def test():
model = Darknet()
print('Size of the test set:{}'.format(len(testset)))
# load ckpt
ckpt = torch.load(os.path.join(EXPT_DIR, CKPT_FILE))
model.load_state_dict(ckpt['model'])
fw = Flow(model, trainset, testset, hyp)
# write prediction
fw.validate(batch_size=16)
def infer():
model = Darknet()
ckpt = torch.load(os.path.join(EXPT_DIR, CKPT_FILE))
model.load_state_dict(ckpt['model'])
fw = Flow(model, trainset, testset, hyp)
# get the indices of unlabeled data
with open(os.path.join(EXPT_DIR, 'unlabeled.pkl'), 'rb') as f:
dt = pickle.load(f)
unlabeled = [i for i in dt]
fw.infer(unlabeled)
return
def train():
# Evaluation pipeline
files = glob.glob(
os.path.join('/home/iotsc_group1/ChangxingDENG/det/datasets/',
'PretrainImageNet', 'ILSVRC2012_img_val', '*.JPEG'))
files = sorted(files,
key=lambda f: f.split('/')[-1].split('_')[-1].split('.')[0])
labels = loadlabel(
os.path.join(
'/home/iotsc_group1/ChangxingDENG/det/datasets/',
'PretrainImageNet',
'ILSVRC2012_devkit_t12/data/ILSVRC2012_validation_ground_truth.txt'
))
eval_pipeline = EvalImageDecoderPipeline(files=files, labels=labels)
eval_pipeline.build()
eval_pii = pytorchIterator(eval_pipeline,
last_batch_policy=LastBatchPolicy.PARTIAL,
reader_name='Reader',
auto_reset=True)
model = Darknet()
state_dict = torch.load(
'logs/PretrainImageNet_20210316173822_1/13_70056.pth')
model.load_state_dict(state_dict=state_dict)
model = model.cuda()
criterion = nn.CrossEntropyLoss()
model.eval()
epoch_loss = 0
prediciton = []
target = []
with torch.no_grad():
for iter, data in enumerate(eval_pii):
x = data[0]['data']
label = data[0]['label'].squeeze(-1).long().cuda()
output = model(x)
loss = criterion(output, label).item()
epoch_loss += loss * x.shape[0]
prediciton.append(output)
target.append(label)
loss = epoch_loss / 50000
prediciton = torch.cat(prediciton, dim=0)
target = torch.cat(target, dim=0)
acc = top1accuracy(prediciton, target)
acctop5 = top5accuracy(prediciton, target)
print(f'Top1 ACC: {acc} Top5 ACC {acctop5} loss: {loss}')
def main():
# Initiate model
eval_model = Darknet("config/net/resnet_dropout.cfg").to('cuda')
eval_model.load_state_dict(torch.load("./logs/model/model_params_99.ckpt"))
precision, recall, AP, f1, ap_class = evaluate(eval_model, ['2012', 'val'],
[0.5, 0.5, 0.5],
4,
True,
diagnosis_code=1)
evaluation_metrics = [
("val_precision", precision.mean()),
("val_recall", recall.mean()),
("val_mAP", AP.mean()),
("val_f1", f1.mean()),
]
for tag, value in evaluation_metrics:
print("{}: {}".format(tag, value.item()))
def train():
epochs = 1
model = Darknet()
if RESUME_FROM:
ckpt = torch.load(os.path.join(EXPT_DIR, RESUME_FROM))
model.load_state_dict(ckpt['model'])
trainset.augment = True
fw = Flow(model, trainset, testset, hyp)
with open(os.path.join(EXPT_DIR, 'labeled.pkl'), 'rb') as f:
selected = pickle.load(f)
# if want to try pollution study
# load labels from EXPT_DIR/labeled.pkl
for epoch in range(epochs):
fw.train(epoch, samples=list(selected.keys()), prebias=False)
return
def train(params):
params = Params(params)
set_random_seeds(params.seed)
time_now = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
params.save_root = params.save_root + f'/{params.project_name}_{time_now}_{params.version}'
os.makedirs(params.save_root, exist_ok=True)
logging.basicConfig(
filename=
f'{params.save_root}/{params.project_name}_{time_now}_{params.version}.log',
filemode='a',
format='%{asctime}s - %(levalname)s: %(message)s')
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
