def model_init(model_name):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if model_name == 'retinanet' :
#weight_file_path = '/content/retinanet/resnet34-333f7ec4.pth'
#weight_file_path = '/content/retinanet/CP_epoch5.pth'
weight_file_path = '/content/retinanet/retinanet50_pretrained.pth'
total_keys = len(list(torch.load(weight_file_path).keys()))
# Create the model
if total_keys >= 102 and total_keys < 182 :
retinanet = model.resnet18(num_classes=num_classes, pretrained=False)
elif total_keys >= 182 and total_keys < 267:
retinanet = model.resnet34(num_classes=num_classes, pretrained=False)
elif total_keys >= 267 and total_keys < 522:
retinanet = model.resnet50(num_classes=num_classes, pretrained=False)
elif total_keys >= 522 and total_keys < 777:
retinanet = model.resnet101(num_classes=num_classes, pretrained=False)
elif total_keys >= 777:
retinanet = model.resnet152(num_classes=num_classes, pretrained=False)
else:
raise ValueError('Unsupported model backbone, must be one of resnet18, resnet34, resnet50, resnet101, resnet152')
retinanet.load_state_dict(torch.load(weight_file_path, map_location=device), strict=False) # Initialisng Model with loaded weights
print('model initialized..')
return retinanet, device
def main():
weights_path = "./resNet34.pth"
model = resnet34(num_classes=5)
model.load_state_dict(torch.load(weights_path, map_location=device))
model.to(device)
# validate_model(model)
# module = model.conv1
# print(list(module.named_parameters()))
# # print(list(module.named_buffers()))
#
# # 裁剪50%的卷积核
# prune.ln_structured(module, name="weight", amount=0.5, n=2, dim=0)
# print(list(module.weight))
# print(module.weight.shape)
# # print(list(module.named_buffers()))
#
# prune.remove(module, "weight")
# print(module.weight.shape)
# 收集所有需要裁剪的卷积核
parameters_to_prune = []
for name, module in model.named_modules():
if isinstance(module, torch.nn.Conv2d):
parameters_to_prune.append((module, "weight"))
# 对卷积核进行剪枝处理
prune.global_unstructured(parameters_to_prune,
pruning_method=prune.L1Unstructured,
amount=0.5)
# 统计剪枝比例
count_sparsity(model, p=False)
# 验证剪枝后的模型
validate_model(model)
def pytorch_model_speed(data_loader, val_num):
net = resnet34(num_classes=5)
# load weights
model_weight_path = "./resNet34.pth"
check_path_exist(model_weight_path)
net.load_state_dict(torch.load(model_weight_path, map_location=device),
strict=False)
net.eval()
test_data = torch.rand((1, 3, 224, 224))
net(test_data.to(device))
forward_time = 0
acc = 0.0 # accumulate accurate number / epoch
with torch.no_grad():
for val_data in tqdm(data_loader, desc="Running pytorch model..."):
val_images, val_labels = val_data
t1 = time.time()
outputs = net(val_images.to(
device)) # eval model only have last output layer
t2 = time.time()
forward_time += (t2 - t1)
predict_y = torch.max(outputs, dim=1)[1]
acc += (predict_y == val_labels.to(device)).sum().item()
val_accurate = acc / val_num
fps = round(val_num / forward_time, 1)
print("pytorch info:\nfps: {}/s accuracy: {}\n".format(fps, val_accurate))
return fps, val_accurate, "Pytorch(not opt)"
def __init__(self, model_weight_path, json_file):
# self.data_transform = transforms.Compose([transforms.Resize(256),
# transforms.CenterCrop(224),
# transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
#
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
print(self.device)
self.device_cpu = torch.device("cpu")
self.data_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
json_filef = open(json_file, "r", encoding="utf-8")
self.class_indict = json.load(json_filef)
# create model
if (str.find(model_weight_path, "res50.pth") > 0):
self.model = resnet50(num_classes=len(self.class_indict)) # res50
else:
self.model = resnet34(num_classes=len(self.class_indict))
# load model weights
self.model.load_state_dict(torch.load(model_weight_path))
self.model.eval()
self.model.to(self.device)
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
data_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# load image
img_path_list = ["./tulip.jpg", "./rose.jpg"]
img_list = []
# img_list= Image.open(img_path_list).convert('RGB')
for img_path in img_path_list:
assert os.path.exists(img_path), "file: '{}' dose not exist.".format(
img_path)
img = Image.open(img_path)
img = data_transform(img)
img_list.append(img)
# batch img
batch_img = torch.stack(img_list,
dim=0) #torch.stack()是将原来的几个tensor按照一定方式进行堆叠，
# 然后在按照堆叠后的维度进行切分。一共有0，1，2三个维度
# read class_indict
json_path = './class_indices.json'
assert os.path.exists(json_path), "file: '{}' dose not exist.".format(
json_path)
json_file = open(json_path, "r")
class_indict = json.load(json_file)
# create model
model = resnet34(num_classes=5).to(device)
# load model weights
weights_path = "./resNet34.pth"
assert os.path.exists(weights_path), "file: '{}' dose not exist.".format(
weights_path)
model.load_state_dict(torch.load(weights_path, map_location=device))
# prediction
model.eval()
with torch.no_grad():
# predict class
output = model(batch_img.to(device)).cpu()
predict = torch.softmax(output, dim=1)
probs, classes = torch.max(predict, dim=1)
for idx, (pro, cla) in enumerate(zip(probs, classes)):
print("image: {} class: {} prob: {:.3}".format(
img_path_list[idx], class_indict[str(cla.numpy())],
pro.numpy()))
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
data_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
# load image
img_path = "../tulip.jpg"
assert os.path.exists(img_path), "file: '{}' dose not exist.".format(
img_path)
img = Image.open(img_path)
plt.imshow(img)
# [N, C, H, W]
img = data_transform(img)
# expand batch dimension
img = torch.unsqueeze(img, dim=0)
# read class_indict
json_path = './class_indices.json'
assert os.path.exists(json_path), "file: '{}' dose not exist.".format(
json_path)
json_file = open(json_path, "r")
class_indict = json.load(json_file)
# create model
model = resnet34(num_classes=5).to(device)
# load model weights
weights_path = "./resNet34.pth"
assert os.path.exists(weights_path), "file: '{}' dose not exist.".format(
weights_path)
model.load_state_dict(torch.load(weights_path, map_location=device))
# prediction
model.eval()
with torch.no_grad():
# predict class
output = torch.squeeze(model(img.to(device))).cpu()
predict = torch.softmax(output, dim=0)
predict_cla = torch.argmax(predict).numpy()
print_res = "class: {} prob: {:.3}".format(
class_indict[str(predict_cla)], predict[predict_cla].numpy())
plt.title(print_res)
for i in range(len(predict)):
print("class: {:10} prob: {:.3}".format(class_indict[str(i)],
predict[i].numpy()))
plt.show()
def load_model(self):
self.checkpoint = torch.load(self.model_checkpoint_file_path,
map_location=lambda storage, loc: storage)
self.model_args = self.checkpoint['args']
self.num_classes = None
if self.model_args.model_type == 'food179':
self.num_classes = 179
elif self.model_args.model_type == 'nsfw':
self.num_classes = 5
else:
raise ('Not Implemented!')
if self.model_args.model_arc == 'resnet18':
self.model = model.resnet18(num_classes=self.num_classes,
zero_init_residual=True)
elif self.model_args.model_arc == 'resnet34':
self.model = model.resnet34(num_classes=self.num_classes,
zero_init_residual=True)
elif self.model_args.model_arc == 'resnet50':
self.model = model.resnet50(num_classes=self.num_classes,
zero_init_residual=True)
elif self.model_args.model_arc == 'resnet101':
self.model = model.resnet101(num_classes=self.num_classes,
zero_init_residual=True)
elif self.model_args.model_arc == 'resnet152':
self.model = model.resnet152(num_classes=self.num_classes,
zero_init_residual=True)
elif self.model_args.model_arc == 'mobilenet':
self.model = model.MobileNetV2(n_class=self.num_classes,
input_size=256)
else:
raise ('Not Implemented!')
