def dump_model():
def get_bn_params(bn):
return [bn.running_mean,bn.running_var,bn.weight.data,bn.bias.data]
def get_block_params(block):
params=[block.conv1.weight.data]+get_bn_params(block.bn1)+[block.conv2.weight.data]+get_bn_params(block.bn2)
if block.downsample:
params+=[block.downsample._modules['0'].weight.data]+get_bn_params(block.downsample._modules['1'])
return params
def get_BN_block_params(block):
params=[block.conv1.weight.data]+get_bn_params(block.bn1)+[block.conv2.weight.data]+get_bn_params(block.bn2)+[block.conv3.weight.data]+get_bn_params(block.bn3)
if block.downsample:
params+=[block.downsample._modules['0'].weight.data]+get_bn_params(block.downsample._modules['1'])
return params
def get_layer_params(layer):
params=[]
for block in layer._modules:
params+=get_block_params(layer._modules[block])
return params
model = resnet.ResNet(resnet.BasicBlock, [2, 2, 2, 2])
# model.load('resnet18\\mergedresnet18.sd')
checkpoint = torch.load('resnet18\\checkpoint_blu35103.pth.tar')
#model.load_state_dict(moduledict_to_dict(checkpoint['state_dict']))
model.load_state_dict(checkpoint['state_dict'])
#model.quantize_from('quant_param89.data')
params=[model.conv1.weight.data]+get_bn_params(model.bn1)#(64,3,7,7)
params+=get_layer_params(model.layer1)
params+=get_layer_params(model.layer2)
params += get_layer_params(model.layer3)
params += get_layer_params(model.layer4)
params+=[model.fc.weight.data,model.fc.bias.data]
with open('checkpoint_blu35103.data','wb') as f:
for para_tensor in params:
para_np=para_tensor.numpy()
para_np.tofile(f)
return
def main():
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model = resnet.ResNet(resnet.Bottleneck, [3,8,36,3])
model.load_state_dict(torch.load('resnet152_best-148.ckpt', map_location=device))
loader = transforms.Compose([transforms.Resize((32, 32)), transforms.ToTensor()])
model.cuda()
model.eval()
with torch.no_grad():
# ResNet152 Python Pytorch Test...
'''
img_path = "9/" + i
image = Image.open(img_path)
image = image.convert('RGB')
image_tensor = loader(image).unsqueeze(0)
image_numpy = image_tensor.numpy()
image_tensor = image_tensor.cuda()
output = model(image_tensor)
_, predict = torch.max(output.data, 1)
print("predicted : ", predict, i)
'''
# Save Script Module
example = torch.rand(1,3,32,32).cuda()
traced_script_module = torch.jit.trace(model, example)
traced_script_module.save('script_module.pt')
def conv_validation():
#model = resnet.__dict__['resnet18'](pretrained=True)
model = resnet.ResNet(resnet.BasicBlock, [2, 2, 2, 2])
#model.load('resnet18\\mergedresnet18.sd')
checkpoint = torch.load('resnet18\\checkpoint_renorm89.pth.tar')
model.load_state_dict(moduledict_to_dict(checkpoint['state_dict']))
#model.load_state_dict(checkpoint['state_dict'])
model.quantize('quant_param89.data')
model.load_blu('3sigma.blu')
#normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])
normalize = transforms.Normalize(mean=[128/255,128/255,128/255], std=[0.226, 0.226, 0.226])
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder('..\\data\\imagenet\\val', transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=10, shuffle=False,
num_workers=4, pin_memory=True)
acc1_t = torch.tensor([0.0])
acc5_t = torch.tensor([0.0])
for i, (input, target) in enumerate(val_loader):
#input=(input-0.5)/0.226
output=model(input)
#output=model.forward_blu(input)
acc1, acc5 = accuracy(output, target, topk=(1, 5))
print('{}:acc1:{},acc5:{}'.format(i,acc1,acc5))
acc1_t+=acc1
acc5_t+=acc5
print('acc1:{}\nacc5:{}'.format(acc1_t/(i+1), acc5_t/(i+1)))
def unittest():
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess = tf.Session(config=sess_config)
model = resnet.ResNet(args)
sess.run(tf.global_variables_initializer())
def __init__( self, T, C ):
self.T = T
self.C = C
self.X = tf.placeholder( tf.float32, [None, self.T, 224, 224, 3] )
self.Y = tf.placeholder( tf.int32, [None, self.T] )
# The Joint annotations.
self.J = 18 # Using CMU Openpose
self.P = tf.placeholder( tf.float32, [None, self.T, 7, 7, self.J] )
self.LR = tf.placeholder( tf.float32 ) # Learning rate
self.lambda_l2 = tf.placeholder( tf.float32 ) # Regularization factor
self.phase = tf.placeholder( tf.bool ) # Training phase
self.BATCH = tf.shape( self.X )[0]
self.BT = self.BATCH * self.T
self.scope = "Model" # Train only variables in scope
self.l_action = 1.0
self.l_pose = 1.0
self.DIM_LSTM = 512 # Dimensionality of LSTM
self.DIM_ATT = 32 # Either 32 (Sub-JHMDB) or 128 (PennAction)
# Init ResNet
self.net = resnet.ResNet() # We are using ResNet as base DCN. Change here.
self.net.phase = self.phase
def train():
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess = tf.Session(config=sess_config)
model = resnet.ResNet(args)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(max_to_keep=100)
model.load_model(sess, saver)
logger.info('model loaded')
train_config = {'data_dir':args.train_dir}
dtrain = dataset.DataSet(train_config)
logger.info('dataset loaded')
for epoch in range(args.beg_epoch, args.epochs):
train_acc,train_loss = run_epoch(epoch, model, sess, dtrain)
saver.save(sess, os.path.join(args.model_dir, 'model.ckpt'), global_step=epoch)
logger.info('model of epoch {} saved'.format(epoch))
s = '[epoch {}]train-acc:{:.3} train-loss:{:.3}'
s = s.format(epoch,train_acc,train_loss)
logger.warning(s)
print(s)
#if epoch!=args.beg_epoch and (epoch-args.beg_epoch)%10==0:
if True:
valid_acc,valid_time = run_test(epoch, model, sess, 10, 20)
s = '[epoch {}]test-acc:{:.3} test-time:{}'.format(epoch, valid_acc, valid_time)
logger.warning(s)
print(s)
logger.info('train end')
def netMaker(args):
'''Build a network
Args:
args (Namespace): the namespace containing all the arguments required for training and building the network
Returns:
the built network
'''
if args.dataset == "MNIST" or args.dataset == "FAKENIST":
inSize = 28
inChan = 1
numClasses = 10
elif args.dataset == "CIFAR10":
inSize = 32
inChan = 3
numClasses = 10
elif args.dataset == "IMAGENET":
inSize = 224
inChan = 3
numClasses = 1000
else:
raise ValueError("Unknown dataset : {}".format(args.dataset))
if args.model == "cnn":
net = resnet.ResNet(resnet.BasicBlock, [1, 1, 1, 1],geom=False,inChan=inChan, width_per_group=args.dechan,\
strides=[2,2,2,2],firstConvKer=args.deker,inPlanes=args.dechan,num_classes=numClasses,conv=True)
elif args.model == "gnn":
net = GNN(inChan, args.chan_gnn, args.nb_lay_gnn, numClasses,
args.res_con_gnn, args.batch_norm_gnn, args.max_pool_pos,
args.max_pool_ker, False)
elif args.model == "gnn_resnet_stri":
net = resnet.ResNet(resnet.BasicBlock, [1, 1, 1, 1],geom=True,inChan=inChan, width_per_group=args.dechan,\
strides=[2,2,2,2],firstConvKer=args.deker,inPlanes=args.dechan,num_classes=numClasses,multiChan=False,conv=False)
elif args.model == "gnn_resnet":
net = resnet.ResNet(resnet.BasicBlock, [1, 1, 1, 1],geom=True,inChan=inChan, width_per_group=args.dechan,\
strides=[1,1,1,1],firstConvKer=args.deker,inPlanes=args.dechan,num_classes=numClasses,multiChan=False,conv=False)
elif args.model == "gnn_resnet_mc":
net = resnet.ResNet(resnet.BasicBlock, [1, 1, 1, 1],geom=True,inChan=inChan, width_per_group=args.dechan,\
strides=[1,1,1,1],firstConvKer=args.deker,inPlanes=args.dechan,num_classes=numClasses,multiChan=True,conv=False)
elif args.model == "gcnn_resnet":
net = resnet.ResNet(resnet.BasicBlock, [1, 1, 1, 1],geom=True,inChan=inChan, width_per_group=args.dechan,\
strides=[2,2,2,2],firstConvKer=args.deker,inPlanes=args.dechan,num_classes=numClasses,multiChan=False,conv=True)
else:
raise ValueError("Unknown model type : {}".format(args.model))
return net
def load_model(self):
# Load the model as defined by ` self.trained_model_prefix + "latest_model.pth" ` and already trained by train_model.py
nb_outputs = ut.nbOutputs(self.label_style, self.environment)
if self.model_type == 'dk_resnet18_CP':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = RewardFunctionHeadCartPole()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model_type == 'dk_resnet18_CP_weird':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = WeirdRewardFunctionHeadCartPole()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model_type == 'dk_resnet18_DT':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = RewardFunctionHeadDuckieTown()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model_type == 'resnet18':
net = models.resnet18(pretrained=False, num_classes=nb_outputs)
#### To use in case want the pretrained model: (remove num_classes as pretrained model only comes with original 1000 classes)
# dim_feats = net.fc.in_features # =1000
# net.fc = nn.Linear(dim_feats, nb_outputs)
elif self.model_type == 'resnet34':