logging.info(f'Available GPUs: {torch.cuda.device_count()}')
# Train pipeline
files = glob.glob(
os.path.join(params.data_root, params.project_name, params.train_set,
'*/*.JPEG'))
labels = []
for fp in files:
label = int(fp.split('/')[-2]) - 1
labels.append(label)
assert len(files) == len(labels)
train_pipeline = TrainImageDecoderPipeline(params=params,
device_id=0,
files=files,
labels=labels)
train_pipeline.build()
train_pii = pytorchIterator(train_pipeline,
last_batch_policy=LastBatchPolicy.DROP,
reader_name='Reader',
auto_reset=True)
# Evaluation pipeline
files = glob.glob(
os.path.join(params.data_root, params.project_name, params.val_set,
'*.JPEG'))
files = sorted(files,
key=lambda f: f.split('/')[-1].split('_')[-1].split('.')[0])
labels = loadlabel(
os.path.join(
params.data_root, params.project_name,
'ILSVRC2012_devkit_t12/data/ILSVRC2012_validation_ground_truth.txt'
))
eval_pipeline = EvalImageDecoderPipeline(params=params,
device_id=0,
files=files,
labels=labels)
eval_pipeline.build()
eval_pii = pytorchIterator(eval_pipeline,
last_batch_policy=LastBatchPolicy.PARTIAL,
reader_name='Reader',
auto_reset=True)
model = Darknet()
last_step = 0
last_epoch = 0
if params.load_weights != 'None':
try:
state_dict = torch.load(params.load_weights)
model.load_state_dict(state_dict)
last_step = int(params.load_weights.split('_')[-1].split('.')[0])
last_epoch = int(params.load_weights.split('_')[-2])
except:
logging.error('Fail to resuming from weight!')
exit()
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = nn.DataParallel(model)
if params.optim == 'Adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=params.learning_rate)
else:
optimizer = torch.optim.SGD(model.parameters(),
lr=params.learning_rate,
momentum=0.9,
nesterov=True)
criterion = nn.CrossEntropyLoss()
# ls_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer=optimizer, factor=0.5, verbose=True, patience=8)
epoch = 0
begin_epoch = max(0, last_epoch)
step = max(0, last_step)
best_loss = 100
logging.info('Begin to train...')
model.train()
try:
for epoch in range(begin_epoch, params.epoch):
for iter, data in enumerate(train_pii):
x = data[0]['data']
label = data[0]['label'].squeeze(-1).long().cuda()
output = model(x)
loss = criterion(output, label)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if iter % params.save_interval == 0:
logging.info(
f'{datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")} '
f'Train Epoch: {epoch} iter: {iter} loss: {loss.item()}'
)
step += 1
if epoch % params.eval_interval == 0:
model.eval()
epoch_loss = 0
prediciton = []
target = []
with torch.no_grad():
for iter, data in enumerate(eval_pii):
x = data[0]['data']
label = data[0]['label'].squeeze(-1).long().cuda()
output = model(x)
loss = criterion(output, label).item()
epoch_loss += loss * x.shape[0]
prediciton.append(output)
target.append(label)
loss = epoch_loss / 50000
prediciton = torch.cat(prediciton, dim=0)
target = torch.cat(target, dim=0)
acc = top1accuracy(prediciton, target)
acctop5 = top5accuracy(prediciton, target)
logging.info(
f'{datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")} '
f'Eval Epoch: {epoch} loss: {loss} accuracy: {acc} Top5 acc: {acctop5}'
)
if loss < best_loss:
best_loss = loss
save_checkpoint(
model, f'{params.save_root}/{epoch}_{step}.pth')
model.train()
except KeyboardInterrupt:
save_checkpoint(model,
f'{params.save_root}/Interrupt_{epoch}_{step}.pth')
os.makedirs("checkpoints", exist_ok=True)
# Get data configuration
data_config = parse_data_config(opt.data_config)
