def all_docs():
with open(filename, newline='') as doc_file:
fields = get_fieldnames(doc_file)
dict_reader = csv.DictReader(doc_file, fieldnames=fields)
if 'ticket' in doc_type:
fields.append("ticket_time")
echo('Using the following ' + str(len(fields)) + ' fields:',
quiet)
for field in fields:
echo(field, quiet)
i = 0
for row in dict_reader:
# Prepare meta info for each indexed document.
meta = {
'index': idx_name,
'type': doc_type,
}
if id_field_idx is not None:
meta['id'] = row[fields[int(id_field_idx)]]
# Convert tim inteval to an integer in minutes.
for k, v in row.items():
if isinstance(v, str) and isperiod(v):
row[k] = t2i(v)
if 'ticket' in doc_type:
row['ticket_time'] = time_interval(row['create_time'],
row['close_time'],
'%m/%d/%Y %I:%M:%S %p')
i += 1
echo('Sending item %s to ES ...' % i, quiet)
yield index_op(row, meta)
def all_docs():
with open(filename, newline='') as doc_file:
fields = get_fieldnames(doc_file)
dict_reader = csv.DictReader(doc_file, fieldnames=fields)
if 'ticket' in doc_type:
fields.append("ticket_time")
echo('Using the following ' + str(len(fields)) + ' fields:',
quiet)
for field in fields:
echo(field, quiet)
i = 0
for row in dict_reader:
# Prepare meta info for each indexed document.
meta = {
'index': idx_name,
'type': doc_type,
}
if id_field_idx is not None:
meta['id'] = row[fields[int(id_field_idx)]]
# Convert tim inteval to an integer in minutes.
for k, v in row.items():
if isinstance(v, str) and isperiod(v):
row[k] = t2i(v)
if 'ticket' in doc_type:
row['ticket_time'] = time_interval(row['create_time'],
row['close_time'],
'%m/%d/%Y %I:%M:%S %p')
i += 1
echo('Sending item %s to ES ...' % i, quiet)
yield index_op(row, meta)
def grad_desc(original_image, distorted_image, init_kernel, criteria, dump):
kernel = np.expand_dims(init_kernel, axis=0)
kernel = np.expand_dims(kernel, axis=0)
kernel = torch.tensor(kernel.astype(np.float32), requires_grad=True)
optimizer = torch.optim.Adam([kernel], 0.5)
for i in range(1000):
optimizer.zero_grad()
loss_v = criteria(conv(original_image, kernel, True), distorted_image,
True) + 0.01 * torch.mean(kernel**2)
loss_v.backward()
optimizer.step()
print('I: {} | LOSS: {:.10f}'.format(i, loss_v))
# kernel.data /= torch.sum(kernel.data)
dump['found_kernel'] = t2i(kernel)
def test_2_time2integer(self):
self.assertEqual(t2i('03:30:18'), 210)
def test_2_time2integer(self):
self.assertEqual(t2i('03:30:18'), 210)
def val(model, dataloader, vis, boxShow=False):
model.eval()
segCFM = EvalMatrix()
ancCFM = EvalMatrix()
seg_Loss, anc_Loss, reg_Loss, dice = 0, 0, 0, 0
seg_Precision, seg_Sensitive, anc_Precision, anc_Sensitive = 0, 0, 0, 0
RegCriterion = BBoxRegLoss()
AncCriterion = AnchorLoss()
SegCriterion = SegLoss()
totalNum = len(dataloader)
with torch.no_grad():
for ii, (img, mask, anchors_labels, bbox_labels, ancLossIdx,
bboxLossIdx, segIdx, GtBoxes) in enumerate(tqdm(dataloader)):
segCFM.reset()
img = V(img).float().cuda()
mask = V(mask).float().cuda()
seg, ancOutput, boxOutput = model(img)
vis.show_hist(name="Val Anchor Output", tensor=ancOutput)
# ----------------------Show Anchor Location--------
for threshold in np.arange(0.4, 0.6, 0.1):
image_ = mapAnc2Img(ancOutput, img, threshold=threshold)
vis.show_anchor(name="Anc Loc {}".format(threshold),
imgs=image_)
# ----------------------Segmentation-------------
temp_seg = 0
if seg is not None:
temp_seg = SegCriterion(seg, mask, segIdx)
if temp_seg != 0:
dice += Iou_loss(seg, mask)
vis.plot("Dice in Iteration:", t2i(Iou_loss(seg, mask)))
# --------------------Loss Computation--------------
if temp_seg != 0 and not t.isnan(temp_seg):
seg_Loss += temp_seg
vis.plot("Follow Seg Loss:", t2i(temp_seg))
for i in range(config.Output_features):