self.model = nn.DataParallel(self.model)
self.model.load_state_dict(self.checkpoint['model_state_dict'])
self.model_epoch = self.checkpoint['epoch']
self.model_test_acc = self.checkpoint['test_acc']
self.model_best_acc = self.checkpoint['best_acc']
self.model_test_acc_top5 = self.checkpoint['test_acc_top5']
self.model_class_to_idx = self.checkpoint['class_to_idx']
self.model_idx_to_class = {
v: k
for k, v in self.model_class_to_idx.items()
}
self.model_train_history_dict = self.checkpoint['train_history_dict']
self.mean = self.checkpoint['NORM_MEAN']
self.std = self.checkpoint['NORM_STD']
self.model.eval()
return
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# load pretrain weights
# download url: https://download.pytorch.org/models/resnet34-333f7ec4.pth
model_weight_path = "./resnet34-pre.pth"
assert os.path.exists(model_weight_path), "file {} does not exist.".format(
model_weight_path)
# option1
net = resnet34()
net.load_state_dict(torch.load(model_weight_path, map_location=device))
# change fc layer structure
in_channel = net.fc.in_features
net.fc = nn.Linear(in_channel, 5)
def main(args=None):
data_set = {
x: guipang(cfg=cfg['dataset_guipang'], part=x) for x in ['train', 'val']
}
# data_set = {
# x: qiafan(cfg=cfg['dataset_qiafan'], part=x) for x in ['train', 'val']
# }
data_loader = {
x: data.DataLoader(data_set[x], batch_size=cfg['batch_size'],
num_workers=4, shuffle=True, pin_memory=False)
for x in ['train', 'val']
}
# Create the model
if cfg['depth'] == 18:
retinanet = model.resnet18(
num_classes=dataset_train.num_classes(), pretrained=True)
elif cfg['depth'] == 34:
retinanet = model.resnet34(
num_classes=dataset_train.num_classes(), pretrained=True)
elif cfg['depth'] == 50:
retinanet = model.resnet50(
num_classes=dataset_train.num_classes(), pretrained=True)
elif cfg['depth'] == 101:
retinanet = model.resnet101(
num_classes=dataset_train.num_classes(), pretrained=True)
elif cfg['depth'] == 152:
retinanet = model.resnet152(
num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
def __init__(self, model_weight_path, json_file):
# self.data_transform = transforms.Compose([transforms.Resize(256),
# transforms.CenterCrop(224),
# transforms.ToTensor(),
# transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
self.data_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
json_filef = open(json_file, "r", encoding="utf-8")
self.class_indict = json.load(json_filef)
# create model
self.model = resnet34(num_classes=len(self.class_indict))
# load model weights
self.model.load_state_dict(torch.load(model_weight_path))
self.model.eval()
def get_network(args):
'''
return the required network
'''
if args.net == 'resnet34':
from model import resnet34
net = resnet34(args.num_classes)
elif args.net == 'resnet50':
from model import resnet50
net = resnet50(args.num_classes)
else:
print(args.net +
' is not supported, please enter resnet34 or resnet50')
if args.gpu:
net = net.cuda()
return net
def main(args=None):
from dataloader import JinNanDataset, Augmenter, UnNormalizer, Normalizer,Resizer
from torch.utils.data import Dataset, DataLoader
from torchvision import datasets, models, transforms
import model
import torch
import argparse
parser = argparse.ArgumentParser(description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset',default='jingnan', help='Dataset type, must be one of csv or coco.')
parser.add_argument('--threshold',help='treshold')
parser.add_argument('--dataset_path', help='Path to file containing training and validation annotations (optional, see readme)')
parser.add_argument('--model_path',help=('the model path'))
parser.add_argument('--depth', help='Resnet depth, must be one of 18, 34, 50, 101, 152', type=int, default=50)
parser = parser.parse_args(args)
dataset_val=JinNanDataset(parser.dataset_path, set_name='val', transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_val.num_classes(), pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_val.num_classes(), pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_val.num_classes(), pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_val.num_classes(), pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_val.num_classes(), pretrained=True)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
retinanet=torch.load(parser.model_path)
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet.eval()
print('Evaluating dataset')
evaluate_jinnan(dataset_val, retinanet)
def build(self, depth=50, learning_rate=1e-5, ratios=[0.5, 1, 2],
scales=[2 ** 0, 2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)]):
# Create the model
if depth == 18:
retinanet = model.resnet18(num_classes=self.dataset_train.num_classes(), ratios=ratios, scales=scales,
weights_dir=self.weights_dir_path,
pretrained=True)
elif depth == 34:
retinanet = model.resnet34(num_classes=self.dataset_train.num_classes(), ratios=ratios, scales=scales,
weights_dir=self.weights_dir_path,
pretrained=True)
elif depth == 50:
retinanet = model.resnet50(num_classes=self.dataset_train.num_classes(), ratios=ratios, scales=scales,
weights_dir=self.weights_dir_path,
pretrained=True)
elif depth == 101:
retinanet = model.resnet101(num_classes=self.dataset_train.num_classes(), ratios=ratios, scales=scales,
weights_dir=self.weights_dir_path,
pretrained=True)
elif depth == 152:
retinanet = model.resnet152(num_classes=self.dataset_train.num_classes(), ratios=ratios, scales=scales,
weights_dir=self.weights_dir_path,
pretrained=True)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
self.retinanet = retinanet.to(device=self.device)
self.retinanet.training = True
self.optimizer = optim.Adam(self.retinanet.parameters(), lr=learning_rate)
self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, patience=3, verbose=True)
if self.checkpoint is not None:
self.retinanet.load_state_dict(self.checkpoint['model'])
self.optimizer.load_state_dict(self.checkpoint['optimizer'])
self.scheduler.load_state_dict(self.checkpoint['scheduler']) # TODO: test this, is it done right?
# TODO is it right to resume_read_trial optimizer and schedular like this???
self.ratios = ratios
self.scales = scales
self.depth = depth
def set_models(self, dataset_train):
# Create the model
if self.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif self.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif self.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif self.depth == 101:
retinanet = model.resnet101(
num_classes=dataset_train.num_classes(), pretrained=True)
elif self.depth == 152:
retinanet = model.resnet152(
num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
# dim = 0 [30, xxx] -> [10, ...], [10, ...], [10, ...] on 3 GPUs
retinanet = nn.DataParallel(retinanet)
self.retinanet = retinanet.to(self.device)
self.retinanet.training = True
self.optimizer = optim.Adam(self.retinanet.parameters(), lr=self.lr)
# This lr_shceduler reduce the learning rate based on the models's validation loss
self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer,
patience=3,
verbose=True)
self.loss_hist = collections.deque(maxlen=500)
def load_model(model_arch,numOfClasses):
if model_arch.endswith('resnet34') or model_arch.endswith('resnet34') :
NN = resnet34(pretrained = False)
num_features = NN.fc.in_features
NN.fc = nn.Linear(num_features,numOfClasses)
elif model_arch.endswith('resnet50_pretrained') or model_arch.endswith('resnet50'):
NN = models.resnet50(pretrained = True)
num_features = NN.fc.in_features
NN.fc = nn.Linear(num_features,numOfClasses)
elif model_arch.endswith('resnet152_pretrained'):
NN = models.resnet152(pretrained = True)
num_features = NN.fc.in_features
NN.fc = nn.Linear(num_features,numOfClasses)
elif model_arch.endswith('resnet152'):
NN = models.resnet152(pretrained = False)
num_features = NN.fc.in_features
NN.fc = nn.Linear(num_features,numOfClasses)
elif model_arch.endswith('densenet121_pretrained'):
NN = models.densenet121(pretrained=True)
num_ftrs = NN.classifier.in_features
NN.classifier = nn.Linear(num_ftrs, numOfClasses)
return NN
def main(args=None):
parser = argparse.ArgumentParser(description='Simple testing script for RetinaNet network.')
parser.add_argument('--dataset', help='Dataset type, must be one of csv or coco.',default = "csv")
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument('--csv_classes', help='Path to file containing class list (see readme)',default="binary_class.csv")
parser.add_argument('--csv_val', help='Path to file containing validation annotations (optional, see readme)')
parser.add_argument('--csv_box_annot', help='Path to file containing predicted box annotations ')
parser.add_argument('--depth', help='Resnet depth, must be one of 18, 34, 50, 101, 152', type=int, default=18)
parser.add_argument('--epochs', help='Number of epochs', type=int, default=500)
parser.add_argument('--model', help='Path of .pt file with trained model',default = 'esposallescsv_retinanet_0.pt')
parser.add_argument('--model_out', help='Path of .pt file with trained model to save',default = 'trained')
parser.add_argument('--score_threshold', help='Score above which boxes are kept',default=0.15)
parser.add_argument('--nms_threshold', help='Score above which boxes are kept',default=0.2)
parser.add_argument('--max_epochs_no_improvement', help='Max epochs without improvement',default=100)
parser.add_argument('--max_boxes', help='Max boxes to be fed to recognition',default=50)
parser.add_argument('--seg_level', help='Line or word, to choose anchor aspect ratio',default='line')
parser.add_argument('--htr_gt_box',help='Train recognition branch with box gt (for debugging)',default=False)
parser = parser.parse_args(args)
# Create the data loaders
if parser.dataset == 'csv':
if parser.csv_classes is None:
raise ValueError('Must provide --csv_classes when training on COCO,')
if parser.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=parser.csv_val, class_list=parser.csv_classes, transform=transforms.Compose([Normalizer(), Resizer()]))
if parser.csv_box_annot is not None:
box_annot_data = CSVDataset(train_file=parser.csv_box_annot, class_list=parser.csv_classes, transform=transforms.Compose([Normalizer(), Resizer()]))
else:
box_annot_data = None
else:
raise ValueError('Dataset type not understood (must be csv or coco), exiting.')
