def __init__(self,
model_path: str = "efficientnet-b0",
init_model=None,
**kwargs):
Finetunable.__init__(self)
self.image_size = EfficientNet_pt.get_image_size(model_path)
self.num_classes = 1000
if init_model is None:
init_model = partial(EfficientNet_pt.from_pretrained,
num_classes=self.num_classes)
with open(download(IMAGENET_LABELS_URL, "efficientnet"), "r") as f:
self.labels = json.load(f)
self.tfms = transforms.Compose([
transforms.Resize(self.image_size, interpolation=Image.BICUBIC),
transforms.CenterCrop(self.image_size),
lambda image: image.convert("RGB"),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
PathModel.__init__(self, model_path, init_model, **kwargs)
self.name = model_path
self.description = "EfficientNet is an image classification model that achieves state-of-the-art accuracy while being an order-of-magnitude smaller and faster than previous models. Trained on ImageNet's 1000 categories."
self.tasks = ["image-classification"]
def main(argv):
usage = """ python3 app.py -m <model_type>
for instance :
python3 app.py -m 0 for b0 efficient net
Details about the models are below:
Name # Params Top-1 Acc. Pretrained?
efficientnet-b0 5.3M 76.3 ✓
efficientnet-b1 7.8M 78.8 ✓
efficientnet-b2 9.2M 79.8 ✓
efficientnet-b3 12M 81.1 ✓
efficientnet-b4 19M 82.6 ✓
efficientnet-b5 30M 83.3 ✓
efficientnet-b6 43M 84.0 -
efficientnet-b7 66M 84.4 -
"""
global model, labels_map, image_size, directory
try:
opts, args = getopt.getopt(argv, "hm:", ["help", "model"])
except getopt.GetoptError:
print(usage)
sys.exit(2)
for opt, arg in opts:
print(opt)
print(arg)
if opt in ("-h", "--help"):
print(usage)
sys.exit()
elif opt in ("-m", "--model"):
model_type = int(arg)
directory = 'upload'
if not os.path.exists(directory):
os.makedirs(directory)
try:
model_name = 'efficientnet-b' + str(model_type)
except:
print("Error : missing model number")
print(usage)
sys.exit(2)
print("API running with " + model_name)
model = EfficientNet.from_pretrained(model_name)
image_size = EfficientNet.get_image_size(model_name)
labels_map = json.load(open('labels_map.txt'))
labels_map = [labels_map[str(i)] for i in range(1000)]
model.eval()
port = 5000
host = '0.0.0.0'
app.run(host=host, port=port, threaded=True)
def get_imagenet_validation(args):
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if 'efficientnet' in args.arch:
image_size = EfficientNet.get_image_size(args.arch.replace('_', '-'))
val_transforms = transforms.Compose([
transforms.Resize(image_size, interpolation=PIL.Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
normalize,
])
else:
val_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])
val_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
os.path.join(args.val_dir, 'imagenet', 'val'), val_transforms),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
return val_loader
def load_efficientnet(saved_model_path):
model_name = "efficientnet-b5"
model = EfficientNetClassifier(n_classes=10)
# model = EfficientNet.from_pretrained(model_name)
image_size = EfficientNet.get_image_size(model_name)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
model.load_state_dict(torch.load(saved_model_path))
return model, image_size, device
def save_efficientnet_from_pytorch(model_name, model_path):
image_size = EfficientNet.get_image_size(model_name)
# Load model
model = EfficientNet.from_pretrained(model_name)
model.set_swish(memory_efficient=False) # swish->x*sigmoid(x)
model.eval()
# Dummy input for ONNX
dummy_input = torch.randn(1, 3, 224, 224)
# Export with ONNX
torch.onnx.export(model, dummy_input, model_path, verbose=True)
def make_data():
imagelist = SSIC.get_train_imagelist(args.validate_ratio)
imagelist = imagelist.transform(
# TODO: I saw some function somehwere that added proven fastai defaults
tfms=([rotate(degrees=(-90, 90), p=1)], []),
size=EfficientNet.get_image_size(args.model_name),
resize_method=ResizeMethod.SQUISH)
data = imagelist.databunch(bs=args.batch_size).normalize(imagenet_stats)
# TODO: data.show_batch(3, figsize=(9, 9))
return data
def main_worker():
writer = SummaryWriter()
model = EfficientNet.from_name(opt.arch,
override_params={
'num_classes': opt.num_class