train_path = data_config["train"]
valid_path = data_config["valid"]
class_names = load_classes(data_config["names"])
# Initiate model
model = Darknet(opt.model_def).to(device)
model.apply(weights_init_normal)
# If specified we start from checkpoint
if opt.pretrained_weights:
if opt.pretrained_weights.endswith(".pth"):
model.load_state_dict(torch.load(opt.pretrained_weights))
else:
model.load_darknet_weights(opt.pretrained_weights)
# Get dataloader
dataset = ListDataset(train_path,
augment=True,
multiscale=opt.multiscale_training)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_cpu,
pin_memory=True,
collate_fn=dataset.collate_fn,
)
def train(params):
params = Params(params)
set_random_seeds(params.seed)
time_now = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
params.save_root = params.save_root + f'/{params.project_name}_{time_now}_{params.version}'
os.makedirs(params.save_root, exist_ok=True)
logging.basicConfig(
filename=
f'{params.save_root}/{params.project_name}_{time_now}_{params.version}.log',
filemode='a',
format='%{asctime}s - %(levalname)s: %(message)s')
if params.num_gpus == 0:
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
logging.info(f'Available GPUs: {torch.cuda.device_count()}')
data_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=params.mean, std=params.std)
])
train_set = TrainDataset(root=os.path.join(params.data_root,
params.project_name,
params.train_set),
transform=data_transform)
# val_set = EvalDataset(root=os.path.join(params.data_root, params.project_name, params.val_set),
# label_path=os.path.join(params.data_root, params.project_name,
# 'ILSVRC2012_devkit_t12/data/ILSVRC2012_validation_ground_truth.txt'),
# transform=data_transform)
train_params = {
'batch_size': params.batch_size,
'shuffle': False,
'num_workers': 4,
'drop_last': True
}
# val_params = {'batch_size': params.batch_size, 'shuffle': False, 'num_workers': params.num_gpus * 4,
# 'drop_last': False}
train_loader = DataLoader(train_set, **train_params)
# val_loader = DataLoader(val_set, **val_params)
eli = ExternalInputIterator(params.batch_size)
params.mean = torch.Tensor(params.mean).unsqueeze(0).unsqueeze(0)
params.std = torch.Tensor(params.std).unsqueeze(0).unsqueeze(0)
pipe = ExternalSourcePipeline(params=params,
num_threads=4,
device_id=0,
external_date=eli,
seed=params.seed)
# pipe.build()
# images, _ = pipe.run()
# print(np.array(images[0].as_cpu()).shape)
import matplotlib.pyplot as plt
# plt.imsave('image[0].jpg', np.array(images[0].as_cpu()).transpose((1, 2, 0)))
# plt.imsave('image[0].jpg', images[0].as_cpu())
# exit()
pii = pytorchIterator(pipe,
last_batch_padded=True,
last_batch_policy=LastBatchPolicy.DROP)
model = Darknet()
criterion = nn.CrossEntropyLoss()
last_step = 0
last_epoch = 0
if params.load_weights != 'None':
try:
state_dict = torch.load(params.load_weights)
model.load_state_dict(state_dict)
last_step = int(params.load_weights.split('_')[-1].split('.')[0])
last_epoch = int(params.load_weights.split('_')[-2])
except:
logging.error('Fail to resuming from weight!')
exit()
if params.num_gpus > 0:
model = model.cuda()
if params.num_gpus > 1:
model = nn.DataParallel(model)
if params.optim == 'Adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=params.learning_rate)
else:
optimizer = torch.optim.SGD(model.parameters(),
lr=params.learning_rate,
momentum=0.9,
nesterov=True)
epoch = 0
begin_epoch = max(0, last_epoch)
step = max(0, last_step)
logging.info('Begin to train...')