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
dataloader_val = DataLoader(dataset_val, num_workers=0, collate_fn=collater, batch_sampler=sampler_val)
if box_annot_data is not None:
sampler_val = AspectRatioBasedSampler(box_annot_data, batch_size=1, drop_last=False)
dataloader_box_annot = DataLoader(box_annot_data, num_workers=0, collate_fn=collater, batch_sampler=sampler_val)
else:
dataloader_box_annot = dataloader_val
if not os.path.exists('trained_models'):
os.mkdir('trained_models')
# Create the model
alphabet=dataset_val.alphabet
if os.path.exists(parser.model):
retinanet = torch.load(parser.model)
else:
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_val.num_classes(), pretrained=True,max_boxes=int(parser.max_boxes),score_threshold=float(parser.score_threshold),seg_level=parser.seg_level,alphabet=alphabet)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
#retinanet = torch.load('../Documents/TRAINED_MODELS/pytorch-retinanet/esposallescsv_retinanet_99.pt')
#print "LOADED pretrained MODEL\n\n"
optimizer = optim.Adam(retinanet.parameters(), lr=1e-4)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=4, verbose=True)
loss_hist = collections.deque(maxlen=500)
ctc = CTCLoss()
retinanet.module.freeze_bn()
best_cer = 1000
epochs_no_improvement=0
cers=[]
retinanet.eval()
retinanet.module.epochs_only_det = 0
#retinanet.module.htr_gt_box = False
retinanet.training=False
if parser.score_threshold is not None:
retinanet.module.score_threshold = float(parser.score_threshold)
'''if parser.dataset == 'csv' and parser.csv_val is not None:
print('Evaluating dataset')
'''
mAP = csv_eval.evaluate(dataset_val, retinanet,score_threshold=retinanet.module.score_threshold)
aps = []
for k,v in mAP.items():
aps.append(v[0])
print ("VALID mAP:",np.mean(aps))
print("score th",retinanet.module.score_threshold)
for idx,data in enumerate(dataloader_box_annot):
print("Eval CER on validation set:",idx,"/",len(dataloader_box_annot),"\r")
if box_annot_data:
image_name = box_annot_data.image_names[idx].split('/')[-1].split('.')[-2]
else:
image_name = dataset_val.image_names[idx].split('/')[-1].split('.')[-2]
#generate_pagexml(image_name,data,retinanet,parser.score_threshold,parser.nms_threshold,dataset_val)
text_gt_path="/".join(dataset_val.image_names[idx].split('/')[:-1])
text_gt = os.path.join(text_gt_path,image_name+'.txt')
f =open(text_gt,'r')
text_gt_lines=f.readlines()[0]
transcript_pred = get_transcript(image_name,data,retinanet,retinanet.module.score_threshold,float(parser.nms_threshold),dataset_val,alphabet)
cers.append(float(editdistance.eval(transcript_pred,text_gt_lines))/len(text_gt_lines))
print("GT",text_gt_lines)
print("PREDS SAMPLE:",transcript_pred)
print("VALID CER:",np.mean(cers),"best CER",best_cer)
print("GT",text_gt_lines)
print("PREDS SAMPLE:",transcript_pred)
print("VALID CER:",np.mean(cers),"best CER",best_cer)
def main(args=None):
#def main(epoch):
parser = argparse.ArgumentParser(description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset', help='Dataset type, must be one of csv or coco.')
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument('--csv_train', help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes', help='Path to file containing class list (see readme)')
parser.add_argument('--csv_val', help='Path to file containing validation annotations (optional, see readme)')
parser.add_argument('--depth', help='Resnet depth, must be one of 18, 34, 50, 101, 152', type=int, default=50)
parser.add_argument('--epochs', help='Number of epochs', type=int, default=100)
#parser.add_argument('--resume', '-r', action='store_true', help='resume from checkpoint')
parser.add_argument('--start-epoch', default=0, type=int, help='manual epoch number (useful on restarts)')
parser.add_argument('--resume', default='', type=str, metavar='PATH', help='path to latest checkpoint (default: none)')
parser = parser.parse_args(args)
#args = parser.parse_args()
#parser = parser.parse_args(epoch)
# Create the data loaders
if parser.dataset == 'coco':
if parser.coco_path is None:
raise ValueError('Must provide --coco_path when training on COCO,')
dataset_train = CocoDataset(parser.coco_path, set_name='train2017', transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
dataset_val = CocoDataset(parser.coco_path, set_name='val2017', transform=transforms.Compose([Normalizer(), Resizer()]))
elif parser.dataset == 'csv':
if parser.csv_train is None:
raise ValueError('Must provide --csv_train when training on COCO,')
if parser.csv_classes is None:
raise ValueError('Must provide --csv_classes when training on COCO,')
dataset_train = CSVDataset(train_file=parser.csv_train, class_list=parser.csv_classes, transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
if parser.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=parser.csv_val, class_list=parser.csv_classes, transform=transforms.Compose([Normalizer(), Resizer()]))
else:
raise ValueError('Dataset type not understood (must be csv or coco), exiting.')
sampler = AspectRatioBasedSampler(dataset_train, batch_size=4, drop_last=False)
dataloader_train = DataLoader(dataset_train, num_workers=3, collate_fn=collater, batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
dataloader_val = DataLoader(dataset_val, num_workers=3, collate_fn=collater, batch_sampler=sampler_val)
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
#retinanet().load_state_dict(torch.load('/users/wenchi/ghwwc/Pytorch-retinanet-master/resnet50-19c8e357.pth'))
#if True:
#print('==> Resuming from checkpoint..')
#checkpoint = torch.load('/users/wenchi/ghwwc/Pytorch-retinanet-master/coco_retinanet_2.pt')
#retinanet().load_state_dict(checkpoint)
#best_loss = checkpoint['loss']
#start_epoch = checkpoint['epoch']
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
#optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
optimizer = optim.SGD(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
#retinanet.freeze_bn() #for train from a middle state
retinanet.module.freeze_bn() #for train from the very beginning
print('Num training images: {}'.format(len(dataset_train)))
for epoch_num in range(parser.start_epoch, parser.epochs):
if parser.resume:
if os.path.isfile(parser.resume):
print("=>loading checkpoint '{}'".format(parser.resume))
checkpoint = torch.load(parser.resume)
print(parser.start_epoch)
#parser.start_epoch = checkpoint['epoch']
#retinanet.load_state_dict(checkpoint['state_dict'])
retinanet=checkpoint
#retinanet.load_state_dict(checkpoint)
print(retinanet)
#optimizer.load_state_dict(checkpoint)
print("=> loaded checkpoint '{}' (epoch {})".format(parser.resume, checkpoint))
else:
print("=> no checkpoint found at '{}'".format(parser.resume))
retinanet.train()
retinanet.freeze_bn()
#retinanet.module.freeze_bn()
if parser.dataset == 'coco':
print('Evaluating dataset')
coco_eval.evaluate_coco(dataset_val, retinanet)
elif parser.dataset == 'csv' and parser.csv_val is not None:
print('Evaluating dataset')
mAP = csv_eval.evaluate(dataset_val, retinanet)
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet([data['img'].cuda().float(), data['annot'].cuda()])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
print('Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'.format(epoch_num, iter_num, float(classification_loss), float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
if parser.dataset == 'coco':
print('Evaluating dataset')
coco_eval.evaluate_coco(dataset_val, retinanet)
elif parser.dataset == 'csv' and parser.csv_val is not None:
print('Evaluating dataset')
mAP = csv_eval.evaluate(dataset_val, retinanet)
scheduler.step(np.mean(epoch_loss))
#torch.save(retinanet.module, '{}_retinanet_101_{}.pt'.format(parser.dataset, epoch_num))
torch.save(retinanet, '{}_retinanet_dilation_experiment1_{}.pt'.format(parser.dataset, epoch_num))
name = '{}_retinanet_dilation_experiment1_{}.pt'.format(parser.dataset, epoch_num)
parser.resume = '/users/wenchi/ghwwc/pytorch-retinanet-master_new/name'
retinanet.eval()
torch.save(retinanet, 'model_final_dilation_experiment1.pt'.format(epoch_num))
def main(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
print(args)
print(
'Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/'
)
# 实例化SummaryWriter对象
tb_writer = SummaryWriter(log_dir="runs/flower_experiment")
if os.path.exists("./weights") is False:
os.makedirs("./weights")
# 划分数据为训练集和验证集
train_images_path, train_images_label, val_images_path, val_images_label = read_split_data(
args.data_path)
# 定义训练以及预测时的预处理方法
data_transform = {
"train":
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
"val":
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
}
# 实例化训练数据集
train_data_set = MyDataSet(images_path=train_images_path,
images_class=train_images_label,
transform=data_transform["train"])
# 实例化验证数据集
val_data_set = MyDataSet(images_path=val_images_path,
images_class=val_images_label,
transform=data_transform["val"])
batch_size = args.batch_size
# 计算使用num_workers的数量
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using {} dataloader workers every process'.format(nw))
train_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
val_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_data_set.collate_fn)
# 实例化模型
model = resnet34(num_classes=args.num_classes).to(device)
# 将模型写入tensorboard
init_img = torch.zeros((1, 3, 224, 224), device=device)
tb_writer.add_graph(model, init_img)
# 如果存在预训练权重则载入
if os.path.exists(args.weights):
weights_dict = torch.load(args.weights, map_location=device)
load_weights_dict = {
k: v
for k, v in weights_dict.items()
if model.state_dict()[k].numel() == v.numel()
}
model.load_state_dict(load_weights_dict, strict=False)
else:
print("not using pretrain-weights.")
# 是否冻结权重
if args.freeze_layers:
print("freeze layers except fc layer.")
