def Train_Dataset(self, root_dir, coco_dir, img_dir, set_dir, batch_size=8, image_size=512, use_gpu=True, num_workers=3):
self.system_dict["dataset"]["train"]["root_dir"] = root_dir;
self.system_dict["dataset"]["train"]["coco_dir"] = coco_dir;
self.system_dict["dataset"]["train"]["img_dir"] = img_dir;
self.system_dict["dataset"]["train"]["set_dir"] = set_dir;
self.system_dict["params"]["batch_size"] = batch_size;
self.system_dict["params"]["image_size"] = image_size;
self.system_dict["params"]["use_gpu"] = use_gpu;
self.system_dict["params"]["num_workers"] = num_workers;
if(self.system_dict["params"]["use_gpu"]):
if torch.cuda.is_available():
self.system_dict["local"]["num_gpus"] = torch.cuda.device_count()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
self.system_dict["local"]["training_params"] = {"batch_size": self.system_dict["params"]["batch_size"] * self.system_dict["local"]["num_gpus"],
"shuffle": True,
"drop_last": True,
"collate_fn": collater,
"num_workers": self.system_dict["params"]["num_workers"]}
self.system_dict["local"]["training_set"] = CocoDataset(root_dir=self.system_dict["dataset"]["train"]["root_dir"] + "/" + self.system_dict["dataset"]["train"]["coco_dir"],
img_dir = self.system_dict["dataset"]["train"]["img_dir"],
set_dir = self.system_dict["dataset"]["train"]["set_dir"],
transform = transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
self.system_dict["local"]["training_generator"] = DataLoader(self.system_dict["local"]["training_set"],
**self.system_dict["local"]["training_params"]);
def Val_Dataset(self, root_dir, coco_dir, img_dir, set_dir):
self.system_dict["dataset"]["val"]["status"] = True
self.system_dict["dataset"]["val"]["root_dir"] = root_dir
self.system_dict["dataset"]["val"]["coco_dir"] = coco_dir
self.system_dict["dataset"]["val"]["img_dir"] = img_dir
self.system_dict["dataset"]["val"]["set_dir"] = set_dir
self.system_dict["local"]["val_params"] = {
"batch_size": self.system_dict["params"]["batch_size"],
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": self.system_dict["params"]["num_workers"]
}
self.system_dict["local"]["val_set"] = CocoDataset(
root_dir=self.system_dict["dataset"]["val"]["root_dir"] + "/" +
self.system_dict["dataset"]["val"]["coco_dir"],
img_dir=self.system_dict["dataset"]["val"]["img_dir"],
set_dir=self.system_dict["dataset"]["val"]["set_dir"],
transform=transforms.Compose([Normalizer(),
Resizer()]))
self.system_dict["local"]["test_generator"] = DataLoader(
self.system_dict["local"]["val_set"],
**self.system_dict["local"]["val_params"])
def test(opt):
model = torch.load(opt.pretrained_model).module
model.cuda()
dataset = CocoDataset(opt.data_path,
set='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
if os.path.isdir(opt.output):
shutil.rmtree(opt.output)
os.makedirs(opt.output)
for index in range(len(dataset)):
data = dataset[index]
scale = data['scale']
image_info = dataset.coco.loadImgs(dataset.image_ids[index])[0]
with torch.no_grad():
tb = datetime.now()
scores, labels, boxes = model(data['img'].cuda().permute(
2, 0, 1).float().unsqueeze(dim=0))
boxes /= scale
te = datetime.now()
print(te, image_info['file_name'] + " cost " + str(te - tb))
if boxes.shape[0] > 0:
#image_info = dataset.coco.loadImgs(dataset.image_ids[index])[0]
#path = os.path.join(dataset.root_dir, 'images', dataset.set_name, image_info['file_name'])
path = os.path.join(dataset.root_dir, dataset.set_name,
image_info['file_name'])
print("read from ", path)
output_image = cv2.imread(path)
for box_id in range(boxes.shape[0]):
pred_prob = float(scores[box_id])
if pred_prob < opt.cls_threshold:
break
pred_label = int(labels[box_id])
xmin, ymin, xmax, ymax = boxes[box_id, :]
color = colors[pred_label]
cv2.rectangle(output_image, (xmin, ymin), (xmax, ymax), color,
2)
text_size = cv2.getTextSize(
COCO_CLASSES[pred_label] + ' : %.2f' % pred_prob,
cv2.FONT_HERSHEY_PLAIN, 1, 1)[0]
cv2.rectangle(
output_image, (xmin, ymin),
(xmin + text_size[0] + 3, ymin + text_size[1] + 4), color,
-1)
cv2.putText(output_image,