model.train()
try:
import time
for epoch in range(begin_epoch, params.epoch):
for iter, (data_pii,
data_torch) in enumerate(zip(pii, train_loader)):
t = time.time()
# type(x)显示torch.Tensor，但是x已经在显存上
x_pii = data_pii[0]['data']
label_pii = data_pii[0]['label'].cuda()
x_torch = data_torch[0].cuda()
label_torch = data_torch[1].cuda()
x_pii = x_pii.cpu().squeeze(0).numpy().transpose((1, 2, 0))
x_torch = x_torch.cpu().squeeze(0).numpy().transpose((1, 2, 0))
import matplotlib.pyplot as plt
plt.imsave('x_pii.jpg', x_pii)
plt.imsave('x_torch.jpg', x_torch)
exit()
# print('load data time:', time.time() - t)
# t = time.time()
# output = model(x)
# loss = criterion(output, label)
# optimizer.zero_grad()
# loss.backward()
# optimizer.step()
# print('running time:', time.time() - t)
# if iter == 6:
# exit()
except KeyboardInterrupt:
save_checkpoint(model,
f'{params.save_root}/Interrupt_{epoch}_{step}.pth')
def infer(payload):
unlabeled = payload["unlabeled"]
ckpt_file = payload["ckpt_file"]
batch_size = 16
coco = COCO("./data", Transforms(), samples=unlabeled, train=True)
loader = DataLoader(coco,
shuffle=False,
batch_size=batch_size,
collate_fn=collate_fn)
config_file = "yolov3.cfg"
model = Darknet(config_file).to(device)
ckpt = torch.load(os.path.join("./log", ckpt_file))
model.load_state_dict(ckpt["model"])
model.eval()
# batch predictions from the entire test set
predictions = []
with torch.no_grad():
for img, _, _ in loader:
img = img.to(device)
# get inference output
output = model(img)
# batch predictions from 3 yolo layers
batched_prediction = []
for p in output: # (batch_size, 3, gx, gy, 85)
batch_size = p.shape[0]
p = p.view(batch_size, -1, 85)
batched_prediction.append(p)
batched_prediction = torch.cat(batched_prediction, dim=1)
predictions.append(batched_prediction)
predictions = torch.cat(predictions, dim=0)
# apply nms to predicted bounding boxes
predicted_boxes, predicted_objectness, predicted_class_dist = bbox_transform(
predictions)
# the predicted boxes are in log space relative to the anchor priors
# bring them back to normalized xyxy format
cxcy_priors = anchors.normalize("cxcy")
# expand the priors to match the dimension of predicted_boxes
batched_cxcy_priors = cxcy_priors.unsqueeze(0).repeat(
predicted_boxes.shape[0], 1, 1)
predicted_boxes = batched_gcxgcy_to_cxcy(predicted_boxes,
batched_cxcy_priors)
del batched_cxcy_priors
# convert predicted_boxes to xyxy format and perform nms
xyxy = batched_cxcy_to_xy(predicted_boxes)
del predicted_boxes # (no longer need cxcy format)
# class distribution is part of the return
# do notapply softmax to the predicted class distribution
# as we will do it internally for efficiency
outputs = {}
for i in range(len(coco)):
# get boxes, scores, and objects on each image
_xyxy, _scores = xyxy[i], predicted_objectness[i]
_pre_softmax = predicted_class_dist[i]
keep = tv.ops.nms(_xyxy, _scores, 0.5)
boxes, scores, pre_softmax = _xyxy[keep], _scores[keep], _pre_softmax[
keep]
outputs[i] = {
"boxes": boxes.cpu().numpy().tolist(),
"pre_softmax": pre_softmax.cpu().numpy().tolist(),
"scores": scores.cpu().numpy().tolist(),
}
return {"outputs": outputs}
def test(payload):
ckpt_file = payload["ckpt_file"]
batch_size = 16
coco = COCO("./data", Transforms(), train=False)
loader = DataLoader(coco,
shuffle=False,
batch_size=batch_size,
collate_fn=collate_fn)
config_file = "yolov3.cfg"
model = Darknet(config_file).to(device)
ckpt = torch.load(os.path.join("./log", ckpt_file))
model.load_state_dict(ckpt["model"])
model.eval()