for name, para in model.named_parameters():
# 除最后的全连接层外，其他权重全部冻结
if "fc" not in name:
para.requires_grad_(False)
pg = [p for p in model.parameters() if p.requires_grad]
optimizer = optim.SGD(pg, lr=args.lr, momentum=0.9, weight_decay=0.005)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: ((1 + math.cos(x * math.pi / args.epochs)) / 2) * (
1 - args.lrf) + args.lrf # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
for epoch in range(args.epochs):
# train
mean_loss = train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_loader,
device=device,
epoch=epoch)
# update learning rate
scheduler.step()
# validate
acc = evaluate(model=model, data_loader=val_loader, device=device)
# add loss, acc and lr into tensorboard
print("[epoch {}] accuracy: {}".format(epoch, round(acc, 3)))
tags = ["train_loss", "accuracy", "learning_rate"]
tb_writer.add_scalar(tags[0], mean_loss, epoch)
tb_writer.add_scalar(tags[1], acc, epoch)
tb_writer.add_scalar(tags[2], optimizer.param_groups[0]["lr"], epoch)
# add figure into tensorboard
fig = plot_class_preds(net=model,
images_dir="./plot_img",
transform=data_transform["val"],
num_plot=5,
device=device)
if fig is not None:
tb_writer.add_figure("predictions vs. actuals",
figure=fig,
global_step=epoch)
# add conv1 weights into tensorboard
tb_writer.add_histogram(tag="conv1",
values=model.conv1.weight,
global_step=epoch)
tb_writer.add_histogram(tag="layer1/block0/conv1",
values=model.layer1[0].conv1.weight,
global_step=epoch)
# save weights
torch.save(model.state_dict(), "./weights/model-{}.pth".format(epoch))
def main():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("using {} device.".format(device))
data_transform = {
#torchvision.transforms是pytorch中的图像预处理包。一般用Compose把多个步骤整合到一起
"train": transforms.Compose([transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),#依据概率p对PIL图片进行垂直翻转
transforms.ToTensor(),#将PIL Image或者 ndarray 转换为tensor，并且归一化至[0-1]；
# 归一化至[0-1]是直接除以255，若自己的ndarray数据尺度有变化，则需要自行修改。
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),
##transforms.Normalize对图像标准化处理的参数，这里根据官网定义的
#transforms.Normalize使用如下公式进行归一化：channel=（channel-mean）/std。前面三个是mean、后面三个是std
#因为totensor已经转化为[0-1],故最后的结果范围是[-1~1]
"val": transforms.Compose([transforms.Resize(256), #，长宽比固定不动，最小边缩放到256
transforms.CenterCrop(224), #根据中心点剪切到224
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])}#标准化处理，防止突出数值较高的指标在综合分析中的作用。
data_root = os.path.abspath(os.path.join(os.getcwd(), "..\Test5_resnet")) # 返回绝对路径.表示返回上上一层目录
image_path = os.path.join(data_root, "data_set", "flower_data")# flower data set path
# image_path = os.path.join(data_root, "data_set/flower_data") #再从根目录开始向下进行完整目录的拼接。
# image_path = os.path.join("D:\pycharmproject\deep learning for images"
# "\deep-learning-for-image-processing-master\pytorch_classification"
# "\Test5_resnet/flower_data/data_set/flower_data")
assert os.path.exists(image_path), "{} path does not exist.".format(image_path)
#断言（assert）作为一种软件调试的方法，提供了一种在代码中进行正确性检查的机制。
train_dataset = datasets.ImageFolder(root=os.path.join(image_path, "train"),
transform=data_transform["train"])
'''
其构造函数如下：
#ImageFolder假设所有的文件按文件夹保存，每个文件夹下存储同一个类别的图片，文件夹名为类名，
ImageFolder(root, transform=None, target_transform=None, loader=default_loader)
它主要有四个参数：
root：在root指定的路径下寻找图片
transform：对PIL Image进行的转换操作，transform的输入是使用loader读取图片的返回对象
target_transform：对label的转换
loader：给定路径后如何读取图片，默认读取为RGB格式的PIL Image对象
'''
train_num = len(train_dataset)
# {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}
flower_list = train_dataset.class_to_idx #类别名与数字类别的映射关系字典（class_to_idx）
cla_dict = dict((val, key) for key, val in flower_list.items())
# write dict into json file
json_str = json.dumps(cla_dict, indent=4) #json.dumps()是把python对象转换成json对象的一个过程，
# 生成的是字符串。cla_dict是转化成json的对象，indent:参数根据数据格式缩进显示，读起来更加清晰。
with open('class_indices.json', 'w') as json_file:
json_file.write(json_str)
'''
因为传统打开文件是f = open('test.txt','r')如果存在则继续执行，然后f.close()；如果不存在则会报错
改用with open as简化上面过程。无论有无异常均可自动调用f.close()方法。
'''
batch_size = 16
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8]) # number of workers 获取电脑的cpu个数。
print('Using {} dataloader workers every process'.format(nw))
train_loader = torch.utils.data.DataLoader(train_dataset,#从实例化的dataset当中取得图片，
# 然后打包成一个一个batch，然后输入网络进行训练
batch_size=batch_size, shuffle=True,#打乱数据集
num_workers=nw) #如果使用的是linux系统，
# 将num_workers设置成一个大于0的数加快图像预处理进程
validate_dataset = datasets.ImageFolder(root=os.path.join(image_path, "val"),#在image-path中找到val文件夹
transform=data_transform["val"])
val_num = len(validate_dataset)
validate_loader = torch.utils.data.DataLoader(validate_dataset,
batch_size=batch_size, shuffle=False,
num_workers=nw)
print("using {} images for training, {} images fot validation.".format(train_num,
val_num))
net = resnet34()
# load pretrain weights
# download url: https://download.pytorch.org/models/resnet34-333f7ec4.pth
model_weight_path = "./resnet34-pre.pth" #里面是预训练的权重
assert os.path.exists(model_weight_path), "file {} does not exist.".format(model_weight_path)
missing_keys, unexpected_keys = net.load_state_dict(torch.load(model_weight_path), strict=False) #载入模型权重
# for param in net.parameters():
# param.requires_grad = False
# change fc layer structure
in_channel = net.fc.in_features #这里的fc是model.ResNet里的fc;in_features是输入特征矩阵的深度
net.fc = nn.Linear(in_channel, 5)
net.to(device)
loss_function = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.0001)
epochs = 3
best_acc = 0.0
save_path = './resNet34.pth'
train_steps = len(train_loader)
for epoch in range(epochs):
# train
net.train()
running_loss = 0.0
train_bar = tqdm(train_loader) #tqdm显示进度
for step, data in enumerate(train_bar):
images, labels = data
optimizer.zero_grad()
logits = net(images.to(device))
loss = loss_function(logits, labels.to(device))
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
epochs,
loss)
# validate
net.eval()
acc = 0.0 # accumulate accurate number / epoch
with torch.no_grad():
val_bar = tqdm(validate_loader, colour='green')
for val_data in val_bar:
val_images, val_labels = val_data
outputs = net(val_images.to(device))
# loss = loss_function(outputs, test_labels)
predict_y = torch.max(outputs, dim=1)[1]
acc += torch.eq(predict_y, val_labels.to(device)).sum().item()
val_bar.desc = "valid epoch[{}/{}]".format(epoch + 1,
epochs)
val_accurate = acc / val_num
print('[epoch %d] train_loss: %.3f val_accuracy: %.3f' %
(epoch + 1, running_loss / train_steps, val_accurate))
if val_accurate > best_acc:
best_acc = val_accurate
torch.save(net.state_dict(), save_path)
print('Finished Training')
def main(args=None):
parser = argparse.ArgumentParser(
description='Simple training script for training a RetinaNet network.')
parser.add_argument('--coco_path',
help='Path to COCO directory',
type=str,
default='./data/coco')
parser.add_argument(
'--depth',
help='Resnet depth, must be one of 18, 34, 50, 101, 152',
type=int,
default=50)
parser.add_argument('--checkpoint',
help='The path to the checkpoint.',
type=str,
default=None)
parser.add_argument('--epochs',
help='Number of epochs',
type=int,
default=100)
parser.add_argument('--batch_size',
help='Number of batch',
type=int,
default=16)
parser.add_argument('--gpu_ids',
help='Gpu parallel',
type=str,
default='1, 2')
parser = parser.parse_args(args)
# Create the data lodaders
dataset_train = CocoDataset(parser.coco_path,
set_name='train2017',
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
dataset_val = CocoDataset(parser.coco_path,
set_name='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
sampler = AspectRatioBasedSampler(dataset_train,
batch_size=4,
drop_last=False)
dataloader_train = DataLoader(dataset_train,
num_workers=16,
collate_fn=collater,
batch_sampler=sampler)
sampler_val = AspectRatioBasedSampler(dataset_val,
batch_size=1,
drop_last=False)
dataloader_val = DataLoader(dataset_val,
num_workers=3,
collate_fn=collater,
batch_sampler=sampler_val)
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
gpu_ids = parser.gpu_ids.split(',')
device = torch.device("cuda:" + gpu_ids[0])
torch.cuda.set_device(device)
gpu_ids = list(map(int, gpu_ids))
retinanet = torch.nn.DataParallel(retinanet, device_ids=gpu_ids).to(device)
if parser.checkpoint:
pretrained = torch.load(parser.checkpoint).state_dict()
retinanet.module.load_state_dict(pretrained)
# add tensorboard to record train log
retinanet.training = True
writer = SummaryWriter('./log')
# writer.add_graph(retinanet, input_to_model=[images, labels])
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
retinanet.module.freeze_bn()
print('Num training images: {}'.format(len(dataset_train)))
for epoch_num in range(parser.epochs):
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet(
[data['img'].to(device), data['ann'].to(device)])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss))
writer.add_scalar('Loss/train', loss, iter_num)
writer.add_scalar('Loss/reg_loss', regression_loss, iter_num)
writer.add_scalar('Loss/cls_loss', classification_loss,
iter_num)
epoch_loss.append(float(loss))
if (iter_num + 1) % 1000 == 0:
print('Save model')
torch.save(
retinanet.module,
'COCO_retinanet_epoch{}_iter{}.pt'.format(
epoch_num, iter_num))
print(
'Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'
.format(epoch_num, iter_num, float(classification_loss),
float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
print('Evaluating dataset')
coco_eval.evaluate_coco(dataset_val, retinanet, writer)
scheduler.step(np.mean(epoch_loss))
torch.save(retinanet.module, 'COCO_retinanet_{}.pt'.format(epoch_num))
retinanet.eval()
torch.save(retinanet, 'model_final.pt'.format(epoch_num))
# load image
img = Image.open("../../data_set/flower_data/flower_photos/tulips/112428665_d8f3632f36_n.jpg")
plt.imshow(img)
# [N, C, H, W]
img = data_transform(img)
# expand batch dimension
img = torch.unsqueeze(img, dim=0)
# read class_indict
try:
json_file = open('./class_indices.json', 'r')
class_indict = json.load(json_file)
except Exception as e:
print(e)
exit(-1)
# create model
model = resnet34(num_classes=5)
# load model weights
model_weight_path = "./resNet34.pth"
model.load_state_dict(torch.load(model_weight_path))
model.eval()
with torch.no_grad():
# predict class
output = torch.squeeze(model(img))
predict = torch.softmax(output, dim=0)
predict_cla = torch.argmax(predict).numpy()
print(class_indict[str(predict_cla)], predict[predict_cla].numpy())
plt.show()
def main():
global NORM_MEAN, NORM_STD, coconut_model, train_history_dict
for arg in vars(args):
print(str(arg) + ': ' + str(getattr(args, arg)))
print('=' * 100)
# Build Model base on dataset and arc
num_classes = None
if args.model_type == 'food179':
num_classes = 179
NORM_MEAN = FOOD179_MEAN
NORM_STD = FOOD179_STD
elif args.model_type == 'nsfw':
num_classes = 5
NORM_MEAN = NSFW_MEAN
NORM_STD = NSFW_STD
else:
raise ('Not Implemented!')
