photometric loss at multiple scales
for scale in range(len(pred_array)):
pred_ = pred_array[scale]
rgb_curr_ = rgb_curr_array[scale]
rgb_near_ = rgb_near_array[scale]
mask_ = None
if mask is not None:
mask_ = mask_array[scale]
# compute the corresponding intrinsic parameters
height_, width_ = pred_.size(2), pred_.size(3)
intrinsics_ = kitti_intrinsics.scale(height_, width_)
# inverse warp from a nearby frame to the current frame
warped_ = homography_from(rgb_near_, pred_,
batch_data['r_mat'],
batch_data['t_vec'], intrinsics_)
photometric_loss += photometric_criterion(
rgb_curr_, warped_, mask_) * (2**(scale - num_scales))
# Loss 3: the depth smoothness loss
smooth_loss = smoothness_criterion(pred) if args.w2 > 0 else 0
# backprop
loss = depth_loss + args.w1 * photometric_loss + args.w2 * smooth_loss
optimizer.zero_grad()
loss.backward()
zero_params(model)
optimizer.step()
gpu_time = time.time() - start
# counting pixels in each bin
#binned_pixels = np.load("value.npy", allow_pickle=True)
#print(len(binned_pixels))
if (i % 1 == 0 and args.evaluate and args.instancewise) or\
(i % args.every == 0 and not args.evaluate and not args.instancewise): # global training
# print(model.module.conv4[5].conv1.weight[0])
# print(model.conv4.5.bn2.weight)
# print(model.module.parameter.grad)
#print("*************swiches:")
torch.set_printoptions(precision=7, sci_mode=False)
if model.module.phi is not None:
mmp = 1000 * model.module.phi
phi = F.softplus(mmp)
S = phi / torch.sum(phi)
#print("S", S[1, -10:])
S_numpy= S.detach().cpu().numpy()
if args.instancewise:
global Ss
if "Ss" not in globals():
Ss = []
Ss.append(S_numpy)
else:
Ss.append(S_numpy)
# GLOBAL
if (i % args.every ==0 and not args.evaluate and not args.instancewise and model.module.phi is not None):
np.set_printoptions(precision=4)
switches_2d_argsort = np.argsort(S_numpy, None) # 2d to 1d sort torch.Size([9, 31])
switches_2d_sort = np.sort(S_numpy, None)
print("Switches: ")
print(switches_2d_argsort[:10])
print(switches_2d_sort[:10])
print("and")
print(switches_2d_argsort[-10:])
print(switches_2d_sort[-10:])
##### saving global ranks
global_ranks_path = lambda \
ii: f"ranks/{args.type_feature}/global/{folder_and_name[0]}/Ss_val_{folder_and_name[1]}_iter_{ii}.npy"
global old_i
if ("old_i" in globals()):
print("old_i")
if os.path.isfile(global_ranks_path(old_i)):
os.remove(global_ranks_path(old_i))
folder_and_name = args.resume.split(os.sep)[-2:]
os.makedirs(f"ranks/{args.type_feature}/global/{folder_and_name[0]}", exist_ok=True)
np.save(global_ranks_path(i), S_numpy)
old_i = i
print("saving ranks")
if args.type_feature == "sq":
hor = switches_2d_argsort % S_numpy.shape[1]
ver = np.floor(switches_2d_argsort // S_numpy.shape[1])
print(ver[:10],hor[:10])
print("and")
print(ver[-10:], hor[-10:])
# measure accuracy and record loss
with torch.no_grad():
mini_batch_size = next(iter(batch_data.values())).size(0)
result = Result()
if mode != 'test_prediction' and mode != 'test_completion':
result.evaluate(pred.data, gt.data, photometric_loss)
[
m.update(result, gpu_time, data_time, mini_batch_size)
for m in meters
]
logger.conditional_print(mode, i, epoch, lr, len(loader),
block_average_meter, average_meter)
logger.conditional_save_img_comparison(mode, i, batch_data, pred,
epoch)
logger.conditional_save_pred(mode, i, pred, epoch)
draw=False
if draw:
ma = batch_data['rgb'].detach().cpu().numpy().squeeze()
ma = np.transpose(ma, axes=[1, 2, 0])
# ma = np.uint8(ma)
#ma2 = Image.fromarray(ma)
ma2 = Image.fromarray(np.uint8(ma)).convert('RGB')
# create rectangle image
img1 = ImageDraw.Draw(ma2)
if args.type_feature == "sq":
size=40
print_square_num = 20
for ii in range(print_square_num):
s_hor=hor[-ii].detach().cpu().numpy()
s_ver=ver[-ii].detach().cpu().numpy()