}).cuda()
if opt.loss_kind == 'CE':
criterion = nn.CrossEntropyLoss().cuda()
elif opt.loss_kind == 'Focal':
criterion = Focal(alpha=[2.5, 2.5, 2.5, 2.5, 2.5, 2.5],
num_class=opt.num_class).cuda()
optimizer = torch.optim.SGD(model.parameters(),
opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay)
cudnn.benchmark = True
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
opt.image_size = EfficientNet.get_image_size(opt.arch)
train_dataset = MyDataset(opt.train_data,
transform=transforms.Compose([
transforms.Resize(opt.image_size + 100),
transforms.RandomRotation(360),
transforms.CenterCrop(opt.image_size),
transforms.ToTensor(), normalize,
Cutout(opt.cutout_n, opt.cutout_len)
])) #Cutout(opt.cutout_n,opt.cutout_len),
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.worker,
pin_memory=True)
#val data
test_dataset = MyDataset(opt.test_data,
transform=transforms.Compose([
transforms.Resize(opt.image_size + 100),
transforms.RandomRotation(360),
transforms.RandomCrop(opt.image_size),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(), normalize
]))
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=opt.test_batch_size,
shuffle=False,
num_workers=opt.worker,
pin_memory=True)
for i in range(0, opt.epochs):
adjust_learning_rate(optimizer, i, opt)
train(model, train_loader, test_loader, criterion, optimizer, writer)
#val_acc,val_loss = validate(model,test_loader,criterion)
#print("{} epoch,acc:{},loss{}".format(i+1,val_acc,val_loss))
writer.close()
def __init__(self, model_name="efficientnet-b0"):
self.model_name = model_name
self.TARGET_IMG_SIZE = EfficientNet.get_image_size(self.model_name)
self.transform = transforms.Compose([
transforms.Resize((self.TARGET_IMG_SIZE, self.TARGET_IMG_SIZE)),
# transforms.CenterCrop(input_size),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.model = EfficientNet.from_pretrained(self.model_name)
self.model.to(self.device)
self.model.eval()
def get_dataset(tfms_list=list(), model_name="b0"):
resolution = EfficientNet.get_image_size("efficientnet-" + model_name)
tfms = transforms.Compose(
[transforms.Resize(resolution),
transforms.CenterCrop(resolution)] + tfms_list)
cal_tech101 = ImageFolder(
root="/home/loc/projects/efficient_net/data/101_ObjectCategories",
transform=tfms)
# Train 60%, val 20%, test 20%
train_size = int(0.6 * len(cal_tech101))
test_size = int(0.2 * len(cal_tech101))
val_size = len(cal_tech101) - train_size - test_size
train_data, val_data, test_data = random_split(
cal_tech101, [train_size, val_size, test_size])
return train_data, val_data, test_data
def efficientnet_load(model_name):
model_name = model_name
print(model_name, 'load complete')
image_size = EfficientNet.get_image_size(model_name)
model = EfficientNet.from_pretrained(model_name,
in_channels=3,
num_classes=50,
batch_norm_momentum=0.99,
batch_norm_epsilon=1e-3,
dropout_rate=0.4,
drop_connect_rate=0.3,
depth_divisor=8,
include_top=True)
return model
def my_effnet_loader(batch_size):
import PIL
from efficientnet_pytorch import EfficientNet
image_size = EfficientNet.get_image_size('efficientnet-b7')
print('image_size:', image_size)
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
test_transform = transforms.Compose([
transforms.Resize(image_size, interpolation=PIL.Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
normalize,
])
test_dataset = Dataset(transform=test_transform)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False, num_workers=16, pin_memory=True)
return test_loader
def __init__(self, model_info):
super(AttentiveCNN, self).__init__()
self.f_channel = 1408
eff_net_flag = model_info['eff_flag']
is_train = model_info['phase_train']
if eff_net_flag == True:
model_name = model_info['eff_model_name']
# model_name= 'efficientnet-b2'
pretrained_base = '../pretrained/normal/'
image_size = EfficientNet.get_image_size(model_name)
print('==> Building model.. : ', model_name)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if is_train == True:
model = EfficientNet.from_pretrained(model_name,
pretrained_base)
else:
model = EfficientNet.from_name(model_name)
if model_name == 'efficientnet-b2':
self.f_channel = 1408
if model_name == 'efficientnet-b3':