if args.model_arc == 'resnet18':
coconut_model = model.resnet18(num_classes=num_classes,
zero_init_residual=True)
elif args.model_arc == 'resnet34':
coconut_model = model.resnet34(num_classes=num_classes,
zero_init_residual=True)
elif args.model_arc == 'resnet50':
coconut_model = model.resnet50(num_classes=num_classes,
zero_init_residual=True)
elif args.model_arc == 'resnet101':
coconut_model = model.resnet101(num_classes=num_classes,
zero_init_residual=True)
elif args.model_arc == 'resnet152':
coconut_model = model.resnet152(num_classes=num_classes,
zero_init_residual=True)
elif args.model_arc == 'mobilenet':
coconut_model = model.MobileNetV2(n_class=num_classes, input_size=256)
else:
raise ('Not Implemented!')
coconut_model = nn.DataParallel(coconut_model)
if args.cuda:
coconut_model = coconut_model.cuda()
torch.backends.benchmark = True
print("CUDA Enabled")
gpu_count = torch.cuda.device_count()
print('Total of %d GPU available' % (gpu_count))
args.train_batch_size = args.train_batch_size * gpu_count
args.test_batch_size = args.test_batch_size * gpu_count
print('args.train_batch_size: %d' % (args.train_batch_size))
print('args.test_batch_size: %d' % (args.test_batch_size))
model_parameters = filter(lambda p: p.requires_grad,
coconut_model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print('Total of %d parameters' % (params))
# Build Training
start_epoch = 0
best_acc = 0
optimizer = None
scheduler = None
milestones = [50, 150, 250]
if args.train_optimizer == 'sgd':
optimizer = optim.SGD(coconut_model.parameters(),
lr=args.lr,
momentum=0.9,
nesterov=True,
weight_decay=args.l2_reg)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.1)
elif args.train_optimizer == 'adam':
optimizer = optim.Adam(coconut_model.parameters(), lr=args.lr)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=milestones,
gamma=0.1)
elif args.train_optimizer == 'adabound':
optimizer = adabound.AdaBound(coconut_model.parameters(),
lr=1e-3,
final_lr=0.1)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=150,
gamma=0.1,
last_epoch=-1)
global_steps = 0
if not args.start_from_begining:
filename = args.model_checkpoint_path
if args.load_gpu_model_on_cpu:
checkpoint = torch.load(filename,
map_location=lambda storage, loc: storage)
else:
checkpoint = torch.load(filename)
coconut_model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['model_optimizer'])
best_acc = checkpoint['best_acc']
train_history_dict = checkpoint['train_history_dict']
scheduler.optimizer = optimizer # Not sure if this actually works
start_epoch = checkpoint['epoch']
global_steps = checkpoint['global_steps']
print(filename + ' loaded!')
data_loaders = load_datasets()
train_ops(start_epoch=start_epoch,
model=coconut_model,
optimizer=optimizer,
scheduler=scheduler,
data_loaders=data_loaders,
best_acc=best_acc,
global_steps=global_steps)
init.normal(m.weight.data, 1.0, 0.02)
init.constant(m.bias.data, 0.0)
#m.weight.data.normal_(1.0, 0.02)
#m.bias.data.fill_(0)
elif classname.find('Linear') != -1:
init.kaiming_normal(m.weight.data, a=0, mode='fan_in')
init.constant(m.bias.data, 0.0)
#m.weight.data.normal_(0.0, 0.02)
def compute_accuracy(x, y):
_, predicted = torch.max(x, dim=1)
correct = (predicted == y).float()
accuracy = torch.mean(correct) * 100.0
return accuracy
Resnet34 = resnet34(num_classes=opt.out_class)
Resnet34.apply(weights_init)
criterion = nn.CrossEntropyLoss()
if opt.cuda:
Resnet34.cuda()
criterion.cuda()
if ngpu>1:
Resnet34 = nn.DataParallel(Resnet34)
if opt.Resnet34 != '':
Resnet34.load_state_dict(torch.load(opt.Resnet34))
def main_fun(rank, world_size, args):
if torch.cuda.is_available() is False:
raise EnvironmentError("not find GPU device for training.")
# 初始化各进程环境 start
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355"
args.rank = rank
args.world_size = world_size
args.gpu = rank
args.distributed = True
torch.cuda.set_device(args.gpu)
args.dist_backend = 'nccl'
print('| distributed init (rank {}): {}'.format(args.rank, args.dist_url),
flush=True)
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
dist.barrier()
# 初始化各进程环境 end
rank = args.rank
device = torch.device(args.device)
batch_size = args.batch_size
num_classes = args.num_classes
weights_path = args.weights
args.lr *= args.world_size # 学习率要根据并行GPU的数量进行倍增
if rank == 0: # 在第一个进程中打印信息，并实例化tensorboard
print(args)
print(
'Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/'
)
tb_writer = SummaryWriter()
if os.path.exists("./weights") is False:
os.makedirs("./weights")
train_images_path, train_images_label, val_images_path, val_images_label = read_split_data(
args.data_path)
data_transform = {
"train":
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
"val":
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
}
# 实例化训练数据集
train_data_set = MyDataSet(images_path=train_images_path,
images_class=train_images_label,
transform=data_transform["train"])
# 实例化验证数据集
val_data_set = MyDataSet(images_path=val_images_path,
images_class=val_images_label,
transform=data_transform["val"])
# 给每个rank对应的进程分配训练的样本索引
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data_set)
val_sampler = torch.utils.data.distributed.DistributedSampler(val_data_set)
# 将样本索引每batch_size个元素组成一个list
train_batch_sampler = torch.utils.data.BatchSampler(train_sampler,
batch_size,
drop_last=True)
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
if rank == 0:
print('Using {} dataloader workers every process'.format(nw))
train_loader = torch.utils.data.DataLoader(
train_data_set,
batch_sampler=train_batch_sampler,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
val_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
sampler=val_sampler,
pin_memory=True,
num_workers=nw,
collate_fn=val_data_set.collate_fn)
# 实例化模型
model = resnet34(num_classes=num_classes).to(device)
# 如果存在预训练权重则载入
if os.path.exists(weights_path):
weights_dict = torch.load(weights_path, map_location=device)
load_weights_dict = {
k: v
for k, v in weights_dict.items()
if model.state_dict()[k].numel() == v.numel()
}
model.load_state_dict(load_weights_dict, strict=False)
else:
checkpoint_path = os.path.join(tempfile.gettempdir(),
"initial_weights.pt")
# 如果不存在预训练权重，需要将第一个进程中的权重保存，然后其他进程载入，保持初始化权重一致
if rank == 0:
torch.save(model.state_dict(), checkpoint_path)
dist.barrier()
# 这里注意，一定要指定map_location参数，否则会导致第一块GPU占用更多资源
model.load_state_dict(torch.load(checkpoint_path, map_location=device))
# 是否冻结权重
if args.freeze_layers:
for name, para in model.named_parameters():
# 除最后的全连接层外，其他权重全部冻结
if "fc" not in name:
para.requires_grad_(False)
else:
# 只有训练带有BN结构的网络时使用SyncBatchNorm采用意义
if args.syncBN:
# 使用SyncBatchNorm后训练会更耗时
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(
device)
# 转为DDP模型
model = torch.nn.parallel.DistributedDataParallel(model,
device_ids=[args.gpu])
# optimizer
pg = [p for p in model.parameters() if p.requires_grad]
optimizer = optim.SGD(pg, lr=args.lr, momentum=0.9, weight_decay=0.005)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: ((1 + math.cos(x * math.pi / args.epochs)) / 2) * (
1 - args.lrf) + args.lrf # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
for epoch in range(args.epochs):
train_sampler.set_epoch(epoch)
mean_loss = train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_loader,
device=device,
epoch=epoch)
scheduler.step()
sum_num = evaluate(model=model, data_loader=val_loader, device=device)
acc = sum_num / val_sampler.total_size
if rank == 0:
print("[epoch {}] accuracy: {}".format(epoch, round(acc, 3)))
tags = ["loss", "accuracy", "learning_rate"]
tb_writer.add_scalar(tags[0], mean_loss, epoch)
tb_writer.add_scalar(tags[1], acc, epoch)
tb_writer.add_scalar(tags[2], optimizer.param_groups[0]["lr"],
epoch)
torch.save(model.module.state_dict(),
"./weights/model-{}.pth".format(epoch))
# 删除临时缓存文件
if rank == 0:
if os.path.exists(checkpoint_path) is True:
os.remove(checkpoint_path)
cleanup()
def main(args=None):
parser = argparse.ArgumentParser(
description="Simple training script for training a RetinaNet network.")