COCO_CLASSES[pred_label] + ' : %.2f' % pred_prob,
(xmin, ymin + text_size[1] + 4),
cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 255), 1)
cv2.imwrite(
"{}/{}_prediction.jpg".format(opt.output,
image_info["file_name"][:-4]),
output_image)
def test(opt):
model = SSD(backbone=ResNet())
checkpoint = torch.load(opt.pretrained_model)
model.load_state_dict(checkpoint["model_state_dict"])
if torch.cuda.is_available():
model.cuda()
model.eval()
dboxes = generate_dboxes()
test_set = CocoDataset(opt.data_path, 2017, "val",
SSDTransformer(dboxes, (300, 300), val=True))
encoder = Encoder(dboxes)
if os.path.isdir(opt.output):
shutil.rmtree(opt.output)
os.makedirs(opt.output)
for img, img_id, img_size, _, _ in test_set:
if img is None:
continue
if torch.cuda.is_available():
img = img.cuda()
with torch.no_grad():
ploc, plabel = model(img.unsqueeze(dim=0))
result = encoder.decode_batch(ploc, plabel, opt.nms_threshold,
20)[0]
loc, label, prob = [r.cpu().numpy() for r in result]
best = np.argwhere(prob > opt.cls_threshold).squeeze(axis=1)
loc = loc[best]
label = label[best]
prob = prob[best]
if len(loc) > 0:
path = test_set.coco.loadImgs(img_id)[0]["file_name"]
output_img = cv2.imread(
os.path.join(opt.data_path, "val2017", path))
height, width, _ = output_img.shape
loc[:, 0::2] *= width
loc[:, 1::2] *= height
loc = loc.astype(np.int32)
for box, lb, pr in zip(loc, label, prob):
category = test_set.label_info[lb]
color = colors[lb]
xmin, ymin, xmax, ymax = box
cv2.rectangle(output_img, (xmin, ymin), (xmax, ymax),
color, 2)
text_size = cv2.getTextSize(category + " : %.2f" % pr,
cv2.FONT_HERSHEY_PLAIN, 1,
1)[0]
cv2.rectangle(
output_img, (xmin, ymin),
(xmin + text_size[0] + 3, ymin + text_size[1] + 4),
color, -1)
cv2.putText(output_img, category + " : %.2f" % pr,
(xmin, ymin + text_size[1] + 4),
cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 255), 1)
cv2.imwrite(
"{}/{}_prediction.jpg".format(opt.output, path[:-4]),
output_img)
def train(opt):
num_gpus = 1
if torch.cuda.is_available():
num_gpus = torch.cuda.device_count()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
training_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": True,
"drop_last": True,
"collate_fn": collater,
"num_workers": 12
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": 12
}
training_set = CocoDataset(root_dir=opt.data_path,
set="train2017",
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
training_generator = DataLoader(training_set, **training_params)
test_set = CocoDataset(root_dir=opt.data_path,
set="val2017",
transform=transforms.Compose(
[Normalizer(), Resizer()]))
test_generator = DataLoader(test_set, **test_params)
model = EfficientDet(num_classes=training_set.num_classes())
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
writer = SummaryWriter(opt.log_path)
if torch.cuda.is_available():
model = model.cuda()
model = nn.DataParallel(model)
optimizer = torch.optim.Adam(model.parameters(), opt.lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
best_loss = 1e5
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epochs):
model.train()
# if torch.cuda.is_available():
# model.module.freeze_bn()
# else:
# model.freeze_bn()
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
try:
optimizer.zero_grad()
if torch.cuda.is_available():
cls_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0:
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
total_loss = np.mean(epoch_loss)
progress_bar.set_description(
'Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Batch loss: {:.5f} Total loss: {:.5f}'
.format(epoch + 1, opt.num_epochs, iter + 1,
num_iter_per_epoch, cls_loss, reg_loss, loss,
total_loss))
writer.add_scalar('Train/Total_loss', total_loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Regression_loss', reg_loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Classfication_loss (focal loss)',
cls_loss, epoch * num_iter_per_epoch + iter)
except Exception as e:
print(e)
continue
scheduler.step(np.mean(epoch_loss))
if epoch % opt.test_interval == 0:
model.eval()
loss_regression_ls = []
loss_classification_ls = []
for iter, data in enumerate(test_generator):
with torch.no_grad():
if torch.cuda.is_available():
cls_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss_classification_ls.append(float(cls_loss))
loss_regression_ls.append(float(reg_loss))
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
print(
'Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}. Total loss: {:1.5f}'
.format(epoch + 1, opt.num_epochs, cls_loss, reg_loss,
np.mean(loss)))
writer.add_scalar('Test/Total_loss', loss, epoch)
writer.add_scalar('Test/Regression_loss', reg_loss, epoch)
writer.add_scalar('Test/Classfication_loss (focal loss)', cls_loss,
epoch)
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
torch.save(
model,
os.path.join(opt.saved_path,
"signatrix_efficientdet_coco.pth"))
dummy_input = torch.rand(opt.batch_size, 3, 512, 512)
if torch.cuda.is_available():
dummy_input = dummy_input.cuda()
if isinstance(model, nn.DataParallel):
model.module.backbone_net.model.set_swish(
memory_efficient=False)
torch.onnx.export(model.module,
dummy_input,
os.path.join(
opt.saved_path,
"signatrix_efficientdet_coco.onnx"),
verbose=False,
opset_version=11)
model.module.backbone_net.model.set_swish(
memory_efficient=True)
else:
model.backbone_net.model.set_swish(memory_efficient=False)
torch.onnx.export(model,
dummy_input,
os.path.join(
opt.saved_path,
"signatrix_efficientdet_coco.onnx"),
verbose=False,
opset_version=11)
model.backbone_net.model.set_swish(memory_efficient=True)
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, loss))
break
writer.close()
def test(opt):
test_set = CocoDataset(opt.data_path,
set='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
opt.num_classes = test_set.num_classes()
opt.batch_size = opt.batch_size * 4
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": 12
}
test_generator = DataLoader(test_set, **test_params)
model = EfficientDet(opt)
model.load_state_dict(
torch.load(os.path.join(opt.pretrained_model, opt.network + '.pth')))
model.cuda()
model.set_is_training(False)
model.eval()
if os.path.isdir(opt.prediction_dir):
shutil.rmtree(opt.prediction_dir)
os.makedirs(opt.prediction_dir)
progress_bar = tqdm(test_generator)
progress_bar.set_description_str(' Evaluating')
IoU_scores = []
for i, data in enumerate(progress_bar):
scale = data['scale']
with torch.no_grad():
output_list = model(data['img'].cuda().float())
for j, output in enumerate(output_list):
scores, labels, boxes = output
annot = data['annot'][j]
annot = annot[annot[:, 4] != -1]
# print(scores.size(), labels.size(), boxes.size(), annot.size())
if boxes.shape[0] == 0:
if annot.size(0) == 0:
IoU_scores.append(1.0)
else:
IoU_scores.append(0.0)
continue
if annot.size(0) == 0:
IoU_scores.append(0.0)
else:
classes = set(annot[:, 4].tolist())
cat = torch.cat(
[scores.view(-1, 1),
labels.view(-1, 1).float(), boxes],
dim=1)
cat = cat[cat[:, 0] >= opt.cls_threshold]
iou_score = []
for c in classes:
box = cat[cat[:, 1] == c][:, 2:]
if box.size(0) == 0:
iou_score.append(0.0)
continue
tgt = annot[annot[:, 4] == c][:, :4]
iou_s = iou(box, tgt.cuda())
iou_score.append(iou_s.cpu().numpy())
classes_pre = set(cat[:, 1].tolist())
for c in classes_pre:
if c not in classes:
iou_score.append(0)
# print(classes_pre, classes ,iou_score)
IoU_scores.append(sum(iou_score) / len(iou_score))
if writePIC:
annot /= scale[j]
boxes /= scale[j]
image_info = test_set.coco.loadImgs(
test_set.image_ids[i * opt.batch_size + j])[0]
# print(image_info['file_name'])
path = os.path.join(test_set.root_dir, 'images',
test_set.set_name, image_info['file_name'])
output_image = cv2.imread(path)
# print(output_image.shape)
for box_id in range(boxes.shape[0]):
pred_prob = float(scores[box_id])
if pred_prob < opt.cls_threshold:
break
pred_label = int(labels[box_id])
xmin, ymin, xmax, ymax = boxes[box_id, :]
color = colors[pred_label]
cv2.rectangle(output_image, (xmin, ymin), (xmax, ymax),
color, 1)
text_size = cv2.getTextSize(
COCO_CLASSES[pred_label] + ' : %.2f' % pred_prob,