net = resnet.resnet34(pretrained=False, num_classes=nb_outputs)
elif self.model_type == 'resnet50':
net = resnet.ResNet(resnet.Bottleneck, [3, 4, 6, 3],
num_classes=nb_outputs)
elif self.model_type == 'resnet101':
net = resnet.ResNet(resnet.Bottleneck, [3, 4, 23, 3],
num_classes=nb_outputs)
elif self.model_type == 'resnet152':
net = resnet.ResNet(resnet.Bottleneck, [3, 8, 36, 3],
num_classes=nb_outputs)
net.load_state_dict(torch.load(self.model_path))
print('Loaded model')
net.eval()
net = net.to(self.device)
return net
def __init__(self, input_channel,
dropout = None, inp = "log-Filterbank-DCT-26-13",
nodes = [16, 32, 64, 128], num_layers = 2):
super(resnet_extractor, self).__init__()
import resnet
self.model = resnet.ResNet(input_channel, resnet.ResidualBlock, nodes, num_layers,
dropout, inp)
self.out_features = self.model.out_features
def initialize_net(self):
nb_outputs = ut.nbOutputs(self.label_style, self.environment)
if self.model == 'dk_resnet18_CP':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = RewardFunctionHeadCartPole()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model == 'dk_resnet18_CP_weird':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = WeirdRewardFunctionHeadCartPole()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model == 'dk_resnet18_DT':
nb_outputs = 2 # FIXME: hard coded
reward_fn_head = RewardFunctionHeadDuckieTown()
net = RewardFunctionHeadModel(
models.resnet18(pretrained=False, num_classes=nb_outputs),
reward_fn_head)
elif self.model == 'resnet18':
net = models.resnet18(pretrained=False, num_classes=nb_outputs)
#### To use in case want the pretrained model: (remove num_classes as pretrained model only comes with original 1000 classes)
# dim_feats = net.fc.in_features # =1000
# net.fc = nn.Linear(dim_feats, nb_outputs)
elif self.model == 'resnet34':
net = models.resnet34(pretrained=False, num_classes=nb_outputs)
elif self.model == 'resnet50':
net = resnet.ResNet(resnet.Bottleneck, [3, 4, 6, 3],
num_classes=nb_outputs)
elif self.model == 'resnet101':
net = resnet.ResNet(resnet.Bottleneck, [3, 4, 23, 3],
num_classes=nb_outputs)
elif self.model == 'resnet152':
net = resnet.ResNet(resnet.Bottleneck, [3, 8, 36, 3],
num_classes=nb_outputs)
net = net.float()
net = net.to(self.device)
return net
def temperature_scaling():
print("Calibration")
state_dict = torch.load(args.save_path +
'resnet110_{0}.pth'.format(args.epoch))
valid_indices = torch.load(args.save_path + 'valid_indices.pth')
mean = [0.5071, 0.4867, 0.4408]
stdv = [0.2675, 0.2565, 0.2761]
test_transforms = tv.transforms.Compose([
tv.transforms.ToTensor(),
tv.transforms.Normalize(mean=mean, std=stdv),
])
valid_set = tv.datasets.CIFAR100(root='./cifar100_data/',
train=True,
transform=test_transforms,
download=False)
test_set = tv.datasets.CIFAR100(root='./cifar100_data/',
train=False,
transform=test_transforms,
download=False)
valid_loader = torch.utils.data.DataLoader(
valid_set,
batch_size=args.batch_size,
sampler=SubsetRandomSampler(valid_indices))
test_loader = torch.utils.data.DataLoader(test_set,
batch_size=100,
num_workers=4)
#valid_loader = valid_loader_pth
criterion = nn.CrossEntropyLoss()
network = resnet.ResNet(args.layer, 100).cuda()
network.load_state_dict(state_dict)
test_loss, test_acc, test_softmax, test_correct, test_y, test_loss_idv = test(
test_loader, network, criterion, args.epoch + 1, 'test')
for i in range(len(test_softmax)):
test_softmax[i] = test_softmax[i].item()
utlis.save_data('before_cali_correct', test_correct, args.save_path)
utlis.save_data('before_cali_softmax', test_softmax, args.save_path)
model = ModelWithTemperature(network)
model.set_temperature(valid_loader)
test_loss, test_acc, test_softmax, test_correct, test_y, test_loss_idv = test(
test_loader, model, criterion, args.epoch + 1, 'test')
for i in range(len(test_softmax)):
test_softmax[i] = test_softmax[i].item()
utlis.save_data('after_cali_correct', test_correct, args.save_path)
utlis.save_data('after_cali_softmax', test_softmax, args.save_path)
def __init__(self, is_train=True):
self.resnet = resnet.ResNet()
self.resnet.cuda()
if is_train:
self.optimizer = torch.optim.SGD(self.resnet.parameters(),
lr=C.LEARNING_RATE,
momentum=0.9,
weight_decay=1E-4)
self.policy_loss_fn = MultiLableCrossEntropy()
self.value_loss_fn = torch.nn.MSELoss()
def test():
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
net = resnet.ResNet(resnet.Bottleneck, [3, 8, 36, 3])
net = net.to(device)
net.eval()
with torch.no_grad():
img_tensor = torch.randn(1, 3, 32, 32)
img_tensor = img_tensor.to(device)
y = net(img_tensor)
print(y.size())
def main(root_image, root_model):
print("\nInput image ==> ", root_image)
print("Current model ===> ", root_model)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print("Run ", device)
## ResNet-50 : [3, 4, 6, 3]
## ResNet-101 : [3, 4, 23, 3]
## ResNet-152 : [3, 8, 36, 3]
model = resnet.ResNet(resnet.Bottleneck, [3, 8, 36, 3])
model.load_state_dict(torch.load(root_model))
model.to(device)
# Input Image Transforms
loader = transforms.Compose([
transforms.Resize((32, 32)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
# transforms.Normalize([0.5],[0.5]),
])
model.eval()
with torch.no_grad():
# label list : 67 class
label = os.listdir("./dataset/img/")
label.sort()
# to RGB
image = Image.open(root_image)
image = image.convert("RGB")
# to Tensor, Cuda
image_tensor = loader(image).unsqueeze(0)
image_tensor = image_tensor.to(device)
output = model(image_tensor)
# Predict
_, predicted = torch.max(output.data, 1)
# Calc Probability
s_max = torch.nn.Softmax(dim=0)
prob = s_max(output.squeeze(0))
prob_np = prob.cpu().numpy()
index = int(predicted.item())
predicted_name = label[index]
probability = round(prob_np[index] * 100, 2)
print("Predicted ===> name : {}, probability : {}".format(
predicted_name, probability))
'''
def __init__(self, trained_dataset='imagenet', depth=101, num_of_class=1000, path=None):
model = RN.ResNet(trained_dataset, depth, num_of_class)
model = torch.nn.DataParallel(model).cuda()
if path is not None:
print("=> loading checkpoint '{}'".format(path))
checkpoint = torch.load(path)
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}'".format(path))
model.module.avgpool = Identity()
model.module.fc = Identity()
self.model = model
def _build_model(self):
# Neural Net for Deep-Q learning Model
# model = Sequential()
# model.add(Dense(32, input_dim=self.state_size, activation='relu'))
# model.add(Dense(64, activation='relu'))
# model.add(Dense(self.action_size, activation='softmax'))
# model.compile(loss=self._huber_loss,
# optimizer=Adam(lr=self.learning_rate))
# resnet
# model = resnet.resnet18(self.action_size)
# model.build(input_shape=(None, 32, 32, 3))
# model.compile(loss=self._huber_loss,
# optimizer=Adam(lr=self.learning_rate))
# resnet go
model = resnet.ResNet(self.action_size)
model.build(input_shape=(None, 16, 16, 3))
model.compile(loss=self._huber_loss,
optimizer=Adam(lr=self.learning_rate))
return model
def test():
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess = tf.Session(config=sess_config)
model = resnet.ResNet(args)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(max_to_keep=100)
for epoch in range(0, args.beg_epoch):
args.model_path = os.path.join(args.model_dir, 'model.ckpt-{}'.format(epoch))
model.load_model(sess, saver)
valid_acc,valid_time = run_test(epoch, model, sess, 10, 20)
s = '[epoch {}]test-acc:{:.3} test-time:{}'.format(epoch, valid_acc, valid_time)
logger.warning(s)
print(s)
return
logger.info('test end')
def __init__(self, args):
super(Model, self).__init__()
self.conv_layer = resnet.ResNet(31)
self.recg = recog.Recogniton(args)
def main():
word_index_dict = json.load(open(args.word_index_json,
encoding="utf-8")) #一个汉字对应一个编号
num_classes = len(word_index_dict)
# xzy 加载label文件
image_label_dict = json.load(open(args.image_label_json)) #一个图片对应多个其标签的编号
# print(image_label_dict)
cudnn.benchmark = True
# xzy 加载模型
if args.model == 'densenet':
# 两千多种字符，multi-label分类
if use_gpu:
model = DenseNet121(num_classes).cuda()
else:
model = DenseNet121(num_classes)
elif args.model == 'resnet':
# resnet主要用于文字区域的segmentation以及object detection操作
if use_gpu:
model = resnet.ResNet(num_classes=num_classes, args=args).cuda()
else:
model = resnet.ResNet(num_classes=num_classes, args=args)
else:
return
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# model = torch.nn.DataParallel(model).cuda()
if use_gpu:
loss = Loss().cuda()
else:
loss = Loss()
# if args.resume:
if need_resume:
print("let's begin resume")