parser.add_argument(
"--dataset", help="Dataset type, must be one of csv or coco or ycb.")
parser.add_argument("--path", help="Path to dataset directory")
parser.add_argument(
"--csv_train",
help="Path to file containing training annotations (see readme)")
parser.add_argument("--csv_classes",
help="Path to file containing class list (see readme)")
parser.add_argument("--csv_val",
help="Path to file containing validation annotations "
"(optional, see readme)")
parser.add_argument(
"--depth",
help="Resnet depth, must be one of 18, 34, 50, 101, 152",
type=int,
default=50)
parser.add_argument("--epochs",
help="Number of epochs",
type=int,
default=100)
parser.add_argument("--evaluate_every", default=20, type=int)
parser.add_argument("--print_every", default=20, type=int)
parser.add_argument('--distributed',
action="store_true",
help='Run model in distributed mode with DataParallel')
parser = parser.parse_args(args)
# Create the data loaders
if parser.dataset == "coco":
if parser.path is None:
raise ValueError(
"Must provide --path when training on non-CSV datasets")
dataset_train = CocoDataset(parser.path,
ann_file="instances_train2014.json",
set_name="train2014",
transform=transforms.Compose([
Normalizer(),
Augmenter(),
Resizer(min_side=512, max_side=512)
]))
dataset_val = CocoDataset(parser.path,
ann_file="instances_val2014.cars.json",
set_name="val2014",
transform=transforms.Compose(
[Normalizer(), Resizer()]))
elif parser.dataset == "ycb":
dataset_train = YCBDataset(parser.path,
"image_sets/train.txt",
transform=transforms.Compose([
Normalizer(),
Augmenter(),
Resizer(min_side=512, max_side=512)
]),
train=True)
dataset_val = YCBDataset(parser.path,
"image_sets/val.txt",
transform=transforms.Compose(
[Normalizer(), Resizer()]),
train=False)
elif parser.dataset == "csv":
if parser.csv_train is None:
raise ValueError("Must provide --csv_train when training on COCO,")
if parser.csv_classes is None:
raise ValueError(
"Must provide --csv_classes when training on COCO,")
dataset_train = CSVDataset(train_file=parser.csv_train,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
if parser.csv_val is None:
dataset_val = None
print("No validation annotations provided.")
else:
dataset_val = CSVDataset(train_file=parser.csv_val,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Resizer()]))
else:
raise ValueError(
"Dataset type not understood (must be csv or coco), exiting.")
sampler = AspectRatioBasedSampler(dataset_train,
batch_size=12,
drop_last=False)
dataloader_train = DataLoader(dataset_train,
num_workers=8,
collate_fn=collater,
batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val,
batch_size=1,
drop_last=False)
dataloader_val = DataLoader(dataset_val,
num_workers=4,
collate_fn=collater,
batch_sampler=sampler_val)
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
"Unsupported model depth, must be one of 18, 34, 50, 101, 152")
print("CUDA available: {}".format(torch.cuda.is_available()))
if torch.cuda.is_available():
device = "cuda"
else:
device = "cpu"
retinanet = retinanet.to(device)
if parser.distributed:
retinanet = torch.nn.DataParallel(retinanet)
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
loss_hist = collections.deque(maxlen=500)
print("Num training images: {}".format(len(dataset_train)))
best_mean_avg_prec = 0.0
for epoch_num in range(parser.epochs):
retinanet.train()
retinanet.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet(
[data["img"].to(device).float(), data["annot"]])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss.item()))
epoch_loss.append(float(loss.item()))
if parser.print_every % iter_num == 0:
print("Epoch: {} | Iteration: {}/{} | "
"Classification loss: {:1.5f} | "
"Regression loss: {:1.5f} | "
"Running loss: {:1.5f}".format(
epoch_num, iter_num, len(dataloader_train),
float(classification_loss),
float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
if ((epoch_num + 1) % parser.evaluate_every
== 0) or epoch_num + 1 == parser.epochs:
mAP = 0.0
if parser.dataset == "coco":
print("Evaluating dataset")
mAP = coco_eval.evaluate_coco(dataset_val, retinanet)
else:
print("Evaluating dataset")
AP = eval.evaluate(dataset_val, retinanet)
mAP = np.asarray([x[0] for x in AP.values()]).mean()
print("Val set mAP: ", mAP)
if mAP > best_mean_avg_prec:
best_mean_avg_prec = mAP
torch.save(
retinanet.state_dict(),
"{}_retinanet_best_mean_ap_{}.pt".format(
parser.dataset, epoch_num))
scheduler.step(np.mean(epoch_loss))
retinanet.eval()
torch.save(retinanet.state_dict(), "retinanet_model_final.pt")
def main(args=None):
parser = argparse.ArgumentParser(
description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset',
help='Dataset type, must be one of csv or coco.')
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument(
'--csv_train',
help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes',
help='Path to file containing class list (see readme)')
parser.add_argument(
'--csv_val',
help=
'Path to file containing validation annotations (optional, see readme)'
)
parser.add_argument(
'--depth',
help='Resnet depth, must be one of 18, 34, 50, 101, 152',
type=int,
default=50)
parser.add_argument('--epochs',
help='Number of epochs',
type=int,
default=100)
parser.add_argument('--optimizer',
help='[SGD | Adam]',
type=str,
default='SGD')
parser.add_argument('--model', help='Path to model (.pt) file.')
parser = parser.parse_args(args)
# Create the data loaders
print("\n[Phase 1]: Creating DataLoader for {} dataset".format(
parser.dataset))
if parser.dataset == 'coco':
if parser.coco_path is None:
raise ValueError('Must provide --coco_path when training on COCO,')
dataset_train = CocoDataset(parser.coco_path,
set_name='train2014',
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
dataset_val = CocoDataset(parser.coco_path,
set_name='val2014',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
elif parser.dataset == 'csv':
if parser.csv_train is None:
raise ValueError('Must provide --csv_train when training on COCO,')
if parser.csv_classes is None:
raise ValueError(
'Must provide --csv_classes when training on COCO,')
dataset_train = CSVDataset(train_file=parser.csv_train,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
if parser.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=parser.csv_val,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Resizer()]))
else:
raise ValueError(
'Dataset type not understood (must be csv or coco), exiting.')
sampler = AspectRatioBasedSampler(dataset_train,
batch_size=8,
drop_last=False)
dataloader_train = DataLoader(dataset_train,
num_workers=8,
collate_fn=collater,
batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val,
batch_size=16,
drop_last=False)
dataloader_val = DataLoader(dataset_val,
num_workers=8,
collate_fn=collater,
batch_sampler=sampler_val)
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
print('| Num training images: {}'.format(len(dataset_train)))
print('| Num test images : {}'.format(len(dataset_val)))
print("\n[Phase 2]: Preparing RetinaNet Detection Model...")
use_gpu = torch.cuda.is_available()
if use_gpu:
device = torch.device('cuda')
retinanet = retinanet.to(device)
retinanet = torch.nn.DataParallel(retinanet,
device_ids=range(
torch.cuda.device_count()))
print("| Using %d GPUs for Train/Validation!" % torch.cuda.device_count())
retinanet.training = True
if parser.optimizer == 'Adam':
optimizer = optim.Adam(retinanet.parameters(),
lr=1e-5) # not mentioned
print("| Adam Optimizer with Learning Rate = {}".format(1e-5))
elif parser.optimizer == 'SGD':
optimizer = optim.SGD(retinanet.parameters(),
lr=1e-2,
momentum=0.9,
weight_decay=1e-4)
print("| SGD Optimizer with Learning Rate = {}".format(1e-2))
else:
raise ValueError('Unsupported Optimizer, must be one of [SGD | Adam]')
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
retinanet.module.freeze_bn(
) # Freeze the BN parameters to ImageNet configuration
# Check if there is a 'checkpoints' path
if not osp.exists('./checkpoints/'):
os.makedirs('./checkpoints/')
print("\n[Phase 3]: Training Model on {} dataset...".format(
parser.dataset))
for epoch_num in range(parser.epochs):
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet(
[data['img'].to(device), data['annot']])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.001)
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
sys.stdout.write('\r')
sys.stdout.write(
'| Epoch: {} | Iteration: {}/{} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'
.format(epoch_num + 1, iter_num + 1, len(dataloader_train),
float(classification_loss), float(regression_loss),
np.mean(loss_hist)))
sys.stdout.flush()
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
print("\n| Saving current best model at epoch {}...".format(epoch_num +
1))
torch.save(
retinanet.state_dict(),
'./checkpoints/{}_retinanet_{}.pt'.format(parser.dataset,
epoch_num + 1))
if parser.dataset == 'coco':
#print('Evaluating dataset')
coco_eval.evaluate_coco(dataset_val, retinanet, device)
elif parser.dataset == 'csv' and parser.csv_val is not None:
#print('Evaluating dataset')
mAP = csv_eval.evaluate(dataset_val, retinanet, device)
scheduler.step(np.mean(epoch_loss))
retinanet.eval()
torch.save(retinanet.state_dict(), './checkpoints/model_final.pt')
def main(args=None):
parser = argparse.ArgumentParser(
description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset',
help='Dataset type, must be one of csv or coco.',
default="csv")
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument(
'--csv_train',
help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes',
help='Path to file containing class list (see readme)',
default="binary_class.csv")
parser.add_argument(
'--csv_val',
help=
'Path to file containing validation annotations (optional, see readme)'
)
parser.add_argument(
'--depth',
help='Resnet depth, must be one of 18, 34, 50, 101, 152',
type=int,
default=18)
parser.add_argument('--epochs',
help='Number of epochs',
type=int,
default=500)
parser.add_argument('--epochs_only_det',
help='Number of epochs to train detection part',
type=int,
default=1)
parser.add_argument('--max_epochs_no_improvement',
help='Max epochs without improvement',
type=int,
default=100)
parser.add_argument('--pretrained_model',
help='Path of .pt file with pretrained model',
default='esposallescsv_retinanet_0.pt')
parser.add_argument('--model_out',
help='Path of .pt file with trained model to save',
default='trained')
parser.add_argument('--score_threshold',
help='Score above which boxes are kept',
type=float,
default=0.5)
parser.add_argument('--nms_threshold',
help='Score above which boxes are kept',
type=float,
default=0.2)
parser.add_argument('--max_boxes',
help='Max boxes to be fed to recognition',
default=95)
parser.add_argument('--seg_level',
help='[line, word], to choose anchor aspect ratio',
default='word')
parser.add_argument(
'--early_stop_crit',
help='Early stop criterion, detection (map) or transcription (cer)',
default='cer')
parser.add_argument('--max_iters_epoch',
help='Max steps per epoch (for debugging)',
default=1000000)
parser.add_argument('--train_htr',
help='Train recognition or not',
default='True')
parser.add_argument('--train_det',
help='Train detection or not',
default='True')
parser.add_argument(
'--binary_classifier',
help=
'Wether to use classification branch as binary or not, multiclass instead.',
default='False')
parser.add_argument(
'--htr_gt_box',
help='Train recognition branch with box gt (for debugging)',
default='False')
parser.add_argument(
'--ner_branch',
help='Train named entity recognition with separate branch',
default='False')
parser = parser.parse_args(args)
if parser.dataset == 'csv':
if parser.csv_train is None:
raise ValueError('Must provide --csv_train')
dataset_name = parser.csv_train.split("/")[-2]
dataset_train = CSVDataset(train_file=parser.csv_train,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
if parser.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=parser.csv_val,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Resizer()]))
else:
raise ValueError(
'Dataset type not understood (must be csv or coco), exiting.')