cv2.FONT_HERSHEY_PLAIN, 1, 1)[0]
cv2.rectangle(
output_image, (xmin, ymin),
(xmin + text_size[0] + 3, ymin + text_size[1] + 4),
color, -1)
cv2.putText(
output_image,
COCO_CLASSES[pred_label] + ' : %.2f' % pred_prob,
(xmin, ymin + text_size[1] + 4),
cv2.FONT_HERSHEY_PLAIN, 1, (255, 255, 255), 1)
for box_id in range(annot.size(0)):
xmin, ymin, xmax, ymax = annot[box_id, :4]
cv2.rectangle(output_image, (xmin, ymin), (xmax, ymax),
(255, 0, 0), 1)
cv2.imwrite(
"{}/{}_prediction.jpg".format(
opt.prediction_dir, image_info["file_name"][:-4]),
output_image)
print(sum(IoU_scores) / len(IoU_scores))
def main(opt):
if torch.cuda.is_available():
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
num_gpus = torch.distributed.get_world_size()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
num_gpus = 1
train_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": True,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
test_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": False,
"drop_last": False,
"num_workers": opt.num_workers,
"collate_fn": collate_fn
}
if opt.model == "ssd":
dboxes = generate_dboxes(model="ssd")
model = SSD(backbone=ResNet(), num_classes=len(coco_classes))
else:
dboxes = generate_dboxes(model="ssdlite")
model = SSDLite(backbone=MobileNetV2(), num_classes=len(coco_classes))
train_set = CocoDataset(opt.data_path, 2017, "train",
SSDTransformer(dboxes, (300, 300), val=False))
train_loader = DataLoader(train_set, **train_params)
test_set = CocoDataset(opt.data_path, 2017, "val",
SSDTransformer(dboxes, (300, 300), val=True))
test_loader = DataLoader(test_set, **test_params)
encoder = Encoder(dboxes)
opt.lr = opt.lr * num_gpus * (opt.batch_size / 32)
criterion = Loss(dboxes)
optimizer = torch.optim.SGD(model.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay,
nesterov=True)
scheduler = MultiStepLR(optimizer=optimizer,
milestones=opt.multistep,
gamma=0.1)
if torch.cuda.is_available():
model.cuda()
criterion.cuda()
if opt.amp:
from apex import amp
from apex.parallel import DistributedDataParallel as DDP
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
else:
from torch.nn.parallel import DistributedDataParallel as DDP
# It is recommended to use DistributedDataParallel, instead of DataParallel
# to do multi-GPU training, even if there is only a single node.
model = DDP(model)
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.save_folder):
os.makedirs(opt.save_folder)
checkpoint_path = os.path.join(opt.save_folder, "SSD.pth")
writer = SummaryWriter(opt.log_path)
if os.path.isfile(checkpoint_path):
checkpoint = torch.load(checkpoint_path)
first_epoch = checkpoint["epoch"] + 1
model.module.load_state_dict(checkpoint["model_state_dict"])
scheduler.load_state_dict(checkpoint["scheduler"])
optimizer.load_state_dict(checkpoint["optimizer"])
else:
first_epoch = 0
for epoch in range(first_epoch, opt.epochs):
train(model, train_loader, epoch, writer, criterion, optimizer,
scheduler, opt.amp)
evaluate(model, test_loader, epoch, writer, encoder, opt.nms_threshold)
checkpoint = {
"epoch": epoch,
"model_state_dict": model.module.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict()
}
torch.save(checkpoint, checkpoint_path)
def train(opt):
num_gpus = 1
if torch.cuda.is_available():
num_gpus = torch.cuda.device_count()
else:
raise Exception('no GPU')
cudnn.benchmark = True
training_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": True,
"drop_last": True,
"collate_fn": collater,
"num_workers": 12
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": 12
}
training_set = CocoDataset(root_dir=opt.data_path,
set="train2017",
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
training_generator = DataLoader(training_set, **training_params)
test_set = CocoDataset(root_dir=opt.data_path,
set="val2017",
transform=transforms.Compose(
[Normalizer(), Resizer()]))
test_generator = DataLoader(test_set, **test_params)
opt.num_classes = training_set.num_classes()
model = EfficientDet(opt)
if opt.resume:
print('Loading model...')