# batch predictions from the entire test set
predictions = []
# keep track of ground-truth boxes and label
labels = []
with torch.no_grad():
for img, boxes, class_labels in loader:
img = img.to(device)
# get inference output
output = model(img)
for b, c in zip(boxes, class_labels):
labels.append((b, c))
# batch predictions from 3 yolo layers
batched_prediction = []
for p in output: # (bacth_size, 3, gx, gy, 85)
p = p.view(p.shape[0], -1, 85)
batched_prediction.append(p)
batched_prediction = torch.cat(batched_prediction, dim=1)
predictions.append(batched_prediction)
predictions = torch.cat(predictions, dim=0)
# apply nms to predicted bounding boxes
predicted_boxes, predicted_objectness, predicted_class_dist = bbox_transform(
predictions)
# the predicted boxes are in log space relative to the anchor priors
# bring them back to normalized xyxy format
cxcy_priors = anchors.normalize("cxcy")
# expand the priors to match the dimension of predicted_boxes
batched_cxcy_priors = cxcy_priors.unsqueeze(0).repeat(
predicted_boxes.shape[0], 1, 1)
predicted_boxes = batched_gcxgcy_to_cxcy(predicted_boxes,
batched_cxcy_priors)
del batched_cxcy_priors
# convert predicted_boxes to xyxy format and perform nms
xyxy = batched_cxcy_to_xy(predicted_boxes)
del predicted_boxes # (no longer need cxcy format)
# get predicted object
# apply softmax to the predicted class distribution
# note that bbox_tranform does not apply softmax
# because the loss we are using requires us to use raw output
predicted_objects = torch.argmax(F.softmax(predicted_class_dist, dim=-1),
dim=-1)
# predictions on the test set (value of "predictions" of the return)
prd = {}
for i in range(len(coco)):
# get boxes, scores, and objects on each image
_xyxy, _scores = xyxy[i], predicted_objectness[i]
_objects = predicted_objects[i]
keep = tv.ops.nms(_xyxy, _scores, 0.5)
boxes, scores, objects = _xyxy[keep], _scores[keep], _objects[keep]
prd[i] = {
"boxes": boxes.cpu().numpy().tolist(),
"objects": objects.cpu().numpy().tolist(),
"scores": scores.cpu().numpy().tolist(),
}
# ground-truth of the test set
# skip "difficulties" field, because every object in COCO
# should be considered reasonable
lbs = {}
for i in range(len(coco)):
boxes, class_labels = labels[i]
lbs[i] = {
"boxes": boxes.cpu().numpy().tolist(),
"objects": class_labels
}
return {"predictions": prd, "labels": lbs}
def train(payload):
labeled = payload["labeled"]
resume_from = payload["resume_from"]
ckpt_file = payload["ckpt_file"]
# hyperparameters
batch_size = 16
epochs = 2 # just for demo
lr = 1e-2
weight_decay = 1e-2
coco = COCO("./data", Transforms(), samples=labeled, train=True)
loader = DataLoader(coco,
shuffle=True,
batch_size=batch_size,
collate_fn=collate_fn)
config_file = "yolov3.cfg"
model = Darknet(config_file).to(device)
optimizer = optim.Adam(model.parameters(),
lr=lr,
weight_decay=weight_decay)
# resume model and optimizer from previous loop
if resume_from is not None:
ckpt = torch.load(os.path.join("./log", resume_from))
model.load_state_dict(ckpt["model"])
optimizer.load_state_dict(ckpt["optimizer"])
# loss function
priors = anchors.normalize("xyxy")
loss_fn = HardNegativeMultiBoxesLoss(priors, device=device)
model.train()
for img, boxes, labels in loader:
img = img.to(device)
# 3 predictions from 3 yolo layers
output = model(img)
# batch predictions on each image
batched_prediction = []
for p in output: # (batch_size, 3, gx, gy, 85)
batch_size = p.shape[0]
p = p.view(batch_size, -1, 85)
batched_prediction.append(p)
batched_prediction = torch.cat(batched_prediction, dim=1)