self.f_channel = 1536
if model_name == 'efficientnet-b4':
self.f_channel = 1792
if model_name == 'efficientnet-b5':
self.f_channel = 2048
eff_conv = model.to(device)
num_ftrs = model._fc.in_features
self.eff_conv = eff_conv
self.input_img_size = image_size
self.eff_channel = 1024
self.conv_a = nn.Conv2d(self.f_channel,
self.eff_channel,
kernel_size=1,
bias=False)
else:
# ResNet-152 backend
# resnet = models.resnet152( pretrained=True )
resnet = models.resnet101(pretrained=True)
modules = list(resnet.children()
)[:-2] # delete the last fc layer and avg pool.
resnet_conv = nn.Sequential(*modules) # last conv feature
self.resnet_conv = resnet_conv
self.input_img_size = 224
self.eff_net_flag = eff_net_flag
class EfficientClassifier:
MODEL_NAME = "efficientnet-b7"
IMAGE_SIZE = EfficientNet.get_image_size(MODEL_NAME) # 224
model = None
labels_map = None
def __init__(self):
self._initialize_model()
def _initialize_model(self):
path = str(Path(os.path.abspath(__file__)).parents[1])
labels_map = json.load(open(path + "/labels_map.txt"))
self.labels_map = [labels_map[str(i)] for i in range(1000)]
self.model = EfficientNet.from_pretrained(self.MODEL_NAME)
self.model.eval()
def _preprocess_image(self, image):
tfms = transforms.Compose(
[
transforms.Resize(self.IMAGE_SIZE),
transforms.CenterCrop(self.IMAGE_SIZE),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]
)
image = tfms(image).unsqueeze(0)
return image
def predict(self, img_path) -> json:
image = Image.open(img_path)
preprocessed_image = self._preprocess_image(image)
with torch.no_grad():
logits = self.model(preprocessed_image)
preds = torch.topk(logits, k=5).indices.squeeze(0).tolist()
predicts_json = {"predicts": []}
print("-----")
for idx in preds:
label = self.labels_map[idx]
prob = torch.softmax(logits, dim=1)[0, idx].item()
print("{:<75} ({:.2f}%)".format(label, prob * 100))
predicts_json["predicts"].append(
{"label": label, "prob": format(prob * 100, ".2f")}
)
return predicts_json
def tell(img):
# 创建网络模型
model = EfficientNet.from_pretrained(model_name, num_classes=num_classes)
if use_gpu:
model.cuda()
# 损失函数和优化器
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
lr,
momentum=momentum,
weight_decay=weight_decay)
# 是加载已有的模型
print("=> loading checkpoint '{}'".format(resume_file))
checkpoint = torch.load(resume_file)
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
resume_file, checkpoint['epoch']))
cudnn.benchmark = True
image_size = EfficientNet.get_image_size(model_name)
tfms = transforms.Compose([
transforms.Resize(image_size, interpolation=Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
img = tfms(img).unsqueeze(0)
# 训练得到结果
# switch to train mode
model.train()
with torch.no_grad():
logits = model(img.cuda())
preds = torch.topk(logits, k=5)[1].squeeze(0).tolist()
print('-----')
for idx in preds:
label = labels_map[idx]
prob = torch.softmax(logits, dim=1)[0, idx].item()
print('{:<75} ({:.2f}%)'.format(label, prob * 100))
def __init__(self,
model_path: str = "efficientnet-b0",
init_model=None,
name: str = None,
description: str = None,
tasks: List[str] = None,
details: Dict = None,
device=None):
self.image_size = EfficientNet_pt.get_image_size(model_path)
self.num_classes = 1000
if init_model is None:
init_model = partial(EfficientNet_pt.from_pretrained,
num_classes=self.num_classes)
with open(download(IMAGENET_LABELS_URL, "efficientnet"), "r") as f:
self.labels = json.load(f)
self.labels = {int(k): v for k, v in self.labels.items()}
self.tfms = transforms.Compose([
transforms.Resize(self.image_size, interpolation=Image.BICUBIC),
transforms.CenterCrop(self.image_size),
lambda image: image.convert("RGB"),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]),
])
tasks = tasks or ["image-classification"]
self.optimal_batch_size = 128
self.process_image = self.tfms
PathModel.__init__(self,
model_path,
name=name,
description=description,
details=details,
tasks=tasks,
init_model=init_model,
device=device)
def build_transparent(
model_name: str,
pretrained: bool = False,
extract_blocks: Iterable[int] = [5],
num_classes: Optional[int] = None,
freeze: bool = False,
):
"""Build an encoder model, possibly pretrained."""