model.load_state_dict(
torch.load(os.path.join(opt.saved_path, opt.network + '.pth')))
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
model = model.cuda()
model = nn.DataParallel(model)
optimizer = torch.optim.AdamW(model.parameters(), opt.lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
best_loss = 1e5
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
for epoch in range(opt.num_epochs):
print('Epoch: {}/{}:'.format(epoch + 1, opt.num_epochs))
model.train()
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
try:
optimizer.zero_grad()
if torch.cuda.is_available():
cls_loss, cls_2_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, cls_2_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
cls_2_loss = cls_2_loss.mean()
loss = cls_loss + cls_2_loss + reg_loss
if loss == 0:
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
total_loss = np.mean(epoch_loss)
progress_bar.set_description(
'Epoch: {}/{}. Iteration: {}/{}'.format(
epoch + 1, opt.num_epochs, iter + 1,
num_iter_per_epoch))
progress_bar.write(
'Cls loss: {:.5f}\tReg loss: {:.5f}\tCls+Reg loss: {:.5f}\tBatch loss: {:.5f}\tTotal loss: {:.5f}'
.format(cls_loss, reg_loss, cls_loss + reg_loss, loss,
total_loss))
except Exception as e:
print(e)
continue
scheduler.step(np.mean(epoch_loss))
if epoch % opt.test_interval == 0:
model.eval()
loss_regression_ls = []
loss_classification_ls = []
loss_classification_2_ls = []
progress_bar = tqdm(test_generator)
progress_bar.set_description_str(' Evaluating')
for iter, data in enumerate(progress_bar):
with torch.no_grad():
if torch.cuda.is_available():
cls_loss, cls_2_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, cls_2_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
cls_2_loss = cls_2_loss.mean()
reg_loss = reg_loss.mean()
loss_classification_ls.append(float(cls_loss))
loss_classification_2_ls.append(float(cls_2_loss))
loss_regression_ls.append(float(reg_loss))
cls_loss = np.mean(loss_classification_ls)
cls_2_loss = np.mean(loss_classification_2_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + cls_2_loss + reg_loss
print(
'Epoch: {}/{}. \nClassification loss: {:1.5f}. \tClassification_2 loss: {:1.5f}. \tRegression loss: {:1.5f}. \tTotal loss: {:1.5f}'
.format(epoch + 1, opt.num_epochs, cls_loss, cls_2_loss,
reg_loss, np.mean(loss)))
if loss + opt.es_min_delta < best_loss:
print('Saving model...')