# state_dict = torch.load(args.resume)
state_dict = torch.load(resume_ckpt_path) #xzy
model.load_state_dict(state_dict['state_dict']) #xzy 断点训练、测试，都走这里。
best_f1score = state_dict['f1score']
start_epoch = state_dict['epoch'] + 1
print("already resume " + str(resume_ckpt_path.split("/")[-1]))
else:
print("no resume")
best_f1score = 0
if args.model == 'resnet':
start_epoch = 100
else:
start_epoch = 1
args.epoch = start_epoch
print('best_f1score', best_f1score)
#xzy 加载、划分数据集
# test_filelist = sorted(glob(os.path.join(args.data_dir,'test','*')))
# trainval_filelist = sorted(glob(os.path.join(args.data_dir,'train','*')))
test_filelist = sorted(glob(os.path.join(args.data_test_dir, '*')))
trainval_filelist = sorted(glob(os.path.join(args.data_dir, '*')))
print("训练数据集长度是" + str(len(trainval_filelist)))
print("测试数据集长度是" + str(len(test_filelist)))
# 两种输入size训练
# train_filelist1: 长宽比小于8:1的图片，经过padding后变成 64*512 的输入
# train_filelist2: 长宽比大于8:1的图片，经过padding,crop后变成 64*1024的输入
train_filelist1, train_filelist2 = [], []
# 黑名单，这些图片的label是有问题的
black_list = set(json.load(open(args.black_json))['black_list'])
# image_hw_ratio_dict = json.load(open(args.image_hw_ratio_json)) #xzy
for f in trainval_filelist:
image = f.split('/')[-1]
if image in black_list:
continue
# r = image_hw_ratio_dict[image]
# if r == 0:
if True: #xzy 不用考虑判断是否1：8从而分散数据集，进行不同resize
train_filelist1.append(f)
else:
train_filelist2.append(f)
# train_val_filelist = train_filelist1 + train_filelist2
train_val_filelist = train_filelist1
val_filelist = train_filelist1[-2048:]
train_filelist1 = train_filelist1[:-2048]
# train_filelist2 = train_filelist2` ` #取消train_dataset2数据集
# image_size = [512, 64] 32的16倍和2倍
image_size = [288, 64] #xzy 32的9倍和2倍
if args.phase in ['test', 'val', 'train_val']:
# 测试输出文字检测结果
test_dataset = dataloader.DataSet(
test_filelist,
image_label_dict,
num_classes,
# transform=train_transform,
args=args,
image_size=image_size,
phase='test')
test_loader = DataLoader(dataset=test_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
train_filelist = train_filelist1[-2048:]
train_dataset = dataloader.DataSet(train_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='test')
train_loader = DataLoader(dataset=train_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
val_dataset = dataloader.DataSet(val_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='test')
val_loader = DataLoader(dataset=val_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
train_val_dataset = dataloader.DataSet(train_val_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='test')
train_val_loader = DataLoader(dataset=train_val_dataset,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=True)
if args.phase == 'test':
print("now , let's do phase of test")
test(start_epoch - 1, model, val_loader, 'val')
test(start_epoch - 1, model, test_loader, 'test')
# test(start_epoch - 1, model, train_val_loader, 'train_val')
elif args.phase == 'val':
test(start_epoch - 1, model, train_loader, 'train')
test(start_epoch - 1, model, val_loader, 'val')
elif args.phase == 'train_val':
test(start_epoch - 1, model, train_val_loader, 'train_val')
return
elif args.phase == 'train':
# print(train_filelist1[:10])
train_dataset1 = dataloader.DataSet(
train_filelist1,
image_label_dict,
num_classes,
image_size=image_size, # image_size = [512, 64]
args=args,
phase='train')
train_loader1 = DataLoader(dataset=train_dataset1,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
train_dataset2 = dataloader.DataSet(train_filelist2,
image_label_dict,
num_classes,
image_size=(1024, 64),
args=args,
phase='train')
train_loader2 = DataLoader(dataset=train_dataset2,
batch_size=int(args.batch_size / 2),
shuffle=True,
num_workers=8,
pin_memory=True)
val_dataset = dataloader.DataSet(val_filelist,
image_label_dict,
num_classes,
image_size=image_size,
args=args,
phase='val')
val_loader = DataLoader(dataset=val_dataset,
batch_size=min(8, args.batch_size),
shuffle=False,
num_workers=8,
pin_memory=True)
filelist = glob(os.path.join(args.bg_dir, '*'))
pretrain_dataset1 = dataloader.DataSet(filelist,
image_label_dict,
num_classes,
image_size=args.image_size,
word_index_dict=word_index_dict,
args=args,
font_range=[8, 32],
margin=10,
rotate_range=[-10., 10.],
phase='pretrain')
pretrain_loader1 = DataLoader(dataset=pretrain_dataset1,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
pretrain_dataset2 = dataloader.DataSet(filelist,
image_label_dict,
num_classes,
image_size=(256, 128),
word_index_dict=word_index_dict,
args=args,
font_range=[24, 64],
margin=20,
rotate_range=[-20., 20.],
phase='pretrain')
pretrain_loader2 = DataLoader(dataset=pretrain_dataset2,
batch_size=args.batch_size,
shuffle=True,
num_workers=8,
pin_memory=True)
best_f1score = 0
# eval_mode = 'pretrain-2'
eval_mode = 'eval'
for epoch in range(start_epoch, args.epochs):
args.epoch = epoch
if eval_mode == 'eval':
if best_f1score > 0.9:
args.lr = 0.0001
if best_f1score > 0.9:
args.hard_mining = 1
for param_group in optimizer.param_groups:
param_group['lr'] = args.lr
train_eval(epoch, model, train_loader1, loss, optimizer, 2.,
'train-1')
# if best_f1score > 0.9:
# train_eval(epoch, model, train_loader2, loss, optimizer, 2., 'train-2') #取消train_dataset2数据集
best_f1score = train_eval(
epoch, model, val_loader, loss, optimizer, best_f1score,
'eval-{:d}-{:d}'.format(args.batch_size, args.hard_mining))
continue
'''
def main():
global args, best_err1, best_err5, global_epoch_confusion, best_loss
args = parser.parse_args()
if args.dataset.startswith('cifar'):
normalize = transforms.Normalize(
mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
std=[x / 255.0 for x in [63.0, 62.1, 66.7]])
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
transform_test = transforms.Compose([transforms.ToTensor(), normalize])
if args.dataset == 'cifar100':
train_loader = torch.utils.data.DataLoader(
datasets.CIFAR100('../data',
train=True,
download=True,
transform=transform_train),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(
datasets.CIFAR100('../data',
train=False,
transform=transform_test),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
numberofclass = 100
elif args.dataset == 'cifar10':
train_data = datasets.CIFAR10('../data',
train=True,
download=True,
transform=transform_train)
print(train_data.targets[:30])
print(train_data.targets[:30])
print(len(train_data))
class_counts = [9.0, 1.0]
num_samples = sum(class_counts)
labels = [0, 0, ..., 0, 1] #corresponding labels of samples
class_weights = [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
class_weights[args.first] = args.weight
class_weights[args.second] = args.weight
print(class_weights)
weights = [
class_weights[train_data.targets[i]]
for i in range(len(train_data))
]
sampler = WeightedRandomSampler(torch.DoubleTensor(weights),
len(train_data))
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=sampler)
val_loader = torch.utils.data.DataLoader(
datasets.CIFAR10('../data',
train=False,
transform=transform_test),
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers,
pin_memory=True)
numberofclass = 10
else:
raise Exception('unknown dataset: {}'.format(args.dataset))
else:
raise Exception('unknown dataset: {}'.format(args.dataset))
print("=> creating model '{}'".format(args.net_type))
if args.net_type == 'resnet':
model = RN.ResNet(args.dataset, args.depth, numberofclass,
args.bottleneck) # for ResNet
elif args.net_type == 'pyramidnet':
model = PYRM.PyramidNet(args.dataset, args.depth, args.alpha,
numberofclass, args.bottleneck)
else:
raise Exception('unknown network architecture: {}'.format(
args.net_type))
model = torch.nn.DataParallel(model).cuda()
if os.path.isfile(args.pretrained):
print("=> loading checkpoint '{}'".format(args.pretrained))
checkpoint = torch.load(args.pretrained)
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}'".format(args.pretrained))
# print(model)
print('the number of model parameters: {}'.format(
sum([p.data.nelement() for p in model.parameters()])))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss(reduction='none').cuda()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay,
nesterov=True)
cudnn.benchmark = True
#validate(val_loader, model, criterion, 0)
# for checking pre-trained model accuracy and confusion
if args.checkmodel:
global_epoch_confusion.append({})
get_confusion(val_loader, model, criterion)
# cat->dog confusion
log_print(str(args.first) + " -> " + str(args.second))
log_print(global_epoch_confusion[-1]["confusion"][(args.first,
args.second)])
# dog->cat confusion
log_print(str(args.second) + " -> " + str(args.first))
log_print(global_epoch_confusion[-1]["confusion"][(args.second,