if model_name.startswith("resnet"):
model_ft = models.__dict__[model_name](pretrained=pretrained)
if num_classes is not None:
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
return (
TransparentResNet(model_ft,
extract_blocks=extract_blocks,
freeze=freeze),
input_size,
)
elif model_name.startswith("efficientnet"):
if num_classes is None:
num_classes = 1000 # Matching EfficientNet default.
if pretrained:
model_ft = EfficientNet.from_pretrained(model_name,
num_classes=num_classes)
else:
model_ft = EfficientNet.from_name(
model_name, override_params={"num_classes": num_classes})
input_size = EfficientNet.get_image_size(model_name)
return (
TransparentEfficientNet(
model_ft,
input_size,
extract_blocks=extract_blocks,
freeze=freeze,
),
input_size,
)
else:
raise NotImplementedError("Unknown model name {}".format(model_name))
def __init__(self):
super(MyNetwork, self).__init__()
cnn_model_name = 'efficientnet-b1'
self.cnn_model = EfficientNet.from_pretrained(cnn_model_name,
num_classes=2).to(device)
print(EfficientNet.get_image_size(cnn_model_name))
self.cnn = nn.Sequential(
nn.Linear(1280, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(p=0.3),
)
self.meta = nn.Sequential(
nn.Linear(18, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(p=0.3),
nn.Linear(128, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(p=0.3),
nn.Linear(128, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(p=0.3),
)
self.post = nn.Sequential(
nn.Linear(256, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(p=0.3),
nn.Linear(128, 32),
nn.BatchNorm1d(32),
nn.ReLU(),
nn.Dropout(p=0.3),
nn.Linear(32, 2),
)
def initialize_model(
model_name: str, num_classes: int, use_pretrained: bool = True
) -> Tuple[nn.Module, int]:
model_ft = None
input_size = 0
if model_name.startswith("resnet"):
"""Resnet18"""
model_ft = models.__dict__[model_name](pretrained=use_pretrained)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, num_classes)
input_size = 224
# Wrap a FeatureExtractorResnet around this so that it gives the
# feature output as well.
model_ft = FeatureExtractorResnet(model_ft)
elif model_name.startswith("efficientnet"):
""" EfficientNet"""
from efficientnet_pytorch import EfficientNet
from efficientnet_pytorch.utils import (
get_same_padding_conv2d,
round_filters,
)
if num_classes is None:
num_classes = 1000 # Matching EfficientNet default.