temp_reg = RegCriterion(
V(boxOutput[i]).cuda(),
V(bbox_labels[i]).cuda(),
V(bboxLossIdx[i]).cuda())
if temp_reg != 0 and not t.isnan(temp_reg):
reg_Loss += temp_reg
temp_anc = AncCriterion(
V(ancOutput[i]).cuda(),
V(anchors_labels[i]).cuda())
if temp_anc != 0 and not t.isnan(temp_anc):
anc_Loss += temp_anc
# ---------------Show Bounding Box----------------
if boxShow and temp_anc / config.Output_features < 1.0 and temp_anc != 0:
for j in range(config.ValBatchSize):
boxes_input = [
boxOutput[i][j:j + 1]
for i in range(config.Output_features)
] # bbox:[features, batch, 9, h, w, 4], bbox:[batch, features, 9, h, w, 1]
anchors_input = [
ancOutput[i][j:j + 1]
for i in range(config.Output_features)
] # anchor:[features, batch, 9, h, w, 4], anchor:[batch, features, 9, h, w, 1]
output_anchcors, output_bboxs = apply_bboxReg_area(
bboxMat=boxes_input,
img_shape=config.Image_shape,
anchors=anchors_input)
box = []
box_Precision = 0
box_Recall = 0
for i in range(config.Output_features):
finalBox = nms(output_bboxs[i], output_anchcors[i])
if finalBox is None:
continue
box.append(finalBox)
if len(box) != 0:
box = t.stack(box, dim=0)
vis.show_boxes(mask[j], box)
if len(GtBoxes) != 0:
tmp_p, tmp_r = PrecisionRecall(GtBoxes, finalBox)
box_Precision += tmp_p
box_Recall += tmp_r
vis.log("Box Precision:" + myStr(box_Precision))
vis.log("Box Recall:" + myStr(box_Recall))
else:
continue
if seg is not None:
segCFM.genConfusionMat(seg.clone(), mask.clone())
seg_Precision += segCFM.precision()
seg_Sensitive += segCFM.sensitive()
for i in range(config.Output_features):
ancCFM.genConfusionMat(ancOutput[i].clone(),
anchors_labels[i].clone())
anc_Precision += ancCFM.precision() / config.Output_features
anc_Sensitive += ancCFM.sensitive() / config.Output_features
Val_avgLoss = (seg_Loss + anc_Loss + reg_Loss) / totalNum
vis.log("Validation Dice:" + myStr(dice / totalNum), isTrain=False)
vis.log("Validation Seg Loss:" + myStr(seg_Loss / totalNum), isTrain=False)
vis.log("Validation BoxReg Loss:" + myStr(reg_Loss / totalNum),
isTrain=False)
vis.log("Validation Anchor Loss:" + myStr(anc_Loss / totalNum),
isTrain=False)
vis.plot("Avg Val Iou:", t2i(dice / totalNum))
vis.plot("Avg Val Loss for Seg: ", t2i(seg_Loss / totalNum))
vis.plot("Avg Val Loss for Regression:", t2i(reg_Loss / totalNum))
vis.plot("Avg Val Loss for Anchor Net:", t2i(anc_Loss / totalNum))
vis.plot("Avg Val Precision for Seg: ", t2i(seg_Precision / totalNum))
vis.plot("Avg Val Sensitive for Seg: ", t2i(seg_Sensitive / totalNum))
vis.plot("Avg Val Precision for Anchor Net:",
t2i(anc_Precision / totalNum))
vis.plot("Avg Val Sensitive for Anchor Net:",
t2i(anc_Sensitive / totalNum))
# -------------Weight Statistics----------------
for name in model.module.ANCNets.state_dict():
if name.endswith("weight"):
vis.show_hist(name=name,
tensor=model.module.ANCNets.state_dict()[name])
model.train()
return Val_avgLoss
def train(**kwargs):
vis = Visualizer(env=config.Env, port=config.vis_port)
model = nn.DataParallel(
RetinaUNet(base=config.Base,
InChannel=config.InputChannel,
OutChannel=1,
BackBone='fpn',
IncludeTop=False,
Godown=False,
IncludeSeg=False)).to(config.Device)
model.train()
scaler = torch.cuda.amp.GradScaler()
Seg_Matrix = EvalMatrix()
Anc_Matrix = EvalMatrix()
if config.checkpoint:
try:
if isinstance(model, nn.DataParallel):
model.module.load(get_pth(model, config.checkpoint))
else:
model.load(get_pth(model, config.checkpoint))
print("Load Model Successfully")
except:
pass
if config.Data_type.lower() == '2dplus':
img_read = read_2dPlus
img_lst = get_Path_2dPlus()
elif config.Data_type.lower() == '2d':
img_read = read_image_2D
img_lst = get_ImgPath()
else:
raise ValueError