shape = [(s_hor * size, s_ver * size), ((s_hor + 1) * size, (s_ver + 1) * size)]
img1.rectangle(shape, outline="red")
tim = time.time()
lala = ma2.save(f"switches_photos/squares/squares_{tim}.jpg")
print("saving")
elif args.type_feature == "lines":
print_square_num = 20
r=1
parameter_mask = np.load("../kitti_pixels_to_lines.npy", allow_pickle=True)
# for m in range(10,50):
# im = Image.fromarray(parameter_mask[m]*155)
# im = im.convert('1') # convert image to black and white
# im.save(f"switches_photos/lala_{m}.jpg")
for ii in range(print_square_num):
points = parameter_mask[ii]
y = np.where(points==1)[0]
x = np.where(points == 1)[1]
for p in range(len(x)):
img1.ellipse((x[p] - r, y[p] - r, x[p] + r, y[p] + r), fill=(255, 0, 0, 0))
tim = time.time()
lala = ma2.save(f"switches_photos/lines/lines_{tim}.jpg")
print("saving")
every = args.every
if i % every ==0:
print("saving")
avg = logger.conditional_save_info(mode, average_meter, epoch)
is_best = logger.rank_conditional_save_best(mode, avg, epoch)
#is_best = True #saving all the checkpoints
if is_best and not (mode == "train"):
logger.save_img_comparison_as_best(mode, epoch)
logger.conditional_summarize(mode, avg, is_best)
if mode != "val":
helper.save_checkpoint({ # save checkpoint
'epoch': epoch,
'model': model.module.state_dict(),
'best_result': logger.best_result,
'optimizer': optimizer.state_dict(),
'args': args,
}, is_best, epoch, logger.output_directory, args.type_feature, i, every, qnet=True)
if args.evaluate and args.instancewise:
#filename = os.path.split(args.evaluate)[1]
Ss_numpy = np.array(Ss)
folder_and_name = args.evaluate.split(os.sep)[-3:]
os.makedirs(f"ranks/instance/{folder_and_name[0]}", exist_ok=True)
os.makedirs(f"ranks/instance/{folder_and_name[0]}/{folder_and_name[1]}", exist_ok=True)
np.save(f"ranks/instance/{folder_and_name[0]}/{folder_and_name[1]}/Ss_val_{folder_and_name[2]}.npy", Ss)
return avg, is_best
def evaluate(val_loader, model, epoch, write_to_file=True):
#print(model)
average_meter = AverageMeter()
model.eval() # switch to evaluate mode
for i, (input, target, mask, h5name) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
input_var = torch.autograd.Variable(input)
target_var = torch.autograd.Variable(target)
depth_pred = model(input_var)
result = Result()
output1 = torch.index_select(depth_pred.data, 1,
torch.cuda.LongTensor([0]))
result.evaluate(output1, target)
average_meter.update(result, input.size(0))
if args.modality == 'd':
img_merge = None
else:
if args.modality == 'rgb':
rgb = input
elif args.modality == 'rgbd':
rgb = input[:, :3, :, :]
sparse_d = input[:, -1, :, :]
sparse_depth_np = utils.depth_to_np(sparse_d)
sd_filename = output_directory + '/' + str(
epoch) + '_' + '_' + h5name[0] + '_s_depth.png'
utils.save_image(sparse_depth_np, sd_filename)
rgb_np, depth_np, pred_np = utils.image_3_row(
rgb, target, depth_pred)
rgb_filename = output_directory + '/' + str(
epoch) + '_' + '_' + h5name[0] + '_rgb.png'
utils.save_image(rgb_np, rgb_filename)
depth_filename = output_directory + '/' + str(
epoch) + '_' + '_' + h5name[0] + '_depth.png'
utils.save_image(depth_np, depth_filename)
pred_filename = output_directory + '/' + str(
epoch) + '_' + '_' + h5name[0] + '_pred.png'
utils.save_image(pred_np, pred_filename)
avg = average_meter.average()
print('\n\nnew prediction with data:\n*\n'
'RMSE={average.rmse:.5f}\n'
'MAE={average.mae:.5f}\n'
'Delta1={average.delta1:.5f}\n'
'Delta2={average.delta2:.5f}\n'
'Delta3={average.delta3:.5f}\n'
'REL={average.absrel:.5f}\n'
'Lg10={average.lg10:.5f}\n'.format(average=avg))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3
})
def validate(val_loader, model, epoch, write_to_file=True):
average_meter = AverageMeter()
# switch to evaluate mode
model.eval()
end = time.time()
for i, (input, target) in enumerate(val_loader):