if use_pretrained:
model_ft = EfficientNet.from_pretrained(
model_name, num_classes=num_classes
)
else:
model_ft = EfficientNet.from_name(
model_name, override_params={"num_classes": num_classes}
)
input_size = EfficientNet.get_image_size(model_name)
model_ft = FeatureExtractorEfficientnet(model_ft)
else:
raise ValueError("Invalid model name")
return model_ft, input_size
def transform_image(image_bytes):
my_transforms = transforms.Compose([
transforms.Resize(255),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image = Image.open(io.BytesIO(image_bytes)).convert('RGB')
plt.imshow(image)
plt.show()
return my_transforms(image).unsqueeze(0)
model_name = 'efficientnet-b7'
image_size = EfficientNet.get_image_size(model_name)
model = EfficientNet.from_pretrained(model_name, num_classes=2)
model.load_state_dict(torch.load('class2.pt'))
model.eval()
with open("test.png", 'rb') as f:
image_bytes = f.read()
inputs = transform_image(image_bytes=image_bytes)
outputs = model(inputs)
_, predict = torch.max(outputs, 1)
print(predict)
def work():
print(opt_train.use_gpu)
print(opt_train.num_classes)
# 创建网络模型
model = EfficientNet.from_pretrained(opt_train.model_name,
num_classes=opt_train.num_classes)
if opt_train.use_gpu:
model.cuda()
# 损失函数和优化器
criterion = nn.CrossEntropyLoss().cuda()
optimizer = torch.optim.SGD(model.parameters(),
opt_train.lr,
momentum=opt_train.momentum,
weight_decay=opt_train.weight_decay)
# 是否加载已有的模型
if args.resume == 'T':
print("=> loading checkpoint '{}'".format(opt_train.resume_file))
checkpoint = torch.load(opt_train.resume_file)
opt_train.start_epoch = checkpoint['epoch']
opt_train.best_acc1 = checkpoint['best_acc1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
opt_train.resume_file, checkpoint['epoch']))
cudnn.benchmark = True
# 加载训练数据
# 训练集
train_dataset = CarsDataset('train', opt_train.data_dir, None)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=opt_train.batch_size,
shuffle=True,
num_workers=opt_train.num_wokers,
pin_memory=True)
# 验证集
image_size = EfficientNet.get_image_size(opt_train.model_name)
val_dataset = CarsDataset('val', opt_train.data_dir, image_size)
val_loader = DataLoader(val_dataset,
batch_size=opt_train.batch_size,
shuffle=False,
num_workers=opt_train.num_wokers,
pin_memory=True)
for epoch in range(opt_train.start_epoch, opt_train.epochs):
adjust_learning_rate(optimizer, epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion)
# remember best acc@1 and save checkpoint
is_best = acc1 > opt_train.best_acc1
opt_train.best_acc1 = max(acc1, opt_train.best_acc1)
print("save:", epoch)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_acc1': opt_train.best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best)
parser.add_argument('--use32', action='store_true', help='train with 32*32 graph size without resizing')
parser.add_argument('--para', action='store_true', help='use multi GPU')
args = parser.parse_args()
model_file = args.network + "_cifar10.ckpt"
fig_file = args.network + "_cifar10.png"
network = NET.index(args.network)
if network <= 1 or args.use32:
transform = transforms.ToTensor()
elif network <= 7:
transform = transforms.Compose([
transforms.Resize((224,224)),
transforms.ToTensor()
])
elif network <= 11:
size = EfficientNet.get_image_size('efficientnet-b{}'.format(network-8))
transform = transforms.Compose([
transforms.Resize((size,size)),
transforms.ToTensor()
])
else:
sys.exit(1)
# prepare data
train_Data = dsets.CIFAR10(
root='../data_cifar10',
train=True,
transform=transform,
download=False
)
from efficientnet_pytorch import EfficientNet
from sotabencheval.image_classification import ImageNetEvaluator
from sotabencheval.utils import is_server
if is_server():
DATA_ROOT = './.data/vision/imagenet'
else: # local settings
DATA_ROOT = os.environ['IMAGENET_DIR']
assert bool(DATA_ROOT), 'please set IMAGENET_DIR environment variable'
print('Local data root: ', DATA_ROOT)
model_name = 'EfficientNet-B5'
model = EfficientNet.from_pretrained(model_name.lower())
image_size = EfficientNet.get_image_size(model_name.lower())