train_lst, val_lst = data_split_2D(img_lst,
ratio=(1 - config.Val_percent),
shuffle=True)
train_data = RetinaDataSet(train_lst, img_read)
train_dataloader = DataLoader(train_data,
batch_size=config.TrainBatchSize,
shuffle=True,
num_workers=config.Num_workers)
val_data = RetinaDataSet(val_lst, img_read, isTrain=False)
val_dataloader = DataLoader(val_data,
batch_size=1,
shuffle=True,
num_workers=config.Num_workers)
lr = config.lr
RegCriterion = BBoxRegLoss().to(config.Device)
AncCriterion = AnchorLoss().to(config.Device)
SegCriterion = SegLoss().to(config.Device)
optimizer = optim.Adam(params=model.parameters(),
lr=lr,
weight_decay=config.Weight_decay)
# scheduler_exp = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.1)
# scheduler_cosin = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=10, eta_min=1e-5, last_epoch=-1)
scheduler_monitor = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
mode='min',
factor=0.1,
patience=3,
threshold_mode='rel',
threshold=0.0001,
cooldown=5,
min_lr=1e-5,
eps=1e-8)
previousLoss = 0
previousValLoss = 100
debug_patient = config.Debug_Patient
patient = config.Lr_Patient
for epoch in range(config.Max_epoch):
Seg_Matrix.reset()
Anc_Matrix.reset()
start_time = time.time()
Seg_precision, Seg_sensi, loss_counter, AvgAncLoss, AvgAncPrec, AvgAncSensi = 0, 0, 0, 0, 0, 0
# if epoch != 0:
# for param_group in optimizer.param_groups:
# param_group['lr'] = lr * config.lr_decay
for ii, (img, mask, anchors_labels, bbox_labels, ancLossIdx,
bboxLossIdx, segIdx, _) in enumerate(tqdm(train_dataloader)):
input = V(img).float().to(config.Device, dtype=torch.float)
mask = V(mask).float().to(config.Device, dtype=torch.float)
optimizer.zero_grad()
seg, ancOutput, boxOutput = model(input)
# ---------------------Anchor Output Statistics----------------
# vis.show_hist(name="Train Anchor Output", tensor=ancOutput)
Anc_precision, Anc_sensi, Anc_iou = 0, 0, 0
box_loss, anc_loss, seg_loss, loss = 0., 0., 0., 0.
temp_seg = 0
if seg is not None:
temp_seg = SegCriterion(seg, mask, segIdx)
if temp_seg != 0 and not t.isnan(temp_seg):
seg_loss += temp_seg
for i in range(config.Output_features):
# ----------------------Bounding Box Regression-----------------
temp_reg = RegCriterion(
boxOutput[i],
V(bbox_labels[i]).to(config.Device, dtype=torch.float),
V(bboxLossIdx[i]).to(config.Device, dtype=torch.float))
if temp_reg != 0 and not t.isnan(temp_reg):
box_loss += temp_reg / config.Output_features
# ----------------Anchor Loss Computation----------------
temp_anc = AncCriterion(
ancOutput[i],
V(anchors_labels[i]).to(config.Device, dtype=torch.float),
V(ancLossIdx[i]).to(config.Device, dtype=torch.float))
# handle = ancOutput[i].register_hook(utils.get_grad)
if temp_anc != 0 and not t.isnan(temp_anc):
anc_loss += temp_anc
Anc_Matrix.genConfusionMat(ancOutput[i], anchors_labels[i])
Anc_precision += (Anc_Matrix.precision() /
config.Output_features)
Anc_sensi += (Anc_Matrix.sensitive() / config.Output_features)
Anc_iou += (Anc_Matrix.mIoU())
if seg_loss != 0:
loss += seg_loss
if anc_loss != 0:
loss += anc_loss
if box_loss != 0:
loss += box_loss
if loss == 0:
# vis.log("Nan Loss: Seg:{}, Box:{}, Anc:{}".format(myStr(seg_loss), myStr(box_loss), myStr(anc_loss)))
continue
loss.backward()
# ---------------------Anchor Gradient Statistics----------------
# vis.show_hist(name="Train Anchor Gradient", tensor=t.tensor([k.mean() for k in utils.features['loss']]))
loss_counter += t2i(loss)
optimizer.step()
# scheduler_cosin.step()
# if anc_loss != 0:
# scheduler_monitor.step(anc_loss)
# ---------------------Train Evaluation-------------------------
if seg is not None:
Seg_Matrix.genConfusionMat(seg.clone(), mask.clone())
Seg_precision += Seg_Matrix.precision()