args.first)])
exit()
for epoch in range(0, args.epochs):
global_epoch_confusion.append({})
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
err1, err5, val_loss = validate(val_loader, model, criterion, epoch)
# remember best prec@1 and save checkpoint
if epoch // (args.epochs * 0.75):
is_best = err1 <= best_err1
best_err1 = min(err1, best_err1)
if is_best:
best_err5 = err5
best_err1 = err1
print('Current best accuracy (top-1 and 5 error):', best_err1,
best_err5)
save_checkpoint(
{
'epoch': epoch,
'arch': args.net_type,
'state_dict': model.state_dict(),
'best_err1': best_err1,
'best_err5': best_err5,
'optimizer': optimizer.state_dict(),
}, is_best)
get_confusion(val_loader, model, criterion, epoch)
# cat->dog confusion
log_print(str(args.first) + " -> " + str(args.second))
log_print(global_epoch_confusion[-1]["confusion"][(args.first,
args.second)])
# dog->cat confusion
log_print(str(args.second) + " -> " + str(args.first))
log_print(global_epoch_confusion[-1]["confusion"][(args.second,
args.first)])
print('Best accuracy (top-1 and 5 error):', best_err1, best_err5)
directory = "runs/%s/" % (args.expname)
if not os.path.exists(directory):
os.makedirs(directory)
epoch_confusions = 'runs/%s/' % (args.expname) + \
'epoch_confusion_' + args.expid
np.save(epoch_confusions, global_epoch_confusion)
log_print("")
# output best model accuracy and confusion
repaired_model = 'runs/%s/' % (args.expname) + 'model_best.pth.tar'
if os.path.isfile(repaired_model):
print("=> loading checkpoint '{}'".format(repaired_model))
checkpoint = torch.load(repaired_model)
model.load_state_dict(checkpoint['state_dict'])
get_confusion(val_loader, model, criterion)
# dog->cat confusion
log_print(str(args.first) + " -> " + str(args.second))
log_print(global_epoch_confusion[-1]["confusion"][(args.first,
args.second)])
# cat->dog confusion
log_print(str(args.second) + " -> " + str(args.first))
log_print(global_epoch_confusion[-1]["confusion"][(args.second,
args.first)])
def main():
parser = argparse.ArgumentParser(
description='Learning CIFAR-10 using ResNet')
parser.add_argument('--batchsize',
type=int,
default=256,
help='Leaning minibatch size')
parser.add_argument('--epoch',
type=int,
default=365,
help='Number of epochs to train')
parser.add_argument('--gpu',
type=int,
default=-1,
help='GPU id (-1 indicates CPU)')
parser.add_argument('--loaders',
type=int,
default=1,
help='Number of data loading processes')
parser.add_argument('--out',
default='./result',
help='Path of output directory')
parser.add_argument('--path',
default='./cifar-10-batches-py',
help='Path of dataset files')
parser.add_argument('--test', action='store_true')
args = parser.parse_args()
print('==========================================')
print('Num Minibatch-size: {}'.format(args.batchsize))
print('Num Epoch: {}'.format(args.epoch))
print('==========================================')
# Model
model = resnet.ResNet()
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu()
# Dataset
data, labels = load_cifar10_dataset(args.path)
mean = np.mean(data, axis=0)
train = Cifar10Dataset(data[0:45000],
labels[0:45000],
train=True,
mean=mean)
val = Cifar10Dataset(data[45000:50000],
labels[45000:50000],
train=False,
mean=mean)
# Iterators
train_iter = chainer.iterators.MultiprocessIterator(
train, args.batchsize, n_processes=args.loaders)
val_iter = chainer.iterators.MultiprocessIterator(val,
args.batchsize,
repeat=False,
shuffle=False,
n_processes=args.loaders)
# Optimizer
optimizer = chainer.optimizers.MomentumSGD(lr=0.1, momentum=0.9)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(0.0001))
# Trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), args.out)
# Validation
val_interval = (10, 'iteration') if args.test else (1, 'epoch')
trainer.extend(extensions.Evaluator(val_iter, model, device=args.gpu),
trigger=val_interval)
trainer.extend(extensions.snapshot_object(
model, 'model_iter_{.updater.iteration}'),
trigger=val_interval)
# Learning rate decay
lr_interval = training.triggers.ManualScheduleTrigger([32000, 48000],
'iteration')
trainer.extend(extensions.ExponentialShift('lr', 0.1), trigger=lr_interval)
# Log
log_interval = (10, 'iteration') if args.test else (1, 'epoch')
trainer.extend(extensions.observe_lr(), trigger=log_interval)
trainer.extend(extensions.LogReport(trigger=log_interval))
trainer.extend(extensions.PrintReport([
'epoch', 'iteration', 'lr', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'elapsed_time'
]),
trigger=log_interval)
trainer.extend(extensions.ProgressBar(update_interval=10))
trainer.run()
def train():
print('[Dataset Configuration]')
print('\tImageNet training root: %s' % FLAGS.train_image_root)
print('\tImageNet training list: %s' % FLAGS.train_dataset)
print('\tImageNet val root: %s' % FLAGS.val_image_root)
print('\tImageNet val list: %s' % FLAGS.val_dataset)
print('\tNumber of classes: %d' % FLAGS.num_classes)
print('\tNumber of training images: %d' % FLAGS.num_train_instance)
print('\tNumber of val images: %d' % FLAGS.num_val_instance)
print('[Network Configuration]')
print('\tBatch size: %d' % FLAGS.batch_size)
print('\tNumber of GPUs: %d' % FLAGS.num_gpus)
print('\tBasemodel file: %s' % FLAGS.basemodel)
print('[Optimization Configuration]')
print('\tL2 loss weight: %f' % FLAGS.l2_weight)
print('\tThe momentum optimizer: %f' % FLAGS.momentum)
print('\tInitial learning rate: %f' % FLAGS.initial_lr)
print('\tEpochs per lr step: %s' % FLAGS.lr_step_epoch)
print('\tLearning rate decay: %f' % FLAGS.lr_decay)
print('[Training Configuration]')
print('\tTrain dir: %s' % FLAGS.train_dir)
print('\tTraining max steps: %d' % FLAGS.max_steps)
print('\tSteps per displaying info: %d' % FLAGS.display)
print('\tSteps per validation: %d' % FLAGS.val_interval)
print('\tSteps during validation: %d' % FLAGS.val_iter)
print('\tSteps per saving checkpoints: %d' % FLAGS.checkpoint_interval)
print('\tGPU memory fraction: %f' % FLAGS.gpu_fraction)
print('\tLog device placement: %d' % FLAGS.log_device_placement)
with tf.Graph().as_default():
init_step = 0
global_step = tf.Variable(0, trainable=False, name='global_step')
# Get images and labels of ImageNet
import multiprocessing
num_threads = multiprocessing.cpu_count() / FLAGS.num_gpus
print('Load ImageNet dataset(%d threads)' % num_threads)
with tf.device('/cpu:0'):
print('\tLoading training data from %s' % FLAGS.train_dataset)
with tf.variable_scope('train_image'):
train_images, train_labels = data_input.distorted_inputs(
FLAGS.train_image_root,
FLAGS.train_dataset,
FLAGS.batch_size,
True,
num_threads=num_threads,
num_sets=FLAGS.num_gpus)
print('\tLoading validation data from %s' % FLAGS.val_dataset)
with tf.variable_scope('test_image'):
val_images, val_labels = data_input.inputs(
FLAGS.val_image_root,
FLAGS.val_dataset,
FLAGS.batch_size,
False,
num_threads=num_threads,
num_sets=FLAGS.num_gpus)
tf.summary.image('images', train_images[0][:2])
# Build model
lr_decay_steps = map(float, FLAGS.lr_step_epoch.split(','))
lr_decay_steps = list(
map(int, [
s * FLAGS.num_train_instance / FLAGS.batch_size /
FLAGS.num_gpus for s in lr_decay_steps
]))
hp = resnet.HParams(batch_size=FLAGS.batch_size,
num_gpus=FLAGS.num_gpus,
num_classes=FLAGS.num_classes,
weight_decay=FLAGS.l2_weight,
momentum=FLAGS.momentum,
finetune=FLAGS.finetune)
network_train = resnet.ResNet(hp,
train_images,
train_labels,
global_step,
name="train")
network_train.build_model()
network_train.build_train_op()
train_summary_op = tf.summary.merge_all() # Summaries(training)
network_val = resnet.ResNet(hp,
val_images,
val_labels,
global_step,
name="val",
reuse_weights=True)
network_val.build_model()
print('Number of Weights: %d' % network_train._weights)
print('FLOPs: %d' % network_train._flops)
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
allow_soft_placement=False,
# allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Create a saver.
saver = tf.train.Saver(tf.global_variables(), max_to_keep=10000)
if FLAGS.checkpoint is not None:
print('Load checkpoint %s' % FLAGS.checkpoint)
saver.restore(sess, FLAGS.checkpoint)
init_step = global_step.eval(session=sess)
elif FLAGS.basemodel:
# Define a different saver to save model checkpoints
print('Load parameters from basemodel %s' % FLAGS.basemodel)
variables = tf.global_variables()
vars_restore = [
var for var in variables
if not "Momentum" in var.name and not "global_step" in var.name
]
saver_restore = tf.train.Saver(vars_restore, max_to_keep=10000)
saver_restore.restore(sess, FLAGS.basemodel)
else:
print(
'No checkpoint file of basemodel found. Start from the scratch.'