# (batch_size, n_priors, 85)
# the last dim of batched_prediction represent the predicted box
# batched_prediction[...,:4] is the coordinate of the predicted bbox
# batched_prediction[...,4] is the objectness score
# batched_prediction[...,5:] is the pre-softmax class distribution
# we need to apply some transforms to the those predictions
# before we can use HardNegativeMultiBoxesLoss
# In particular, the predicted bbox need to be relative to
# normalized anchor priors
# we will define another function bbox_transform
# to do those transform, since it will be used by other processes
# as well.
# see documentation on HardNegativeMultiBoxesLoss
# on its input parameters
predicted_boxes, predicted_objectness, predicted_class_dist = bbox_transform(
batched_prediction)
loss = loss_fn(predicted_boxes, predicted_objectness,
predicted_class_dist, boxes, labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# save ckpt for this loop
ckpt = {"model": model.state_dict(), "optimizer": optimizer.state_dict()}
torch.save(ckpt, os.path.join("./log", ckpt_file))
return
def main():
# Hyperparameters parser
parser = argparse.ArgumentParser()
parser.add_argument("--year", type=str, default='2012', help="used to select training set")
parser.add_argument("--set", type=str, default='train', help="used to select training set")
parser.add_argument("--epochs", type=int, default=201, help="number of epochs")
parser.add_argument("--batch_size", type=int, default=8, help="size of each image batch")
parser.add_argument("--model_def", type=str, default="config/net/resnet_dropout.cfg", help="path to model definition file")
# parser.add_argument("--model_def", type=str, default="config/net/dqnyolo_large.cfg", help="path to model definition file")
# parser.add_argument("--model_def", type=str, default="config/net/dqnyolo_mini.cfg", help="path to model definition file")
# parser.add_argument("--model_def", type=str, default="config/net/dqnyolo_tiny.cfg", help="path to model definition file")
parser.add_argument("--img_size", type=int, default=416, help="size of each image dimension")
parser.add_argument("--opt_lr", type=float, default=1e-5, help="learning rate for optimizer")
parser.add_argument("--use_gpu", default=True, help="use GPU to accelerate training")
parser.add_argument("--shuffle_train", default=True, help="shuffle the training dataset")
parser.add_argument("--checkpoint_interval", type=int, default=20, help="interval between saving model weights")
parser.add_argument("--evaluation_interval", type=int, default=10, help="interval evaluations on validation set")
# parser.add_argument("--pretrained_weights", type=str, default="data/backbone/darknet53.conv.74", help="if specified starts from checkpoint model")
# parser.add_argument("--pretrained_weights", type=str, default="logs/model/model_params_200.ckpt", help="if specified starts from checkpoint model")
parser.add_argument("--pretrained_weights", default=False, help="if specified starts from checkpoint model")
opt = parser.parse_args()
print(opt)
if opt.use_gpu is True:
if torch.cuda.is_available():
device = torch.device('cuda')
else:
raise RuntimeError("Current Torch doesn't have GPU support.")
else:
device = torch.device('cpu')
logger = SummaryWriter(exist_or_create_folder("./logs/tb/"))
# Initiate model
eval_model = Darknet(opt.model_def).to(device)
if opt.pretrained_weights:
print("Initialize model with pretrained_model")
if opt.pretrained_weights.endswith(".ckpt"):
eval_model.load_state_dict(torch.load(opt.pretrained_weights))
else:
eval_model.load_darknet_weights(opt.pretrained_weights)
else:
print("Initialize model randomly")
eval_model.apply(weights_init_normal)
# eval_model.load_state_dict(torch.load("./logs/saved_exp/master-v2/model_params_80.ckpt"))
print(eval_model)
summary(eval_model, (3, 416, 416))
learn_batch_counter = 0 # for logger update (total numbers)
batch_size = opt.batch_size
# Get dataloader
print("Begin loading train dataset ......")
t_load_data = time.time()
dataset = torchvision.datasets.VOCDetection(root='data/VOC/',
year=opt.year,
image_set=opt.set,
transforms=None,
download=True)
dataset_dict = trans_voc(dataset)
dataset = ListDataset(dataset_dict)
loader = torch.utils.data.DataLoader(
dataset,
batch_size=opt.batch_size,
shuffle=opt.shuffle_train,
pin_memory=True,
collate_fn=dataset.collate_fn,
)
print("Complete loading train dataset in {} s".format(time.time() - t_load_data))
optimizer = torch.optim.Adam(eval_model.parameters(), lr=opt.opt_lr)
# Warmup and learning rate decay
scheduler_cosine = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, opt.epochs)
# 5 epoch warmup, lr from 1e-5 to 1e-4, after that schedule as after_scheduler
scheduler_warmup = GradualWarmupScheduler(optimizer, multiplier=10, total_epoch=10,
after_scheduler=scheduler_cosine)
start_time = time.time()
for i_epoch in range(opt.epochs):
eval_model.train()
for i_batch, (_, imgs, raw_targets, transform_params, tar_boxes) in enumerate(loader):
print("\n++++++++++ i_epoch-i_batch {}-{} ++++++++++".format(i_epoch, i_batch))
batch_step_counter = 0
if len(imgs) != batch_size:
print("Current batch size is smaller than opt.batch_size!")