# Files for training log
experiment_id = str(time.time()).split('.')[0]
valid_cer_f = open('trained_models/' + parser.model_out + 'log.txt', 'w')
for arg in vars(parser):
if getattr(parser, arg) is not None:
valid_cer_f.write(
str(arg) + ' ' + str(getattr(parser, arg)) + '\n')
current_commit = subprocess.check_output(['git', 'rev-parse', 'HEAD'])
valid_cer_f.write(str(current_commit))
valid_cer_f.write(
"epoch_num cer best cer mAP best mAP time\n")
valid_cer_f.close()
sampler = AspectRatioBasedSampler(dataset_train,
batch_size=1,
drop_last=False)
dataloader_train = DataLoader(dataset_train,
num_workers=3,
collate_fn=collater,
batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val,
batch_size=1,
drop_last=False)
dataloader_val = DataLoader(dataset_val,
num_workers=0,
collate_fn=collater,
batch_sampler=sampler_val)
if not os.path.exists('trained_models'):
os.mkdir('trained_models')
# Create the model
train_htr = parser.train_htr == 'True'
htr_gt_box = parser.htr_gt_box == 'True'
ner_branch = parser.ner_branch == 'True'
binary_classifier = parser.binary_classifier == 'True'
torch.backends.cudnn.benchmark = False
alphabet = dataset_train.alphabet
if os.path.exists(parser.pretrained_model):
retinanet = torch.load(parser.pretrained_model)
retinanet.classificationModel = ClassificationModel(
num_features_in=256,
num_anchors=retinanet.anchors.num_anchors,
num_classes=dataset_train.num_classes())
if ner_branch:
retinanet.nerModel = NERModel(
feature_size=256,
pool_h=retinanet.pool_h,
n_classes=dataset_train.num_classes(),
pool_w=retinanet.pool_w)
else:
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True,
max_boxes=int(parser.max_boxes),
score_threshold=float(
parser.score_threshold),
seg_level=parser.seg_level,
alphabet=alphabet,
train_htr=train_htr,
htr_gt_box=htr_gt_box,
ner_branch=ner_branch,
binary_classifier=binary_classifier)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True,
max_boxes=int(parser.max_boxes),
score_threshold=float(
parser.score_threshold),
seg_level=parser.seg_level,
alphabet=alphabet,
train_htr=train_htr,
htr_gt_box=htr_gt_box)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(
num_classes=dataset_train.num_classes(), pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(
num_classes=dataset_train.num_classes(), pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
train_htr = parser.train_htr == 'True'
train_det = parser.train_det == 'True'
retinanet.htr_gt_box = parser.htr_gt_box == 'True'
retinanet.train_htr = train_htr
retinanet.epochs_only_det = parser.epochs_only_det
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-4)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=50,
verbose=True)
loss_hist = collections.deque(maxlen=500)
ctc = CTCLoss()
retinanet.train()
retinanet.module.freeze_bn()
best_cer = 1000
best_map = 0
epochs_no_improvement = 0
verbose_each = 20
optimize_each = 1
objective = 100
best_objective = 10000
print(('Num training images: {}'.format(len(dataset_train))))
for epoch_num in range(parser.epochs):
cers = []
retinanet.training = True
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
if iter_num > int(parser.max_iters_epoch): break
try:
if iter_num % optimize_each == 0:
optimizer.zero_grad()
(classification_loss, regression_loss, ctc_loss,
ner_loss) = retinanet([
data['img'].cuda().float(), data['annot'], ctc, epoch_num
])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
if train_det:
if train_htr:
loss = ctc_loss + classification_loss + regression_loss + ner_loss
else:
loss = classification_loss + regression_loss + ner_loss
elif train_htr:
loss = ctc_loss
else:
continue
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
if iter_num % verbose_each == 0:
print((
'Epoch: {} | Step: {} |Classification loss: {:1.5f} | Regression loss: {:1.5f} | CTC loss: {:1.5f} | NER loss: {:1.5f} | Running loss: {:1.5f} | Total loss: {:1.5f}\r'
.format(epoch_num, iter_num,
float(classification_loss),
float(regression_loss), float(ctc_loss),
float(ner_loss), np.mean(loss_hist),
float(loss), "\r")))
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
torch.cuda.empty_cache()
except Exception as e:
print(e)
continue
if parser.dataset == 'csv' and parser.csv_val is not None and train_det:
print('Evaluating dataset')
mAP, text_mAP, current_cer = csv_eval.evaluate(
dataset_val, retinanet, score_threshold=parser.score_threshold)
#text_mAP,_ = csv_eval_binary_map.evaluate(dataset_val, retinanet,score_threshold=parser.score_threshold)
objective = current_cer * (1 - mAP)
retinanet.eval()
retinanet.training = False
retinanet.score_threshold = float(parser.score_threshold)
'''for idx,data in enumerate(dataloader_val):
if idx>int(parser.max_iters_epoch): break
print("Eval CER on validation set:",idx,"/",len(dataset_val),"\r")
image_name = dataset_val.image_names[idx].split('/')[-1].split('.')[-2]
#generate_pagexml(image_name,data,retinanet,parser.score_threshold,parser.nms_threshold,dataset_val)
text_gt =".".join(dataset_val.image_names[idx].split('.')[:-1])+'.txt'
f =open(text_gt,'r')
text_gt_lines=f.readlines()[0]
transcript_pred = get_transcript(image_name,data,retinanet,float(parser.score_threshold),float(parser.nms_threshold),dataset_val,alphabet)
cers.append(float(editdistance.eval(transcript_pred,text_gt_lines))/len(text_gt_lines))'''
t = str(time.time()).split('.')[0]
valid_cer_f.close()
#print("GT",text_gt_lines)
#print("PREDS SAMPLE:",transcript_pred)
if parser.early_stop_crit == 'cer':
if float(objective) < float(
best_objective): #float(current_cer)<float(best_cer):
best_cer = current_cer
best_objective = objective
epochs_no_improvement = 0
torch.save(
retinanet.module, 'trained_models/' + parser.model_out +
'{}_retinanet.pt'.format(parser.dataset))
else:
epochs_no_improvement += 1
if mAP > best_map:
best_map = mAP
elif parser.early_stop_crit == 'map':
if mAP > best_map:
best_map = mAP
epochs_no_improvement = 0
torch.save(
retinanet.module, 'trained_models/' + parser.model_out +
'{}_retinanet.pt'.format(parser.dataset))
else:
epochs_no_improvement += 1
if float(current_cer) < float(best_cer):
best_cer = current_cer
if train_det:
print(epoch_num, "mAP: ", mAP, " best mAP", best_map)
if train_htr:
print("VALID CER:", current_cer, "best CER", best_cer)
print("Epochs no improvement:", epochs_no_improvement)
valid_cer_f = open('trained_models/' + parser.model_out + 'log.txt',
'a')
valid_cer_f.write(
str(epoch_num) + " " + str(current_cer) + " " + str(best_cer) +
' ' + str(mAP) + ' ' + str(best_map) + ' ' + str(text_mAP) + '\n')
if epochs_no_improvement > 3:
for param_group in optimizer.param_groups:
if param_group['lr'] > 10e-5:
param_group['lr'] *= 0.1
if epochs_no_improvement >= parser.max_epochs_no_improvement:
print("TRAINING FINISHED AT EPOCH", epoch_num, ".")