)
# Start queue runners & summary_writer
tf.train.start_queue_runners(sess=sess)
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
summary_writer = tf.summary.FileWriter(
os.path.join(FLAGS.train_dir, str(global_step.eval(session=sess))),
sess.graph)
# Training!
val_best_acc = 0.0
for step in range(init_step, FLAGS.max_steps):
# val
if step % FLAGS.val_interval == 0:
val_loss, val_acc = 0.0, 0.0
for i in range(FLAGS.val_iter):
loss_value, acc_value = sess.run(
[network_val.loss, network_val.acc],
feed_dict={network_val.is_train: False})
val_loss += loss_value
val_acc += acc_value
val_loss /= FLAGS.val_iter
val_acc /= FLAGS.val_iter
val_best_acc = max(val_best_acc, val_acc)
format_str = ('%s: (val) step %d, loss=%.4f, acc=%.4f')
print(format_str % (datetime.now(), step, val_loss, val_acc))
val_summary = tf.Summary()
val_summary.value.add(tag='val/loss', simple_value=val_loss)
val_summary.value.add(tag='val/acc', simple_value=val_acc)
val_summary.value.add(tag='val/best_acc',
simple_value=val_best_acc)
summary_writer.add_summary(val_summary, step)
summary_writer.flush()
# Train
lr_value = get_lr(FLAGS.initial_lr, FLAGS.lr_decay, lr_decay_steps,
step)
start_time = time.time()
# For timeline profiling
# if step == 153:
# run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# run_metadata = tf.RunMetadata()
# _, loss_value, acc_value, train_summary_str = \
# sess.run([network_train.train_op, network_train.loss, network_train.acc, train_summary_op],
# feed_dict={network_train.is_train:True, network_train.lr:lr_value}
# , options=run_options, run_metadata=run_metadata)
# # Create the Timeline object, and write it to a json
# tl = timeline.Timeline(run_metadata.step_stats)
# ctf = tl.generate_chrome_trace_format()
# with open('timeline.json', 'w') as f:
# f.write(ctf)
# print('Wrote the timeline profile of %d iter training on %s' %(step, 'timeline.json'))
# else:
# _, loss_value, acc_value, train_summary_str = \
# sess.run([network_train.train_op, network_train.loss, network_train.acc, train_summary_op],
# feed_dict={network_train.is_train:True, network_train.lr:lr_value})
_, loss_value, acc_value, train_summary_str = \
sess.run([network_train.train_op, network_train.loss, network_train.acc, train_summary_op],
feed_dict={network_train.is_train:True, network_train.lr:lr_value})
duration = time.time() - start_time
assert not np.isnan(loss_value)
# Display & Summary(training)
if step % FLAGS.display == 0 or step < 10:
num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: (Training) step %d, loss=%.4f, acc=%.4f, lr=%f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, acc_value, lr_value,
examples_per_sec, sec_per_batch))
summary_writer.add_summary(train_summary_str, step)
# Save the model checkpoint periodically.
if (step > init_step and step % FLAGS.checkpoint_interval
== 0) or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
if sys.stdin in select.select([sys.stdin], [], [], 0)[0]:
char = sys.stdin.read(1)
if char == 'b':
embed()
def main():
global args, best_err1, best_err5
args = parser.parse_args()
if args.dataset.startswith('cifar'):
normalize = transforms.Normalize(mean=[x / 255.0 for x in [125.3, 123.0, 113.9]],
std=[x / 255.0 for x in [63.0, 62.1, 66.7]])
transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
transform_test = transforms.Compose([
transforms.ToTensor(),
normalize
])
if args.dataset == 'cifar100':
train_loader = torch.utils.data.DataLoader(
datasets.CIFAR100('../data', train=True, download=True, transform=transform_train),
batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
datasets.CIFAR100('../data', train=False, transform=transform_test),
batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
numberofclass = 100
elif args.dataset == 'cifar10':
train_loader = torch.utils.data.DataLoader(
datasets.CIFAR10('../data', train=True, download=True, transform=transform_train),
batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
val_loader = torch.utils.data.DataLoader(
datasets.CIFAR10('../data', train=False, transform=transform_test),
batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
numberofclass = 10
else:
raise Exception('unknown dataset: {}'.format(args.dataset))
elif args.dataset == 'imagenet':
traindir = os.path.join('/home/data/ILSVRC/train')
valdir = os.path.join('/home/data/ILSVRC/val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
jittering = utils.ColorJitter(brightness=0.4, contrast=0.4,
saturation=0.4)
lighting = utils.Lighting(alphastd=0.1,
eigval=[0.2175, 0.0188, 0.0045],
eigvec=[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]])
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
jittering,
lighting,
normalize,
]))
train_sampler = None
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
num_workers=args.workers, pin_memory=True, sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(
datasets.ImageFolder(valdir, transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])),
batch_size=args.batch_size, shuffle=False,
num_workers=args.workers, pin_memory=True)
numberofclass = 1000
else:
raise Exception('unknown dataset: {}'.format(args.dataset))
print("=> creating model '{}'".format(args.net_type))
if args.net_type == 'resnet':
model = RN.ResNet(args.dataset, args.depth, numberofclass, args.bottleneck) # for ResNet
elif args.net_type == 'pyramidnet':
model = PYRM.PyramidNet(args.dataset, args.depth, args.alpha, numberofclass,
args.bottleneck)
else:
raise Exception('unknown network architecture: {}'.format(args.net_type))
model = torch.nn.DataParallel(model).cuda()
print(model)
print('the number of model parameters: {}'.format(sum([p.data.nelement() for p in model.parameters()])))
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay, nesterov=True)
cudnn.benchmark = True
for epoch in range(0, args.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train_loss = train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
err1, err5, val_loss = validate(val_loader, model, criterion, epoch)
# remember best prec@1 and save checkpoint
is_best = err1 <= best_err1
best_err1 = min(err1, best_err1)
if is_best:
best_err5 = err5
print('Current best accuracy (top-1 and 5 error):', best_err1, best_err5)
save_checkpoint({
'epoch': epoch,
'arch': args.net_type,
'state_dict': model.state_dict(),
'best_err1': best_err1,
'best_err5': best_err5,
'optimizer': optimizer.state_dict(),
}, is_best)
print('Best accuracy (top-1 and 5 error):', best_err1, best_err5)
def train():
print('[Dataset Configuration]')
print('\tCIFAR-100 dir: %s' % FLAGS.data_dir)
print('\tNumber of classes: %d' % FLAGS.num_classes)
print('\tNumber of test images: %d' % FLAGS.num_test_instance)
print('[Network Configuration]')
print('\tBatch size: %d' % FLAGS.batch_size)
print('\tResidual blocks per group: %d' % FLAGS.num_residual_units)
print('\tNetwork width multiplier: %d' % FLAGS.k)
print('[Testing Configuration]')
print('\tCheckpoint path: %s' % FLAGS.ckpt_path)
print('\tDataset: %s' % ('Training' if FLAGS.train_data else 'Test'))
print('\tNumber of testing iterations: %d' % FLAGS.test_iter)
print('\tOutput path: %s' % FLAGS.output)
print('\tGPU memory fraction: %f' % FLAGS.gpu_fraction)
print('\tLog device placement: %d' % FLAGS.log_device_placement)
with tf.Graph().as_default():
# Build a Graph that computes the predictions from the inference model.
images = tf.placeholder(
tf.float32,
[FLAGS.batch_size, data_input.HEIGHT, data_input.WIDTH, 3])
labels = tf.placeholder(tf.int32, [FLAGS.batch_size])
# Build model
decay_step = FLAGS.lr_step_epoch * FLAGS.num_train_instance / FLAGS.batch_size
hp = resnet.HParams(batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
num_residual_units=FLAGS.num_residual_units,
k=FLAGS.k,
weight_decay=FLAGS.l2_weight,
initial_lr=FLAGS.initial_lr,
decay_step=decay_step,
lr_decay=FLAGS.lr_decay,
momentum=FLAGS.momentum)
network = resnet.ResNet(hp, images, labels, None)
network.build_model()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Create a saver.
saver = tf.train.Saver(tf.all_variables(), max_to_keep=10000)
if os.path.isdir(FLAGS.ckpt_path):
ckpt = tf.train.get_checkpoint_state(FLAGS.ckpt_path)
# Restores from checkpoint
if ckpt and ckpt.model_checkpoint_path:
print('\tRestore from %s' % ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found in the dir [%s]' %
FLAGS.ckpt_path)
sys.exit(1)
elif os.path.isfile(FLAGS.ckpt_path):
print('\tRestore from %s' % FLAGS.ckpt_path)
saver.restore(sess, FLAGS.ckpt_path)
else:
print('No checkpoint file found in the path [%s]' %
FLAGS.ckpt_path)
sys.exit(1)
graph = tf.get_default_graph()
block_num = 3
conv_num = 2
old_kernels_to_cluster = []
old_kernels_to_add = []
old_batch_norm = []
for i in range(1, block_num + 1):
for j in range(FLAGS.num_residual_units):
old_kernels_to_cluster.append(get_kernel(i, j, 1, graph, sess))
old_kernels_to_add.append(get_kernel(i, j, 2, graph, sess))
old_batch_norm.append(get_batch_norm(i, j, 2, graph, sess))
#old_batch_norm = old_batch_norm[1:]
#old_batch_norm.append(get_last_batch_norm(graph, sess))
new_params = []
new_width = [
16,
int(16 * FLAGS.new_k),
int(32 * FLAGS.new_k),
int(64 * FLAGS.new_k)
]
for i in range(len(old_batch_norm)):
cluster_num = new_width[int(i / 4) + 1]
cluster_kernels, cluster_indices = cluster_kernel(
old_kernels_to_cluster[i], cluster_num)
add_kernels = add_kernel(old_kernels_to_add[i], cluster_indices,
cluster_num)
cluster_batchs_norm = cluster_batch_norm(old_batch_norm[i],
cluster_indices,
cluster_num)
new_params.append(cluster_kernels)
for p in range(BATCH_NORM_PARAM_NUM):
new_params.append(cluster_batchs_norm[p])
new_params.append(add_kernels)
# save variables
init_params = []
new_param_index = 0
for var in tf.global_variables():
update_match = UPDATE_PARAM_REGEX.match(var.name)
skip_match = SKIP_PARAM_REGEX.match(var.name)
if update_match and not skip_match:
print("update {}".format(var.name))
init_params.append((new_params[new_param_index], var.name))
new_param_index += 1
else:
print("not update {}".format(var.name))
var_vector = sess.run(var)
init_params.append((var_vector, var.name))
#close old graph
sess.close()
tf.reset_default_graph()
# build new graph and eval
with tf.Graph().as_default():
# The CIFAR-100 dataset
with tf.variable_scope('test_image'):
test_images, test_labels = data_input.input_fn(
FLAGS.data_dir,
FLAGS.batch_size,
train_mode=FLAGS.train_data,
num_threads=1)
# The class labels
with open(os.path.join(FLAGS.data_dir, 'fine_label_names.txt')) as fd:
classes = [temp.strip() for temp in fd.readlines()]
images = tf.placeholder(
tf.float32,
[FLAGS.batch_size, data_input.HEIGHT, data_input.WIDTH, 3])
labels = tf.placeholder(tf.int32, [FLAGS.batch_size])
new_network = resnet.ResNet(hp, images, labels, None, init_params,
FLAGS.new_k)
new_network.build_model()
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Testing!