continue
imgs = imgs.to(device)
raw_targets = raw_targets.to(device)
tar_boxes = tar_boxes.to(device)
input_img = imgs
if i_epoch == 0 and i_batch == 0:
logger.add_graph(eval_model, input_img)
# print(raw_targets)
# print(raw_targets.size())
# print(raw_targets[:, :, :, 6:].size())
# print(raw_targets[:, :, :, 0].unsqueeze(3).size())
cls_targets = torch.cat((raw_targets[:, :, :, 0].unsqueeze(3), raw_targets[:, :, :, 6:]), 3)
# print(cls_targets.size())
loss, pred = eval_model(input_img, cls_targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_step_counter += 1
learn_batch_counter += 1
print("Ep-bt: {}-{} | Loss: {}".format(i_epoch, i_batch, loss.item()))
logger.add_scalar('loss/loss', loss.item(), learn_batch_counter)
if (i_epoch + 1) % opt.checkpoint_interval == 0:
print("Saving model in epoch {}".format(i_epoch))
torch.save(eval_model.state_dict(),
exist_or_create_folder("./logs/model/model_params_{}.ckpt".format(i_epoch)))
# Evaluate the model on the validation set
if (i_epoch + 1) % opt.evaluation_interval == 0:
precision, recall, AP, f1, ap_class = evaluate(
eval_model,
[opt.year, 'val'],
[0.5, 0.5, 0.5],
batch_size,
True,
diagnosis_code=1
)
evaluation_metrics = [
("val_precision", precision.mean()),
("val_recall", recall.mean()),
("val_mAP", AP.mean()),
("val_f1", f1.mean()),
]
for tag, value in evaluation_metrics:
logger.add_scalar("val/{}".format(tag), value.item(), i_epoch)
# Print class APs and mAP
ap_table = [["Index", "Class name", "AP"]]
for i, c in enumerate(ap_class):
ap_table += [[c, val2labels(c), "%.5f" % AP[i]]]
print(AsciiTable(ap_table).table)
print(f"---- validation mAP {AP.mean()}")
# Evaluate the model on the training set
if (i_epoch + 1) % opt.evaluation_interval == 0:
precision, recall, AP, f1, ap_class = evaluate(
eval_model,
[opt.year, 'train'],
[0.5, 0.5, 0.5],
batch_size,
True,
diagnosis_code=1
)
evaluation_metrics = [
("train_precision", precision.mean()),
("train_recall", recall.mean()),
("train_mAP", AP.mean()),
("train_f1", f1.mean()),
]
for tag, value in evaluation_metrics:
logger.add_scalar("train/{}".format(tag), value.item(), i_epoch)
# Print class APs and mAP
ap_table = [["Index", "Class name", "AP"]]
for i, c in enumerate(ap_class):
ap_table += [[c, val2labels(c), "%.5f" % AP[i]]]
print(AsciiTable(ap_table).table)
print(f"---- training mAP {AP.mean()}")
# Warmup and lr decay
scheduler_warmup.step()
# Free GPU memory
torch.cuda.empty_cache()
total_train_time = time.time() - start_time
print("Training complete in {} hours".format(total_train_time / 3600))
default=2,
help="每隔几次使用验证集")
args = parser.parse_args()
print(args)
class_names = load_classes(
r"D:\py_pro\YOLOv3-PyTorch\data\kalete\dnf_classes.txt") # 加载所有种类名称
train_path = r'D:\py_pro\YOLOv3-PyTorch\data\kalete\train.txt'
val_path = r'D:\py_pro\YOLOv3-PyTorch\data\kalete\val.txt'
print("载入网络...")
model = Darknet(args.cfg)
pretrained = True
if pretrained:
model.load_state_dict(torch.load(args.weights))
else:
# 随机初始化权重,会对模型进行高斯随机初始化
model.apply(weights_init_normal)
print("网络权重加载成功.")
# 设置网络输入图片尺寸大小与学习率
reso = int(model.net_info["height"])
lr = float(model.net_info["learning_rate"])
assert reso % 32 == 0 # 判断如果不是32的整数倍就抛出异常
assert reso > 32 # 判断如果网络输入图片尺寸小于32也抛出异常
if CUDA:
model.cuda()
print(opt)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
data_config = parse_data_config(opt.data_config)
valid_path = data_config["valid"]
class_names = load_classes(data_config["names"])
# Initiate model
model = Darknet(opt.model_def).to(device)
if opt.weights_path.endswith(".weights"):
# Load darknet weights
model.load_darknet_weights(opt.weights_path)
else:
# Load checkpoint weights
model.load_state_dict(torch.load(opt.weights_path))
print("Compute mAP...")
precision, recall, AP, f1, ap_class = evaluate(
model,
path=valid_path,
iou_thres=opt.iou_thres,
conf_thres=opt.conf_thres,
nms_thres=opt.nms_thres,
img_size=opt.img_size,
batch_size=8,
)
print("Average Precisions:")
for i, c in enumerate(ap_class):