input_transform = transforms.Compose([
transforms.Resize(image_size, PIL.Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224,
0.225]),
])
test_dataset = ImageNet(
DATA_ROOT,
split="val",
transform=input_transform,
target_transform=None,
)
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# create model
if 'efficientnet' in args.arch: # NEW
if args.pretrained:
model = EfficientNet.from_pretrained(args.arch,
advprop=args.advprop)
print("=> using pre-trained model '{}'".format(args.arch))
else:
print("=> creating model '{}'".format(args.arch))
model = EfficientNet.from_name(args.arch)
else:
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
if args.arch.find('alexnet') != -1:
model = models.__dict__[args.arch](pretrained=True)
elif args.arch.find('inception_v3') != -1:
model = models.inception_v3(pretrained=True)
elif args.arch.find('densenet121') != -1:
model = models.densenet121(pretrained=True)
elif args.arch.find('resnet') != -1: # ResNet
model = models.__dict__[args.arch](pretrained=True)
else:
print(
'### please check the args.arch for load model in training###'
)
exit(-1)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch]()
if args.fine_tuning:
print(
"=> transfer-learning mode + fine-tuning (train only the last FC layer)"
)
# Freeze Previous Layers(now we are using them as features extractor)
#jiangjiewei
# for param in model.parameters():
# param.requires_grad = False
# Fine Tuning the last Layer For the new task
# juge network: alexnet, inception_v3, densennet, resnet50
if args.arch.find('alexnet') != -1:
num_ftrs = model.classifier[6].in_features
model.classifier[6] = nn.Linear(num_ftrs, 3)
elif args.arch.find('inception_v3') != -1:
num_ftrs = model.fc.in_features
num_auxftrs = model.AuxLogits.fc.in_features
model.fc = nn.Linear(num_ftrs, 3)
model.AuxLogits.fc = nn.Linear(num_auxftrs, 3)
model.aux_logits = False
elif args.arch.find('densenet121') != -1:
num_ftrs = model.classifier.in_features
model.classifier = nn.Linear(num_ftrs, 3)
elif args.arch.find('resnet') != -1: # ResNet
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 3)
else:
print(
"###Error: Fine-tuning is not supported on this architecture.###"
)
exit(-1)
print(model)
else:
parameters = model.parameters()
# name, parma_1 = model.classifier[6].parameters()
# for name, param in model.named_parameters():
# if param.requires_grad:
# print(name)
if args.distributed:
# For multiprocessing distributed, DistributedDataParallel constructor
# should always set the single device scope, otherwise,
# DistributedDataParallel will use all available devices.
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.workers = int(args.workers / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
# DataParallel will divide and allocate batch_size to all available GPUs
if args.arch.startswith('alexnet-1') or args.arch.startswith('vgg-1'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
# criterion = nn.CrossEntropyLoss().cuda(args.gpu)
#jiangjiewei add weight for crossentropyloss
criterion = nn.CrossEntropyLoss(weight=torch.Tensor([1.5, 1.0, 3.0])).cuda(
args.gpu)
# use_cuda = True
# device = torch.device("cuda" if use_cuda else "cpu")
# class_weights = torch.FloatTensor([1.0, 0.2, 1.0]).cuda()
# criterion = nn.CrossEntropyLoss(weight=class_weights).to(device)
if args.arch.find('alexnet') != -1:
fine_tune_parameters = model.classifier[6].parameters()
elif args.arch.find('inception_v3') != -1:
fine_tune_parameters = model.module.fc.parameters()
elif args.arch.find('densenet121') != -1:
fine_tune_parameters = model.module.classifier.parameters()
elif args.arch.find('resnet') != -1: # ResNet
fine_tune_parameters = model.module.fc.parameters()
else:
print('### please check the ignored params ###')
exit(-1)
ignored_params = list(map(id, fine_tune_parameters))
if args.arch.find('alexnet') != -1:
base_params = filter(lambda p: id(p) not in ignored_params,
model.parameters())
else:
base_params = filter(lambda p: id(p) not in ignored_params,