result_ll = [[0, 0] for _ in range(FLAGS.num_classes)
] # Correct/wrong counts for each class
test_loss = 0.0, 0.0
for i in range(FLAGS.test_iter):
test_images_val, test_labels_val = sess.run(
[test_images, test_labels])
preds_val, loss_value, acc_value = sess.run(
[new_network.preds, new_network.loss, new_network.acc],
feed_dict={
new_network.is_train: False,
images: test_images_val,
labels: test_labels_val
})
test_loss += loss_value
for j in range(FLAGS.batch_size):
correct = 0 if test_labels_val[j] == preds_val[j] else 1
result_ll[test_labels_val[j] % FLAGS.num_classes][correct] += 1
test_loss /= FLAGS.test_iter
# Summary display & output
acc_list = [float(r[0]) / float(r[0] + r[1]) for r in result_ll]
result_total = np.sum(np.array(result_ll), axis=0)
acc_total = float(result_total[0]) / np.sum(result_total)
print('Class \t\t\tT\tF\tAcc.')
format_str = '%-31s %7d %7d %.5f'
for i in range(FLAGS.num_classes):
print(format_str %
(classes[i], result_ll[i][0], result_ll[i][1], acc_list[i]))
print(format_str %
('(Total)', result_total[0], result_total[1], acc_total))
# Output to file(if specified)
if FLAGS.output.strip():
with open(FLAGS.output, 'w') as fd:
fd.write('Class \t\t\tT\tF\tAcc.\n')
format_str = '%-31s %7d %7d %.5f'
for i in range(FLAGS.num_classes):
t, f = result_ll[i]
format_str = '%-31s %7d %7d %.5f\n'
fd.write(format_str %
(classes[i].replace(' ', '-'), t, f, acc_list[i]))
fd.write(
format_str %
('(Total)', result_total[0], result_total[1], acc_total))
def train():
def Load():
if (FLAGS.is_Simple or FLAGS.is_Train == False):
train_images = np.load("Input/test_data.npy")
train_labels = np.load("Input/test_label.npy")
else:
train_data = np.load("Input/train_data.npy")
train_label = np.load("Input/train_label.npy")
permutation = np.random.permutation(train_data.shape[0])
train_images = train_data[permutation, :, :, :]
train_labels = train_label[permutation, :]
test_data = np.load("Input/test_data.npy")
test_label = np.load("Input/test_label.npy")
mean_data = np.array(np.load("Input/mean_data.npy"), dtype=np.float16)
mean_label = np.load("Input/mean_label.npy")
std_label = np.load("Input/std_label.npy")
return train_images, train_labels, test_data, test_label, mean_data, mean_label, std_label
with tf.Graph().as_default():
init_step = 0
global_step = tf.Variable(0, trainable=False, name='global_step')
X = tf.placeholder(tf.float32, [None, 224, 224, 3],
name='Input_Images')
SHAPE = tf.placeholder(tf.float32, [None, 100], name='SHAPE')
EXP = tf.placeholder(tf.float32, [None, 79], name='EXP')
EULAR = tf.placeholder(tf.float32, [None, 3], name='EULAR')
T = tf.placeholder(tf.float32, [None, 2], name='T')
S = tf.placeholder(tf.float32, [None], name='S')
# Build model
hp = resnet.HParams(batch_size=FLAGS.batch_size,
num_gpus=FLAGS.num_gpus,
num_output=FLAGS.dim_output,
weight_decay=FLAGS.l2_weight,
momentum=FLAGS.momentum,
finetune=FLAGS.finetune)
network_train = resnet.ResNet(hp,
X,
SHAPE,
EXP,
EULAR,
T,
S,
global_step,
name="train")
network_train.build_model()
network_train.build_train_op()
train_summary_op = tf.summary.merge_all() # Summaries(training)
print('Number of Weights: %d' % network_train._weights)
print('FLOPs: %d' % network_train._flops)
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
print("sess 0")
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
allow_soft_placement=False,
# allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
print("sess 1")
sess.run(init)
print("sess done")
# Create a saver.
saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)
if (FLAGS.is_Train == False):
checkpoint_dir = FLAGS.train_dir # os.path.join(export_dir, 'checkpoint')
checkpoints = tf.train.get_checkpoint_state(checkpoint_dir)
if checkpoints and checkpoints.model_checkpoint_path:
checkpoints_name = os.path.basename(
checkpoints.model_checkpoint_path)
saver.restore(sess,
os.path.join(checkpoint_dir, checkpoints_name))
print('Load checkpoint %s' % checkpoints_name)
init_step = global_step.eval(session=sess)
else:
checkpoint_dir = FLAGS.train_dir # os.path.join(export_dir, 'checkpoint')
checkpoints = tf.train.get_checkpoint_state(checkpoint_dir)
if checkpoints and checkpoints.model_checkpoint_path:
checkpoints_name = os.path.basename(
checkpoints.model_checkpoint_path)
saver.restore(sess,
os.path.join(checkpoint_dir, checkpoints_name))
print('Load checkpoint %s' % checkpoints_name)
init_step = global_step.eval(session=sess)
#print('Start from the scratch.')
# if not os.path.exists(FLAGS.train_dir):
# os.mkdir(FLAGS.train_dir)
# summary_writer = tf.summary.FileWriter(os.path.join(FLAGS.train_dir, str(global_step.eval(session=sess))),
# sess.graph)
# Training!
train_images, train_labels, test_data, test_label, mean_data, mean_label, std_label = Load(
)
one_epoch_step = int(len(train_labels) / FLAGS.batch_size)
train_images = (train_images - mean_data) / 255.0
#train_labels = (train_labels-mean_label)/std_label
test_data = (test_data - mean_data) / 255.0
#test_label= (test_label - mean_label)/std_label
print("data done")
if (FLAGS.is_CustomTest == True):
batch_data = (np.load('Input/gx.npy') - mean_data) / 255.0
batch_labels = np.zeros((len(batch_data), 185))
tmp = np.zeros((len(batch_data), 185))
shape_logits, exp_logits, eular_logits, t_logits, s_logits = sess.run(
[network_train.shape_logits, network_train.exp_logits,
network_train.eular_logits, network_train.t_logits,
network_train.s_logits, ], \
feed_dict={network_train.is_train: False, X: batch_data,
SHAPE: batch_labels[:, :100], EXP: batch_labels[:, 100:179],
EULAR: batch_labels[:, 179:182], T: batch_labels[:, 182:184],
S: batch_labels[:, 184]})
tmp[:, 0:100] = np.array(exp_logits)
tmp[:, 100:179] = np.array(shape_logits)
tmp[:, 179:182] = np.array(eular_logits)
tmp[:, 182:184] = np.array(t_logits)
tmp[:, 184][:, None] = np.array(s_logits)
np.savetxt("tmp/gx.txt", tmp)
elif (FLAGS.is_Train == False):
max_iteration = int(len(test_label) / FLAGS.batch_size)
print("max iteration is " + str(max_iteration))
loss_ = 0
tmp = np.zeros([185])
for i in range(10):
print(i)
offset = (i * FLAGS.batch_size) % (test_data.shape[0] -
FLAGS.batch_size)
batch_data = test_data[offset:(offset +
FLAGS.batch_size), :, :, :]
batch_labels = test_label[offset:(offset +
FLAGS.batch_size), :]
shape_logits,exp_logits,eular_logits,t_logits,s_logits = sess.run([network_train.shape_logits,network_train.exp_logits,
network_train.eular_logits,network_train.t_logits,
network_train.s_logits,],\
feed_dict={network_train.is_train: False, X: batch_data,
SHAPE: batch_labels[:,:100],EXP:batch_labels[:,100:179],
EULAR:batch_labels[:,179:182],T:batch_labels[:,182:184],
S:batch_labels[:,184]})
tmp[0:100] = np.array(exp_logits[0, :])
tmp[100:179] = np.array(shape_logits[0, :])
tmp[179:182] = np.array(eular_logits[0, :])
tmp[182:184] = np.array(t_logits[0, :])
tmp[184] = np.array(s_logits[0, :])
#loss_+=loss_value[0]
#print("test loss = " +str(loss_/ max_iteration))
np.savetxt("tmp/" + str(i) + ".txt", tmp)
fig = np.array((batch_data[0, :, :, :] * 255 + mean_data),
dtype=np.uint8)
cv2.imwrite("tmp/" + str(i) + ".jpg", fig)
else:
for step in range(init_step, FLAGS.max_steps):
offset = (step * FLAGS.batch_size) % (train_labels.shape[0] -
FLAGS.batch_size)
batch_data = train_images[offset:(offset +
FLAGS.batch_size), :, :, :]
batch_labels = train_labels[offset:(offset +
FLAGS.batch_size), :]
# Train
lr_value = get_lr(FLAGS.initial_lr, FLAGS.lr_decay,
one_epoch_step, step)
start_time = time.time()
_, loss_value, shape_loss, exp_loss, eular_loss, t_loss, s_loss, points_loss, geo_loss, pose_loss = sess.run(
[
network_train.train_op, network_train.loss,
network_train.shape_loss, network_train.exp_loss,
network_train.eular_loss, network_train.t_loss,
network_train.s_loss, network_train.points_loss,
network_train.geo_loss, network_train.pose_loss
],
feed_dict={
network_train.is_train: True,
network_train.lr: lr_value,
X: batch_data,
SHAPE: batch_labels[:, :100],
EXP: batch_labels[:, 100:179],
EULAR: batch_labels[:, 179:182],
T: batch_labels[:, 182:184],
S: batch_labels[:, 184]
})
duration = time.time() - start_time
assert not np.isnan(loss_value)
# Display & Summary(training)
if step % FLAGS.display == 0 or step < 10:
num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: (Training) step %d, loss=%.4f, lr=%f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, lr_value,
examples_per_sec, sec_per_batch))
format_str = (
'shape_loss=%.4f, exp_loss=%.4f,eular_loss=%.4f,t_loss=%.4f,s_loss=%.4f,points_loss=%.4f,geo_loss=%.4f,pose_loss=%.4f'
)
print(format_str %
(shape_loss, exp_loss, eular_loss, t_loss, s_loss,
points_loss, geo_loss, pose_loss))
elapse = time.time() - start_time
time_left = (FLAGS.max_steps - step) * elapse
print("\tTime left: %02d:%02d:%02d" %
(int(time_left / 3600), int(
time_left % 3600 / 60), time_left % 60))