input, target = input.cuda(), target.cuda()
# torch.cuda.synchronize()
data_time = time.time() - end
# compute output
end = time.time()
with torch.no_grad():
pred = model(input)
# torch.cuda.synchronize()
gpu_time = time.time() - end
# measure accuracy and record loss
result = Result()
result.evaluate(pred.data, target.data)
average_meter.update(result, gpu_time, data_time, input.size(0))
end = time.time()
# save 8 images for visualization
skip = 50
if args.modality == 'd':
img_merge = None
else:
if args.modality == 'rgb':
rgb = input
elif args.modality == 'rgbd':
rgb = input[:, :3, :, :]
depth = input[:, 3:, :, :]
if i == 0:
if args.modality == 'rgbd':
img_merge = utils.merge_into_row_with_gt(
rgb, depth, target, pred)
else:
img_merge = utils.merge_into_row(rgb, target, pred)
elif (i < 8 * skip) and (i % skip == 0):
if args.modality == 'rgbd':
row = utils.merge_into_row_with_gt(rgb, depth, target,
pred)
else:
row = utils.merge_into_row(rgb, target, pred)
img_merge = utils.add_row(img_merge, row)
elif i == 8 * skip:
filename = output_directory + '/comparison_' + str(
epoch) + '.png'
utils.save_image(img_merge, filename)
if (i + 1) % args.print_freq == 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f})\n\t'
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(val_loader),
gpu_time=gpu_time,
result=result,
average=average_meter.average()))
avg = average_meter.average()
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if write_to_file:
with open(test_csv, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
return avg, img_merge
def evaluate(params, loader, model, experiment):
print("Testing...")
with experiment.test():
with torch.no_grad():
average = AverageMeter()
img_idxs = np.random.randint(0,
len(loader),
size=min(len(loader), 50))
end = time.time()
for i, (inputs, targets) in enumerate(loader):
inputs, targets = inputs.to(params["device"]), targets.to(
params["device"])
data_time = time.time() - end
# Predict
end = time.time()
outputs = model(inputs)
gpu_time = time.time() - end
result = Result()
result.evaluate(outputs.data, targets.data)
average.update(result, gpu_time, data_time, inputs.size(0))
end = time.time()
# Log images to comet
if i in img_idxs:
img_merged = utils.log_image_to_comet(
inputs[0],
targets[0],
outputs[0],
epoch=0,
id=i,
experiment=experiment,
result=result,
prefix="test")
if params["save_test_images"]:
filename = os.path.join(
params["experiment_dir"],
"comparison_epoch_{}_{}.png".format(
str(params["start_epoch"]),
np.where(img_idxs == i)[0][0]))
utils.save_image(img_merged, filename)
if (i + 1) % params["stats_frequency"] == 0 and i != 0:
print('Test: [{0}/{1}]\t'
't_GPU={gpu_time:.3f}({average.gpu_time:.3f})\n\t'
'RMSE={result.rmse:.2f}({average.rmse:.2f}) '
'MAE={result.mae:.2f}({average.mae:.2f}) '
'Delta1={result.delta1:.3f}({average.delta1:.3f}) '
'REL={result.absrel:.3f}({average.absrel:.3f}) '
'Lg10={result.lg10:.3f}({average.lg10:.3f}) '.format(
i + 1,
len(loader),
gpu_time=gpu_time,
result=result,
average=average.average()))
# Mean validation loss
avg = average.average()
utils.log_comet_metrics(experiment, avg, None)
print('\n*\n'
'RMSE={average.rmse:.3f}\n'
'MAE={average.mae:.3f}\n'
'Delta1={average.delta1:.3f}\n'
'REL={average.absrel:.3f}\n'
'Lg10={average.lg10:.3f}\n'
't_GPU={time:.3f}\n'.format(average=avg, time=avg.gpu_time))
if params["save_test_metrics"]:
filename = os.path.join(params["experiment_dir"],
"results.csv")
with open(filename, 'a') as csvfile:
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writerow({
'mse': avg.mse,
'rmse': avg.rmse,
'absrel': avg.absrel,
'lg10': avg.lg10,
'mae': avg.mae,
'delta1': avg.delta1,
'delta2': avg.delta2,
'delta3': avg.delta3,
'irmse': avg.irmse,
'imae': avg.imae,
'data_time': avg.data_time,
'gpu_time': avg.gpu_time
})