model.module.parameters())
optimizer = torch.optim.SGD(
[
{
'params': base_params
}, #model.parameters()
{
'params': fine_tune_parameters,
'lr': 10 * args.lr
}
],
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
if args.gpu is not None:
# best_acc1 may be from a checkpoint from a different GPU
best_acc1 = best_acc1.to(args.gpu)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
traindir = os.path.join(args.data, 'train1')
valdir = os.path.join(args.data, 'val1')
if args.advprop:
normalize = transforms.Lambda(lambda img: img * 2.0 - 1.0)
else:
normalize = transforms.Normalize(
mean=[0.5765036, 0.34929818, 0.2401832],
std=[0.2179051, 0.19200659, 0.17808074])
if 'efficientnet' in args.arch:
image_size = EfficientNet.get_image_size(args.arch)
else:
image_size = args.image_size
train_dataset = datasets.ImageFolder(
traindir,
transforms.Compose([
# transforms.Resize((256, 256), interpolation=PIL.Image.BICUBIC),
# transforms.Resize((224, 224)),
transforms.Resize((args.image_size, args.image_size),
interpolation=PIL.Image.BICUBIC),
# transforms.RandomResizedCrop((image_size, image_size) ), #RandomRotation scale=(0.9, 1.0)
transforms.RandomRotation(90),
transforms.RandomHorizontalFlip(),
transforms.RandomVerticalFlip(),
transforms.ToTensor(),
normalize,
]))
print('classes:', train_dataset.classes)
# Get number of labels
labels_length = len(train_dataset.classes)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
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_transforms = transforms.Compose([
transforms.Resize((args.image_size, args.image_size),
interpolation=PIL.Image.BICUBIC),
# transforms.CenterCrop((image_size,image_size)),
transforms.ToTensor(),
normalize,
])
val_loader = torch.utils.data.DataLoader(datasets.ImageFolder(
valdir, val_transforms),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
if args.evaluate:
res = validate(val_loader, model, criterion, args)
with open('res.txt', 'w') as f:
print(res, file=f)
return
for epoch in range(args.start_epoch, args.epochs):
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion, args)
# remember best acc@1 and save checkpoint
is_best = acc1 > best_acc1
best_acc1 = max(acc1, best_acc1)
if args.arch.find('alexnet') != -1:
pre_name = './alexnet'
elif args.arch.find('inception_v3') != -1:
pre_name = './inception_v3'
elif args.arch.find('densenet121') != -1:
pre_name = './densenet121'
elif args.arch.find('resnet50') != -1:
pre_name = './resnet50'
else:
print('### please check the args.arch for pre_name###')
exit(-1)
if not args.multiprocessing_distributed or (
args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best, pre_name)
# PATH = pre_name + '_fundus_net.pth'
# torch.save(model.state_dict(), PATH)
print('Finished Training')
parser.add_argument('-v', '--valid', required=True)
parser.add_argument('-b', '--batchsize', required=False, default=8)
parser.add_argument('-mn', '--modelname', required=False, default="b5")
args = parser.parse_args()
MODEL_PATH = args.modeldir
TRAIN_PATH = args.train
VALID_PATH = args.valid
BATCH_SIZE = int(args.batchsize)
MODEL_NAME = "efficientnet-" + args.modelname
src = (ImageList.from_folder(
path=TRAIN_PATH).split_by_rand_pct(0.2).label_from_folder())
transforms = ([rotate(degrees=(-90, 90), p=0.8)], [])
image_size = EfficientNet.get_image_size(MODEL_NAME)
data = (src.transform(transforms,
size=image_size,
resize_method=ResizeMethod.SQUISH).databunch(
bs=BATCH_SIZE).normalize(imagenet_stats))
data.show_batch(3, figsize=(9, 9))
"""
Load Model
"""
if torch.cuda.is_available():
state_dict = torch.load(MODEL_PATH)
else:
state_dict = torch.load(MODEL_PATH, map_location=torch.device('cpu'))
model = EfficientNet.from_pretrained(MODEL_NAME)
#========================torch to onnx Gender ===================================================
# '''
import torch
from efficientnet_pytorch import EfficientNet
import sys
import torch.nn as nn
from collections import OrderedDict
# Specify which model to use
model_file = sys.argv[1]
model_name = model_file.split('.')[0]