# summary_writer.add_summary(train_summary_str, step)
# Save the model checkpoint periodically.
if (step > init_step and step % FLAGS.checkpoint_interval
== 0) or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir,
'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False, name='global_step')
images = [tf.placeholder(tf.float32, [2, 224, 224, 3])]
labels = [tf.placeholder(tf.int32, [2])]
# Build model
print("Build ResNet-18 model")
hp = resnet.HParams(batch_size=2,
num_gpus=1,
num_classes=1000,
weight_decay=0.001,
momentum=0.9,
finetune=False)
network_train = resnet.ResNet(hp,
images,
labels,
global_step,
name="train")
network_train.build_model()
print('Number of Weights: %d' % network_train._weights)
print('FLOPs: %d' % network_train._flops)
# Build an initialization operation to run below.
init = tf.global_variables_initializer()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=0.96),
allow_soft_placement=True,
log_device_placement=False))
sess.run(init)
n_classes = len(train_yy)
if len(args.bst) == n_classes:
bs_list = args.bst
elif len(args.bst) == 1:
bs_list = [args.bst[0]] * n_classes
else:
raise RuntimeError()
train_loaders = [utils.cycle(D.DataLoader(ds, bs, shuffle=True)) \
for ds, bs in zip(train_datasets, bs_list)]
if args.model == 'linear':
model = th.nn.Linear(train_x.size(1), n_classes)
elif args.model == 'mlp':
model = mlp.MLP([train_x.size(1), 64, 64, 64, n_classes], th.relu, bn=True)
elif args.model == 'resnet':
model = resnet.ResNet(18, n_classes)[args.model]
else:
raise RuntimeError()
dev = th.device('cpu') if args.gpu < 0 else th.device('cuda:%d' % args.gpu)
model = model.to(dev)
params = list(model.parameters())
kwargs = {'params' : params, 'lr' : args.lr, 'weight_decay' : args.wd}
opt = {'sgd' : optim.SGD(**kwargs),
'adam' : optim.Adam(amsgrad=True, **kwargs)}[args.opt]
metric = getattr(utils, args.metric)
if args.tb:
path = 'tb/%s' % args.id
writer = tb.SummaryWriter(path)
train_writer = tb.SummaryWriter(path + '/a')
val_writer = tb.SummaryWriter(path + '/b')
def train():
print('[Dataset Configuration]')
print('\tCIFAR-100 dir: %s' % FLAGS.data_dir)
print('\tNumber of classes: %d' % FLAGS.num_classes)
print('\tNumber of test images: %d' % FLAGS.num_test_instance)
print('[Network Configuration]')
print('\tBatch size: %d' % FLAGS.batch_size)
print('\tResidual blocks per group: %d' % FLAGS.num_residual_units)
print('\tNetwork width multiplier: %d' % FLAGS.k)
print('[Optimization Configuration]')
print('\tL2 loss weight: %f' % FLAGS.l2_weight)
print('\tThe momentum optimizer: %f' % FLAGS.momentum)
print('\tInitial learning rate: %f' % FLAGS.initial_lr)
print('\tEpochs per lr step: %f' % FLAGS.lr_step_epoch)
print('\tLearning rate decay: %f' % FLAGS.lr_decay)
print('[Training Configuration]')
print('\tTrain dir: %s' % FLAGS.train_dir)
print('\tTraining max steps: %d' % FLAGS.max_steps)
print('\tSteps per displaying info: %d' % FLAGS.display)
print('\tSteps per testing: %d' % FLAGS.test_interval)
print('\tSteps during testing: %d' % FLAGS.test_iter)
print('\tSteps per saving checkpoints: %d' % FLAGS.checkpoint_interval)
print('\tGPU memory fraction: %f' % FLAGS.gpu_fraction)
print('\tLog device placement: %d' % FLAGS.log_device_placement)
with tf.Graph().as_default():
# Build a Graph that computes the predictions from the inference model.
images = tf.placeholder(
tf.float32,
[FLAGS.batch_size, data_input.HEIGHT, data_input.WIDTH, 3])
labels = tf.placeholder(tf.int32, [FLAGS.batch_size])
# Build model
decay_step = FLAGS.lr_step_epoch * FLAGS.num_train_instance / FLAGS.batch_size
hp = resnet.HParams(batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
num_residual_units=FLAGS.num_residual_units,
k=FLAGS.k,
weight_decay=FLAGS.l2_weight,
initial_lr=FLAGS.initial_lr,
decay_step=decay_step,
lr_decay=FLAGS.lr_decay,
momentum=FLAGS.momentum)
network = resnet.ResNet(hp, images, labels, None)
network.build_model()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Create a saver.
saver = tf.train.Saver(tf.all_variables(), max_to_keep=10000)
if os.path.isdir(FLAGS.ckpt_path):
ckpt = tf.train.get_checkpoint_state(FLAGS.ckpt_path)
# Restores from checkpoint
if ckpt and ckpt.model_checkpoint_path:
print('\tRestore from %s' % ckpt.model_checkpoint_path)
saver.restore(sess, ckpt.model_checkpoint_path)
else:
print('No checkpoint file found in the dir [%s]' %
FLAGS.ckpt_path)
sys.exit(1)
elif os.path.isfile(FLAGS.ckpt_path):
print('\tRestore from %s' % FLAGS.ckpt_path)
saver.restore(sess, FLAGS.ckpt_path)
else:
print('No checkpoint file found in the path [%s]' %
FLAGS.ckpt_path)
sys.exit(1)
graph = tf.get_default_graph()
block_num = 3
conv_num = 2
old_kernels_to_cluster = []
old_kernels_to_add = []
old_batch_norm = []
for i in range(1, block_num + 1):
for j in range(FLAGS.num_residual_units):
old_kernels_to_cluster.append(get_kernel(i, j, 1, graph, sess))
old_kernels_to_add.append(get_kernel(i, j, 2, graph, sess))
old_batch_norm.append(get_batch_norm(i, j, 2, graph, sess))
#old_batch_norm = old_batch_norm[1:]
#old_batch_norm.append(get_last_batch_norm(graph, sess))
new_params = []
new_width = [
16,
int(16 * FLAGS.new_k),
int(32 * FLAGS.new_k),
int(64 * FLAGS.new_k)
]
for i in range(len(old_batch_norm)):
cluster_num = new_width[int(i / 4) + 1]
cluster_kernels, cluster_indices = cluster_kernel(
old_kernels_to_cluster[i], cluster_num)
add_kernels = add_kernel(old_kernels_to_add[i], cluster_indices,
cluster_num)
cluster_batchs_norm = cluster_batch_norm(old_batch_norm[i],
cluster_indices,
cluster_num)
new_params.append(cluster_kernels)
for p in range(BATCH_NORM_PARAM_NUM):
new_params.append(cluster_batchs_norm[p])
new_params.append(add_kernels)
# save variables
init_params = []
new_param_index = 0
for var in tf.global_variables():
update_match = UPDATE_PARAM_REGEX.match(var.name)
skip_match = SKIP_PARAM_REGEX.match(var.name)
if update_match and not skip_match:
print("update {}".format(var.name))
init_params.append((new_params[new_param_index], var.name))
new_param_index += 1
else:
print("not update {}".format(var.name))
var_vector = sess.run(var)
init_params.append((var_vector, var.name))
#close old graph
sess.close()
tf.reset_default_graph()
# build new graph and eval
with tf.Graph().as_default():
init_step = 0
global_step = tf.Variable(0, trainable=False, name='global_step')
# Get images and labels of CIFAR-100
with tf.variable_scope('train_image'):
train_images, train_labels = data_input.input_fn(FLAGS.data_dir,
FLAGS.batch_size,
train_mode=True)
with tf.variable_scope('test_image'):
test_images, test_labels = data_input.input_fn(FLAGS.data_dir,
FLAGS.batch_size,
train_mode=False)
# The class labels
with open(os.path.join(FLAGS.data_dir, 'fine_label_names.txt')) as fd:
classes = [temp.strip() for temp in fd.readlines()]
images = tf.placeholder(
tf.float32,
[FLAGS.batch_size, data_input.HEIGHT, data_input.WIDTH, 3])
labels = tf.placeholder(tf.int32, [FLAGS.batch_size])
new_network = resnet.ResNet(hp, images, labels, global_step,
init_params, FLAGS.new_k)
new_network.build_model()
new_network.build_train_op()
train_summary_op = tf.summary.merge_all()
init = tf.initialize_all_variables()
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
log_device_placement=FLAGS.log_device_placement))
sess.run(init)
# Create a saver.
saver = tf.train.Saver(tf.all_variables(), max_to_keep=10000)
ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir)
if ckpt and ckpt.model_checkpoint_path:
print('\tRestore from %s' % ckpt.model_checkpoint_path)
# Restores from checkpoint
saver.restore(sess, ckpt.model_checkpoint_path)
init_step = int(
ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
else:
print('No checkpoint file found. Start from the scratch.')