Seg_sensi += Seg_Matrix.sensitive()
AvgAncLoss += anc_loss
AvgAncPrec += Anc_precision
AvgAncSensi += Anc_sensi
if seg_loss != 0:
vis.log('train loss in Segmentation: ' + myStr(seg_loss))
vis.plot('Segmentation loss:', t2i(seg_loss))
if ii % config.Print_freq == config.Print_freq - 1:
vis.log('train loss in bounding box regression: ' +
myStr(box_loss))
vis.log('train loss in Anchor Net: ' + myStr(anc_loss))
vis.plot("Anchor Precision:", t2i(Anc_precision))
vis.plot("Anchor Sensitive:", t2i(Anc_sensi))
vis.plot('Bounding box regression loss', t2i(box_loss))
vis.plot('Anchor Net Loss: ', t2i(anc_loss))
vis.plot('Training Loss: ', t2i(loss))
end_time = time.time()
# -----------------Validation--------------------
val_loss = val(model, val_dataloader, vis, boxShow=False)
avg_loss = loss_counter / len(train_dataloader)
vis.plot("Avg Train Anchor Loss:",
t2i(AvgAncLoss / len(train_dataloader)))
vis.plot("Avg Train Anchor Precision:",
t2i(AvgAncPrec / len(train_dataloader)))
vis.plot("Avg Train Anchor Sensitive:",
t2i(AvgAncSensi / len(train_dataloader)))
epoch_str = (
'Epoch: {}, Train Loss: {:.5f}, Train Seg Mean Precision: {:.5f}, Train Seg Mean Sensitive:{:.5f},\
Valid Loss: {:.5f}'.format(
epoch, avg_loss, Seg_precision / len(train_dataloader),
Seg_sensi / len(train_dataloader), val_loss.item()))
print(
epoch_str + " Time:" + str(end_time - start_time) +
' lr: {}'.format(optimizer.state_dict()['param_groups'][0]['lr']))
if avg_loss >= previousLoss != 0:
debug_patient -= 1
if patient == 0:
patient = config.Lr_Patient
lr = lr * config.lr_decay
for param_group in optimizer.param_groups:
param_group['lr'] = lr
else:
patient -= 1
if debug_patient == 0:
pdb.set_trace()
debug_patient = config.Debug_Patient
previousLoss = avg_loss
if val_loss < previousValLoss / 2 and previousLoss != 100:
model.module.save()
previousValLoss = val_loss
def main():
args = get_args()
if not args.gray:
raise NotImplementedError(
'Work with color images is not implemented yet. Use --gray flag.')
dump = {}
original_image = i2t(read_image(args.original_image, args.gray))
if args.gen_distortion:
assert len(args.xp) == len(args.yp)
args.kernel_size += (1 - args.kernel_size % 2)
generated_kernel = generate_kernel(np.array(args.xp),
np.array(args.yp), args.kernel_size)
if generated_kernel is None:
print('Poor conditioned kernel.')
exit(-1)
dump['generated_kernel'] = generated_kernel
distorted_image = conv(original_image, generated_kernel)
else:
distorted_image = i2t(read_image(args.distorted_image, args.gray))
criteria = Loss(args.loss)
kernel, kernel_size, kernel_angle = initialize_kernel(t2i(distorted_image))
kernel = kernel_threshold(kernel)
print('ESTIMATED KERNEL PARAMETERS: size - {} | angle - {}'.format(
kernel_size, kernel_angle))
if args.init_with_grid:
m_kernel = MutableKernel(kernel_size, order=1)
save = True if args.dump_path is not None else False
kernel, _, _ = grid_search(original_image, distorted_image, m_kernel,
criteria)
dump['estimated_kernel'] = kernel
dump['original_image'] = t2i(original_image)
dump['distorted_image'] = t2i(distorted_image)
if args.grid_search:
m_kernel = MutableKernel(kernel_size, kernel)
save = True if args.dump_path is not None else False
found_kernel, _, all_data = grid_search(original_image,
distorted_image,
m_kernel,
criteria,
save=save)
dump['found_kernel'] = found_kernel
if save:
dump['all_data'] = all_data
else:
grad_desc(original_image, distorted_image, kernel, criteria, dump)
if 'found_kernel' not in dump or dump['found_kernel'] is None:
print('Kernel is not found.')
exit()
dump['restored_image'] = restoration.richardson_lucy(
t2i(distorted_image), dump['found_kernel'], args.lr_itr)
process_dump(args.dump_path, dump)