best_loss = loss
best_epoch = epoch
torch.save(model.module.state_dict(),
os.path.join(opt.saved_path, opt.network + '.pth'))
# torch.save(model, os.path.join(opt.saved_path, opt.network+'.pth'))
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, loss))
break
if not len(results):
return
# write output
json.dump(results,
open('{}_bbox_results.json'.format(dataset.set_name), 'w'),
indent=4)
# load results in COCO evaluation tool
coco_true = dataset.coco
coco_pred = coco_true.loadRes('{}_bbox_results.json'.format(
dataset.set_name))
# run COCO evaluation
coco_eval = COCOeval(coco_true, coco_pred, 'bbox')
coco_eval.params.imgIds = image_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
efficientdet = torch.load(
"trained_models/signatrix_efficientdet_coco.pth").module
efficientdet.cuda()
dataset_val = CocoDataset("data/COCO",
set='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
evaluate_coco(dataset_val, efficientdet)
def Train_Dataset(self, root_dir, coco_dir, img_dir, set_dir, batch_size=8, image_size=512, use_gpu=True, num_workers=3):
'''
User function: Set training dataset parameters
Dataset Directory Structure
root_dir
|
|------coco_dir
| |
| |----img_dir
| |
| |------<set_dir_train> (set_dir) (Train)
| |
| |---------img1.jpg
| |---------img2.jpg
| |---------..........(and so on)
|
|
| |---annotations
| |----|
| |--------------------instances_Train.json (instances_<set_dir_train>.json)
| |--------------------classes.txt
- instances_Train.json -> In proper COCO format
- classes.txt -> A list of classes in alphabetical order
For TrainSet
- root_dir = "../sample_dataset";
- coco_dir = "kangaroo";
- img_dir = "images";
- set_dir = "Train";
Note: Annotation file name too coincides against the set_dir
Args:
root_dir (str): Path to root directory containing coco_dir
coco_dir (str): Name of coco_dir containing image folder and annotation folder
img_dir (str): Name of folder containing all training and validation folders
set_dir (str): Name of folder containing all training images
batch_size (int): Mini batch sampling size for training epochs
image_size (int): Either of [512, 300]
use_gpu (bool): If True use GPU else run on CPU
num_workers (int): Number of parallel processors for data loader
Returns:
None
'''
self.system_dict["dataset"]["train"]["root_dir"] = root_dir;
self.system_dict["dataset"]["train"]["coco_dir"] = coco_dir;
self.system_dict["dataset"]["train"]["img_dir"] = img_dir;
self.system_dict["dataset"]["train"]["set_dir"] = set_dir;
self.system_dict["params"]["batch_size"] = batch_size;
self.system_dict["params"]["image_size"] = image_size;
self.system_dict["params"]["use_gpu"] = use_gpu;
self.system_dict["params"]["num_workers"] = num_workers;
if(self.system_dict["params"]["use_gpu"]):
if torch.cuda.is_available():
self.system_dict["local"]["num_gpus"] = torch.cuda.device_count()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
self.system_dict["local"]["training_params"] = {"batch_size": self.system_dict["params"]["batch_size"] * self.system_dict["local"]["num_gpus"],
"shuffle": True,
"drop_last": True,
"collate_fn": collater,
"num_workers": self.system_dict["params"]["num_workers"]}
self.system_dict["local"]["training_set"] = CocoDataset(root_dir=self.system_dict["dataset"]["train"]["root_dir"] + "/" + self.system_dict["dataset"]["train"]["coco_dir"],
img_dir = self.system_dict["dataset"]["train"]["img_dir"],
set_dir = self.system_dict["dataset"]["train"]["set_dir"],
transform = transforms.Compose([Normalizer(), Resizer()])) #Augmenter(),
self.system_dict["local"]["training_generator"] = DataLoader(self.system_dict["local"]["training_set"],
**self.system_dict["local"]["training_params"]);
def Val_Dataset(self, root_dir, coco_dir, img_dir, set_dir):
'''
User function: Set training dataset parameters
Dataset Directory Structure
root_dir
|
|------coco_dir
| |
| |----img_dir
| |
| |------<set_dir_val> (set_dir) (Validation)
| |
| |---------img1.jpg
| |---------img2.jpg
| |---------..........(and so on)
|
|
| |---annotations
| |----|
| |--------------------instances_Val.json (instances_<set_dir_val>.json)
| |--------------------classes.txt
- instances_Train.json -> In proper COCO format