sys.stdout.flush()
# Start queue runners & summary_writer
tf.train.start_queue_runners(sess=sess)
if not os.path.exists(FLAGS.train_dir):
os.mkdir(FLAGS.train_dir)
summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
# Training!
test_best_acc = 0.0
for step in range(init_step, FLAGS.max_steps):
# Test
if step % FLAGS.test_interval == 0:
test_loss, test_acc = 0.0, 0.0
for i in range(FLAGS.test_iter):
test_images_val, test_labels_val = sess.run(
[test_images, test_labels])
loss_value, acc_value = sess.run(
[new_network.loss, new_network.acc],
feed_dict={
new_network.is_train: False,
images: test_images_val,
labels: test_labels_val
})
test_loss += loss_value
test_acc += acc_value
test_loss /= FLAGS.test_iter
test_acc /= FLAGS.test_iter
test_best_acc = max(test_best_acc, test_acc)
format_str = ('%s: (Test) step %d, loss=%.4f, acc=%.4f')
print(format_str % (datetime.now(), step, test_loss, test_acc))
sys.stdout.flush()
test_summary = tf.Summary()
test_summary.value.add(tag='test/loss', simple_value=test_loss)
test_summary.value.add(tag='test/acc', simple_value=test_acc)
test_summary.value.add(tag='test/best_acc',
simple_value=test_best_acc)
summary_writer.add_summary(test_summary, step)
summary_writer.flush()
# Train
start_time = time.time()
train_images_val, train_labels_val = sess.run(
[train_images, train_labels])
_, lr_value, loss_value, acc_value, train_summary_str = \
sess.run([new_network.train_op, new_network.lr, new_network.loss, new_network.acc, train_summary_op],
feed_dict={new_network.is_train:True, images:train_images_val, labels:train_labels_val})
duration = time.time() - start_time
assert not np.isnan(loss_value)
# Display & Summary(training)
if step % FLAGS.display == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: (Training) step %d, loss=%.4f, acc=%.4f, lr=%f (%.1f examples/sec; %.3f '
'sec/batch)')
print(format_str %
(datetime.now(), step, loss_value, acc_value, lr_value,
examples_per_sec, sec_per_batch))
sys.stdout.flush()
summary_writer.add_summary(train_summary_str, step)
# Save the model checkpoint periodically.
if (step > init_step and step % FLAGS.checkpoint_interval
== 0) or (step + 1) == FLAGS.max_steps:
checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
def eval(root_image, root_model):
print("\nCurrent test image path ==> ", root_image)
print("Current model ==> ", root_model)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# resnet50 : [3, 4, 6, 3]
# resnet101 : [3, 4, 23, 3]
# resnet152 : [3, 8, 36, 3]
model = resnet.ResNet(resnet.Bottleneck, [3, 8, 36, 3])
model.load_state_dict(torch.load(root_model))
model.to(device)
transforms_test = transforms.Compose([
# transforms.Pad(4),
# transforms.RandomHorizontalFlip(),
# transforms.RandomCrop(10),
transforms.Resize((32, 32)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
# test loader
testset = torchvision.datasets.ImageFolder(root=root_image,
transform=transforms_test)
test_loader = torch.utils.data.DataLoader(testset,
batch_size=16,
shuffle=False,
num_workers=4)
class_correct = list(0. for i in range(67))
class_total = list(0. for i in range(67))
classes = os.listdir("./data/rename_headline_piap/")
classes.sort()
class_count = len(classes)
model.eval()
with torch.no_grad():
total_corr = 0.
correct = 0.
total = 0.
for images, labels in test_loader:
images = images.to(device)
labels = labels.to(device)
outputs = model(images)
# best percentage
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
total_corr += (predicted == labels).sum().item()
correct = (predicted == labels).squeeze()
for i in range(len(labels)):
label = labels[i]
class_correct[label] += correct[i].item()
class_total[label] += 1
for i in range(class_count):
### For euc-kr decoding
# unhexlify(classes[i]).decode('euc-kr')
# print('Accuracy of %s : %2d %%' %(unhexlify(classes[i]).decode('euc-kr')[0], 100*class_correct[i]/class_total[i]))
print('Accuracy of %s ==> %2d %%' %
(classes[i], 100 * class_correct[i] / class_total[i]))
print("Accuracy of the network ======> %2d %%" %
(100 * total_corr / total))
def train():
print('[Dataset Configuration]')
print('\tImageNet test root: %s' % FLAGS.test_image_root)
print('\tImageNet test list: %s' % FLAGS.test_dataset)
print('\tNumber of classes: %d' % FLAGS.num_classes)
print('\tNumber of test images: %d' % FLAGS.num_test_instance)
print('[Network Configuration]')
print('\tBatch size: %d' % FLAGS.batch_size)
print('\tCheckpoint file: %s' % FLAGS.checkpoint)
print('[Optimization Configuration]')
print('\tL2 loss weight: %f' % FLAGS.l2_weight)
print('\tThe momentum optimizer: %f' % FLAGS.momentum)
print('\tInitial learning rate: %f' % FLAGS.initial_lr)
print('\tEpochs per lr step: %s' % FLAGS.lr_step_epoch)
print('\tLearning rate decay: %f' % FLAGS.lr_decay)
print('[Evaluation Configuration]')
print('\tOutput file path: %s' % FLAGS.output_file)
print('\tTest iterations: %d' % FLAGS.test_iter)
print('\tSteps per displaying info: %d' % FLAGS.display)
print('\tGPU memory fraction: %f' % FLAGS.gpu_fraction)
print('\tLog device placement: %d' % FLAGS.log_device_placement)
with tf.Graph().as_default():
global_step = tf.Variable(0, trainable=False, name='global_step')
# Get images and labels of ImageNet
print('Load ImageNet dataset')
with tf.device('/cpu:0'):
print('\tLoading test data from %s' % FLAGS.test_dataset)
with tf.variable_scope('test_image'):
test_images, test_labels = data_input.inputs(
FLAGS.test_image_root,
FLAGS.test_dataset,
FLAGS.batch_size,
False,
num_threads=1,
center_crop=True)
# Build a Graph that computes the predictions from the inference model.
images = tf.placeholder(tf.float32, [
FLAGS.batch_size, data_input.IMAGE_HEIGHT, data_input.IMAGE_WIDTH,
3
])
labels = tf.placeholder(tf.int32, [FLAGS.batch_size])
# Build model
with tf.device('/GPU:0'):
hp = resnet.HParams(batch_size=FLAGS.batch_size,
num_classes=FLAGS.num_classes,
weight_decay=FLAGS.l2_weight,
momentum=FLAGS.momentum,
finetune=FLAGS.finetune)
network = resnet.ResNet(hp, images, labels, global_step)
network.build_model()
print('\tNumber of Weights: %d' % network._weights)
print('\tFLOPs: %d' % network._flops)
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# Start running operations on the Graph.
sess = tf.Session(config=tf.ConfigProto(
gpu_options=tf.GPUOptions(
per_process_gpu_memory_fraction=FLAGS.gpu_fraction),
allow_soft_placement=True,
log_device_placement=FLAGS.log_device_placement))
'''debugging attempt
from tensorflow.python import debug as tf_debug
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
def _get_data(datum, tensor):
return tensor == train_images
sess.add_tensor_filter("get_data", _get_data)
'''
sess.run(init)
# Create a saver.
saver = tf.train.Saver(tf.all_variables(), max_to_keep=10000)
if FLAGS.checkpoint is not None:
saver.restore(sess, FLAGS.checkpoint)
print('Load checkpoint %s' % FLAGS.checkpoint)
else:
print(
'No checkpoint file of basemodel found. Start from the scratch.'
)
# Start queue runners & summary_writer
tf.train.start_queue_runners(sess=sess)
# Test!
test_loss = 0.0
test_acc = 0.0
test_time = 0.0
confusion_matrix = np.zeros((FLAGS.num_classes, FLAGS.num_classes),
dtype=np.int32)
for i in range(FLAGS.test_iter):
test_images_val, test_labels_val = sess.run(
[test_images, test_labels])
start_time = time.time()
loss_value, acc_value, pred_value = sess.run(
[network.loss, network.acc, network.preds],
feed_dict={
network.is_train: False,
images: test_images_val,
labels: test_labels_val
})
duration = time.time() - start_time
test_loss += loss_value
test_acc += acc_value
test_time += duration
for l, p in zip(test_labels_val, pred_value):
confusion_matrix[l, p] += 1
if i % FLAGS.display == 0:
num_examples_per_step = FLAGS.batch_size
examples_per_sec = num_examples_per_step / duration
sec_per_batch = float(duration)
format_str = (
'%s: iter %d, loss=%.4f, acc=%.4f (%.1f examples/sec; %.3f sec/batch)'
)
print(format_str % (datetime.now(), i, loss_value, acc_value,
examples_per_sec, sec_per_batch))
test_loss /= FLAGS.test_iter
test_acc /= FLAGS.test_iter
# Print and save results
sec_per_image = test_time / FLAGS.test_iter / FLAGS.batch_size
print('Done! Acc: %.6f, Test time: %.3f sec, %.7f sec/example' %
(test_acc, test_time, sec_per_image))
print('Saving result... ')
result = {
'accuracy': test_acc,
'confusion_matrix': confusion_matrix,
'test_time': test_time,
'sec_per_image': sec_per_image
}
with open(FLAGS.output_file, 'wb') as fd:
pickle.dump(result, fd)
print('done!')