sys.exit()
scheduler.step(np.mean(epoch_loss))
torch.cuda.empty_cache()
retinanet.eval()
def main(args):
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
print(args)
print(
'Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/'
)
tb_writer = SummaryWriter()
if os.path.exists("./weights") is False:
os.makedirs("./weights")
train_info, val_info, num_classes = read_split_data(args.data_path)
train_images_path, train_images_label = train_info
val_images_path, val_images_label = val_info
# check num_classes
assert args.num_classes == num_classes, "dataset num_classes: {}, input {}".format(
args.num_classes, num_classes)
data_transform = {
"train":
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
"val":
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
}
# 实例化训练数据集
train_data_set = MyDataSet(images_path=train_images_path,
images_class=train_images_label,
transform=data_transform["train"])
# 实例化验证数据集
val_data_set = MyDataSet(images_path=val_images_path,
images_class=val_images_label,
transform=data_transform["val"])
batch_size = args.batch_size
nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0,
8]) # number of workers
print('Using {} dataloader workers every process'.format(nw))
train_loader = torch.utils.data.DataLoader(
train_data_set,
batch_size=batch_size,
shuffle=True,
pin_memory=True,
num_workers=nw,
collate_fn=train_data_set.collate_fn)
val_loader = torch.utils.data.DataLoader(
val_data_set,
batch_size=batch_size,
shuffle=False,
pin_memory=True,
num_workers=nw,
collate_fn=val_data_set.collate_fn)
# 如果存在预训练权重则载入
model = resnet34(num_classes=args.num_classes).to(device)
if args.weights != "":
if os.path.exists(args.weights):
weights_dict = torch.load(args.weights, map_location=device)
load_weights_dict = {
k: v
for k, v in weights_dict.items()
if model.state_dict()[k].numel() == v.numel()
}
print(model.load_state_dict(load_weights_dict, strict=False))
else:
raise FileNotFoundError("not found weights file: {}".format(
args.weights))
# 是否冻结权重
if args.freeze_layers:
for name, para in model.named_parameters():
# 除最后的全连接层外，其他权重全部冻结
if "fc" not in name:
para.requires_grad_(False)
pg = [p for p in model.parameters() if p.requires_grad]
optimizer = optim.SGD(pg, lr=args.lr, momentum=0.9, weight_decay=0.005)
# Scheduler https://arxiv.org/pdf/1812.01187.pdf
lf = lambda x: ((1 + math.cos(x * math.pi / args.epochs)) / 2) * (
1 - args.lrf) + args.lrf # cosine
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)
for epoch in range(args.epochs):
# train
mean_loss = train_one_epoch(model=model,
optimizer=optimizer,
data_loader=train_loader,
device=device,
epoch=epoch)
scheduler.step()
# validate
sum_num = evaluate(model=model, data_loader=val_loader, device=device)
acc = sum_num / len(val_data_set)
print("[epoch {}] accuracy: {}".format(epoch, round(acc, 3)))
tags = ["loss", "accuracy", "learning_rate"]
tb_writer.add_scalar(tags[0], mean_loss, epoch)
tb_writer.add_scalar(tags[1], acc, epoch)
tb_writer.add_scalar(tags[2], optimizer.param_groups[0]["lr"], epoch)
torch.save(model.state_dict(), "./weights/model-{}.pth".format(epoch))
def main(args=None):
parser = argparse.ArgumentParser(
description='Simple training script for training a RetinaNet network.')
parser.add_argument('--dataset',
help='Dataset type, must be one of csv or coco.')
parser.add_argument('--coco_path', help='Path to COCO directory')
parser.add_argument(
'--csv_train',
help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes',
help='Path to file containing class list (see readme)')
parser.add_argument(
'--csv_val',
help=
'Path to file containing validation annotations (optional, see readme)'
)
parser.add_argument(
'--depth',
help='Resnet depth, must be one of 18, 34, 50, 101, 152',
type=int,
default=50)
parser.add_argument('--epochs',
help='Number of epochs',
type=int,
default=100)
parser = parser.parse_args(args)
# Create the data loaders
if parser.dataset == 'coco':
if parser.coco_path is None:
raise ValueError('Must provide --coco_path when training on COCO,')
dataset_train = CocoDataset(parser.coco_path,
set_name='train2017',
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
dataset_val = CocoDataset(parser.coco_path,
set_name='val2017',
transform=transforms.Compose(
[Normalizer(), Resizer()]))
elif parser.dataset == 'csv':
if parser.csv_train is None:
raise ValueError('Must provide --csv_train when training on COCO,')
if parser.csv_classes is None:
raise ValueError(
'Must provide --csv_classes when training on COCO,')
dataset_train = CSVDataset(train_file=parser.csv_train,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Augmenter(),
Resizer()]))
if parser.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=parser.csv_val,
class_list=parser.csv_classes,
transform=transforms.Compose(
[Normalizer(),
Resizer()]))
else:
raise ValueError(
'Dataset type not understood (must be csv or coco), exiting.')
sampler = AspectRatioBasedSampler(dataset_train,
batch_size=2,
drop_last=False)
dataloader_train = DataLoader(dataset_train,
num_workers=3,
collate_fn=collater,
batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val,
batch_size=1,
drop_last=False)
dataloader_val = DataLoader(dataset_val,
num_workers=3,
collate_fn=collater,
batch_sampler=sampler_val)
# Create the model
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
retinanet.module.freeze_bn()
print('Num training images: {}'.format(len(dataset_train)))
for epoch_num in range(parser.epochs):
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
classification_loss, regression_loss = retinanet(
[data['img'].cuda().float(), data['annot']])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
print(
'Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'
.format(epoch_num, iter_num, float(classification_loss),
float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
if parser.dataset == 'coco':
print('Evaluating dataset')
coco_eval.evaluate_coco(dataset_val, retinanet)
elif parser.dataset == 'csv' and parser.csv_val is not None:
print('Evaluating dataset')
mAP = csv_eval.evaluate(dataset_val, retinanet)
scheduler.step(np.mean(epoch_loss))
torch.save(
retinanet.module,
'{}_retinanet_dilation_{}.pt'.format(parser.dataset, epoch_num))
retinanet.eval()
torch.save(retinanet, 'model_final_dilation.pt'.format(epoch_num))
def main(config):
# set seed for reproducibility
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed(0)
# create folder for model
newpath = './models/' + config.model_date
if config.save_model:
os.makedirs(newpath)
# Create the data loaders
if config.csv_train is None:
raise ValueError('Must provide --csv_train when training on csv,')
if config.csv_classes is None:
raise ValueError('Must provide --csv_classes when training on csv,')
train_dataset = datasets.ImageFolder(os.path.join(config.data_dir,
'train'))
dataset_train = GetDataset(train_file=config.csv_train,
class_list=config.csv_classes,
transform=transforms.Compose(
[Augmenter(), Resizer()]),
dataset=train_dataset,
seed=0)
dataloader_train = DataLoader(dataset_train,
batch_size=config.batch_size,
shuffle=True,
num_workers=1,
collate_fn=collater)
if config.csv_val is None:
dataset_val = None
print('No validation annotations provided.')
else:
valid_dataset = datasets.ImageFolder(
os.path.join(config.data_dir, 'valid'))
dataset_val = GetDataset(train_file=config.csv_val,
class_list=config.csv_classes,
transform=transforms.Compose([Resizer()]),
dataset=valid_dataset,
seed=0)
# Create the model
if config.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(),
pretrained=True)
elif config.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(),
pretrained=True)
elif config.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),
pretrained=True)
elif config.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(),
pretrained=True)
elif config.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(),
pretrained=True)
else:
raise ValueError(
'Unsupported model depth, must be one of 18, 34, 50, 101, 152')
if config.use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=1e-5)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
patience=3,
verbose=True)
retinanet.train()
retinanet.module.freeze_bn()
print('Num training images: {}'.format(len(dataset_train)))
best_valid_map = 0
counter = 0
batch_size = config.batch_size
for epoch_num in range(config.epochs):
print('\nEpoch: {}/{}'.format(epoch_num + 1, config.epochs))
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
train_batch_time = AverageMeter()
train_losses = AverageMeter()
tic = time.time()
with tqdm(total=len(dataset_train)) as pbar:
for iter_num, data in enumerate(dataloader_train):
# try:
optimizer.zero_grad()
siamese_loss, classification_loss, regression_loss = retinanet(
[
data['img'].cuda().float(), data['annot'],
data['pair'].cuda().float()
])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss + siamese_loss
if bool(loss == 0):
continue
loss.backward()
torch.nn.utils.clip_grad_norm_(retinanet.parameters(), 0.1)
optimizer.step()
epoch_loss.append(float(loss))
toc = time.time()
train_losses.update(float(loss), batch_size)
train_batch_time.update(toc - tic)
tic = time.time()
pbar.set_description(("{:.1f}s - loss: {:.3f}".format(
train_batch_time.val,
train_losses.val,
)))
pbar.update(batch_size)
del classification_loss
del regression_loss
del siamese_loss
# except Exception as e:
# print('Training error: ', e)
# continue
if config.csv_val is not None:
print('Evaluating dataset')
mAP, correct = eval_new.evaluate(dataset_val, retinanet)
# is_best = mAP[0][0] > best_valid_map
# best_valid_map = max(mAP[0][0], best_valid_map)
is_best = correct > best_valid_map
best_valid_map = max(correct, best_valid_map)
if is_best:
counter = 0
else:
counter += 1
if counter > 3:
print("[!] No improvement in a while, stopping training.")
break
scheduler.step(np.mean(epoch_loss))
if is_best and config.save_model:
torch.save(
retinanet.state_dict(),
'./models/{}/best_retinanet.pt'.format(config.model_date))
if config.save_model:
torch.save(
retinanet.state_dict(),
'./models/{}/{}_retinanet_{}.pt'.format(
config.model_date, config.depth, epoch_num))
msg = "train loss: {:.3f} - val map: {:.3f} - val acc: {:.3f}%"
print(
msg.format(train_losses.avg, mAP[0][0],
(100. * correct) / len(dataset_val)))