- classes.txt -> A list of classes in alphabetical order
For ValSet
- root_dir = "..sample_dataset";
- coco_dir = "kangaroo";
- img_dir = "images";
- set_dir = "Val";
Note: Annotation file name too coincides against the set_dir
Args:
root_dir (str): Path to root directory containing coco_dir
coco_dir (str): Name of coco_dir containing image folder and annotation folder
img_dir (str): Name of folder containing all training and validation folders
set_dir (str): Name of folder containing all validation images
Returns:
None
'''
self.system_dict["dataset"]["val"]["status"] = True;
self.system_dict["dataset"]["val"]["root_dir"] = root_dir;
self.system_dict["dataset"]["val"]["coco_dir"] = coco_dir;
self.system_dict["dataset"]["val"]["img_dir"] = img_dir;
self.system_dict["dataset"]["val"]["set_dir"] = set_dir;
self.system_dict["local"]["val_params"] = {"batch_size": self.system_dict["params"]["batch_size"],
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": self.system_dict["params"]["num_workers"]}
self.system_dict["local"]["val_set"] = CocoDataset(root_dir=self.system_dict["dataset"]["val"]["root_dir"] + "/" + self.system_dict["dataset"]["val"]["coco_dir"],
img_dir = self.system_dict["dataset"]["val"]["img_dir"],
set_dir = self.system_dict["dataset"]["val"]["set_dir"],
transform=transforms.Compose([Normalizer(), Resizer()]))
self.system_dict["local"]["test_generator"] = DataLoader(self.system_dict["local"]["val_set"],
**self.system_dict["local"]["val_params"])
if not len(results):
return
# write output
json.dump(results,
open('{}_bbox_results.json'.format(dataset.set_name), 'w'),
indent=4)
# load results in COCO evaluation tool
coco_true = dataset.coco
coco_pred = coco_true.loadRes('{}_bbox_results.json'.format(
dataset.set_name))
# run COCO evaluation
coco_eval = COCOeval(coco_true, coco_pred, 'bbox')
coco_eval.params.imgIds = image_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
efficientdet = torch.load(
"trained_models/signatrix_efficientdet_coco.pth").module
efficientdet.cuda()
dataset_val = CocoDataset("/disk4t/data/coco/data/coco",
set='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
evaluate_coco(dataset_val, efficientdet)
return
# write output
json.dump(results,
open('{}_bbox_results.json'.format(dataset.set_name), 'w'),
indent=4)
# load results in COCO evaluation tool
coco_true = dataset.coco
coco_pred = coco_true.loadRes('{}_bbox_results.json'.format(
dataset.set_name))
# run COCO evaluation
coco_eval = COCOeval(coco_true, coco_pred, 'bbox')
coco_eval.params.imgIds = image_ids
coco_eval.evaluate()
coco_eval.accumulate()
coco_eval.summarize()
if __name__ == '__main__':
#load model
efficientdet = torch.load(args.model).module
efficientdet.cuda()
dataset_val = CocoDataset(args.dataset,
set='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
evaluate_coco(dataset_val, efficientdet)
def train(opt):
num_gpus = 1
if torch.cuda.is_available():
num_gpus = torch.cuda.device_count()
torch.cuda.manual_seed(123)
else:
torch.manual_seed(123)
training_params = {
"batch_size": opt.batch_size * num_gpus,
"shuffle": True,
"drop_last": True,
"collate_fn": collater,
"num_workers": 12
}
test_params = {
"batch_size": opt.batch_size,
"shuffle": False,
"drop_last": False,
"collate_fn": collater,
"num_workers": 12
}
training_set = CocoDataset(root_dir=opt.data_path,
set="train2017",
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
training_generator = DataLoader(training_set, **training_params)
test_set = CocoDataset(root_dir=opt.data_path,
set="val2017",
transform=transforms.Compose(
[Normalizer(), Resizer()]))
test_generator = DataLoader(test_set, **test_params)
channels_map = {
'efficientnet-b0': [40, 80, 192],
'efficientnet-b1': [40, 80, 192],
'efficientnet-b2': [48, 88, 208],
'efficientnet-b3': [48, 96, 232],
'efficientnet-b4': [56, 112, 272],
'efficientnet-b5': [64, 128, 304],
'efficientnet-b6': [72, 144, 344],
'efficientnet-b7': [80, 160, 384],
'efficientnet-b8': [80, 160, 384]
}
if os.path.isdir(opt.log_path):
shutil.rmtree(opt.log_path)
os.makedirs(opt.log_path)
if not os.path.isdir(opt.saved_path):
os.makedirs(opt.saved_path)
writer = SummaryWriter(opt.log_path)
if opt.resume:
resume_path = os.path.join(opt.saved_path,
'signatrix_efficientdet_coco_latest.pth')
model = torch.load(resume_path).module
print("model loaded from {}".format(resume_path))
else:
model = EfficientDet(
num_classes=training_set.num_classes(),
network=opt.backbone_network,
remote_loading=opt.remote_loading,
advprop=opt.advprop,
conv_in_channels=channels_map[opt.backbone_network])
print("model created with backbone {}, advprop {}".format(
opt.backbone_network, opt.advprop))
if torch.cuda.is_available():
model = model.cuda()
model = nn.DataParallel(model)
if opt.resume:
m = round(opt.start_epoch / 100)
opt.lr = opt.lr * (0.1**m)
optimizer = torch.optim.Adam(model.parameters(), opt.lr)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
best_loss = 1e5
best_epoch = 0
model.train()
num_iter_per_epoch = len(training_generator)
start_epoch = 0
if opt.resume:
start_epoch = opt.start_epoch
for epoch in range(start_epoch, opt.num_epochs):
model.train()
# if torch.cuda.is_available():
# model.module.freeze_bn()
# else:
# model.freeze_bn()
epoch_loss = []
progress_bar = tqdm(training_generator)
for iter, data in enumerate(progress_bar):
try:
optimizer.zero_grad()
if torch.cuda.is_available():
cls_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss = cls_loss + reg_loss
if loss == 0:
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
total_loss = np.mean(epoch_loss)
progress_bar.set_description(
'{} Epoch: {}/{}. Iteration: {}/{}. Cls loss: {:.5f}. Reg loss: {:.5f}. Batch loss: {:.5f} Total loss: {:.5f}'
.format(datetime.now(), epoch + 1, opt.num_epochs,
iter + 1, num_iter_per_epoch, cls_loss, reg_loss,
loss, total_loss))
writer.add_scalar('Train/Total_loss', total_loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Regression_loss', reg_loss,
epoch * num_iter_per_epoch + iter)
writer.add_scalar('Train/Classfication_loss (focal loss)',
cls_loss, epoch * num_iter_per_epoch + iter)
except Exception as e:
print(e)
continue
scheduler.step(np.mean(epoch_loss))
if epoch % opt.test_interval == 0:
model.eval()
loss_regression_ls = []
loss_classification_ls = []
for iter, data in enumerate(test_generator):
with torch.no_grad():
if torch.cuda.is_available():
cls_loss, reg_loss = model(
[data['img'].cuda().float(), data['annot'].cuda()])
else:
cls_loss, reg_loss = model(
[data['img'].float(), data['annot']])
cls_loss = cls_loss.mean()
reg_loss = reg_loss.mean()
loss_classification_ls.append(float(cls_loss))
loss_regression_ls.append(float(reg_loss))
cls_loss = np.mean(loss_classification_ls)
reg_loss = np.mean(loss_regression_ls)
loss = cls_loss + reg_loss
print(
'{} Epoch: {}/{}. Classification loss: {:1.5f}. Regression loss: {:1.5f}. Total loss: {:1.5f}'
.format(datetime.now(), epoch + 1, opt.num_epochs, cls_loss,
reg_loss, np.mean(loss)))
writer.add_scalar('Test/Total_loss', loss, epoch)
writer.add_scalar('Test/Regression_loss', reg_loss, epoch)
writer.add_scalar('Test/Classfication_loss (focal loss)', cls_loss,
epoch)
if loss + opt.es_min_delta < best_loss:
best_loss = loss
best_epoch = epoch
torch.save(
model,
os.path.join(
opt.saved_path,
"signatrix_efficientdet_coco_best_epoch{}.pth".format(
epoch)))
'''
dummy_input = torch.rand(opt.batch_size, 3, 512, 512)
if torch.cuda.is_available():
dummy_input = dummy_input.cuda()
if isinstance(model, nn.DataParallel):
model.module.backbone_net.model.set_swish(memory_efficient=False)
torch.onnx.export(model.module, dummy_input,
os.path.join(opt.saved_path, "signatrix_efficientdet_coco.onnx"),
verbose=False)
model.module.backbone_net.model.set_swish(memory_efficient=True)
else:
model.backbone_net.model.set_swish(memory_efficient=False)
torch.onnx.export(model, dummy_input,
os.path.join(opt.saved_path, "signatrix_efficientdet_coco.onnx"),
verbose=False)
model.backbone_net.model.set_swish(memory_efficient=True)
'''
print("epoch:", epoch, "best_epoch:", best_epoch,
"epoch - best_epoch=", epoch - best_epoch)
# Early stopping
if epoch - best_epoch > opt.es_patience > 0:
print(
"Stop training at epoch {}. The lowest loss achieved is {}"
.format(epoch, loss))
break
if epoch % opt.save_interval == 0:
torch.save(
model,
os.path.join(opt.saved_path,
"signatrix_efficientdet_coco_latest.pth"))
writer.close()