output_name = "{}.onnx".format(model_name)
backbone = 'efficientnet-b7'
image_size = EfficientNet.get_image_size(backbone)
print('Image size: ', image_size)
# Load model
model = EfficientNet.from_pretrained(backbone, num_classes=2)
checkpoint = torch.load(model_file)
# new_state_dict = OrderedDict()
# for k, v in checkpoint['model'].items():
# name = k.replace("module.", "")
# new_state_dict[name]= v
# num_features = model._fc.in_features
# print(num_features)
# model._fc = nn.Linear(num_features, 4) #for emotion
# model._fc = nn.Linear(num_features, 7)
def run():
########################################################################
# Register Prediction Start
########################################################################
aicrowd_helpers.execution_start()
########################################################################
# Gather Input and Output paths from environment variables
########################################################################
test_images_path, predictions_output_path = gather_input_output_path()
########################################################################
# Gather Image Names
########################################################################
image_names = gather_image_names(test_images_path)
########################################################################
# Do your magic here to train the model
########################################################################
# Preprocess first here
src = (ImageList.from_folder(
path='fastai-data').split_by_rand_pct(0.0).label_from_folder()
) # any folder which has sub-folders will work
tfms = get_transforms(do_flip=True,
flip_vert=False,
max_rotate=10.0,
max_zoom=1.1,
max_lighting=0.2,
max_warp=0.2,
p_affine=0.75,
p_lighting=0.75) # some tfms
data = (src.transform(tfms, size=360,
resize_method=ResizeMethod.SQUISH).databunch(
bs=32, num_workers=0).normalize(imagenet_stats))
learn = Learner(
data, SnakeDetector(arch=models.resnet50), loss_func=L1Loss(
)) #temp data, we wont be using this, we will be using src_new
learn.split([learn.model.cnn[:6], learn.model.cnn[6:], learn.model.head])
state_dict = torch.load(
'fastai-data/models/snake-detection-model.pth') #our trained model
learn.model.load_state_dict(state_dict['model'])
if not os.path.exists('preprocessed-images'):
os.makedirs('preprocessed-images') #directory to store files
src_new = (ImageList.from_folder(
path=test_images_path).split_by_rand_pct(0.0).label_from_folder()
) # fetch given test images from data/images
for filename in src_new.items:
try:
im = cv2.imread(f"{filename}", cv2.IMREAD_COLOR)
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
im = cv2.resize(im, (360, 360), interpolation=cv2.INTER_AREA)
im_height, im_width, _ = im.shape
orig_im = cv2.imread(f"{filename}", cv2.IMREAD_COLOR)
orig_im_height, orig_im_width, _ = orig_im.shape
to_pred = open_image(filename)
_, _, bbox = learn.predict(to_pred)
im_original = cv2.imread(f"{filename}", cv2.IMREAD_COLOR)
im_original = cv2.cvtColor(im_original, cv2.COLOR_BGR2RGB)
im_original.shape
im_original_width = im_original.shape[1]
im_original_height = im_original.shape[0]
bbox_new = bbox
bbox_new[0] = bbox_new[0] * im_original_width
bbox_new[2] = bbox_new[2] * im_original_width
bbox_new[1] = bbox_new[1] * im_original_height
bbox_new[3] = bbox_new[3] * im_original_height
x_min, y_min, x_max, y_max = map(int, bbox_new)
#cv2.rectangle(img, (x_min, y_min), (x_max, y_max), color=color, thickness=thickness)
im_original = im_original[
y_min:y_max, x_min:
x_max] #cropping is just slicing opencv uses h,w . which is y,x here
im_original = cv2.cvtColor(im_original, cv2.COLOR_BGR2RGB)
filename_str = str(filename)
to_save = filename_str.replace(
test_images_path, 'preprocessed-images'
) # original file is in /data/images/*.jpg -> store it at preprocessed-images/*.jpg and use preprocessed-images folder later
cv2.imwrite(to_save, im_original)
except:
pass
del learn
gc.collect()
model_name = 'efficientnet-b5'
image_size = EfficientNet.get_image_size(model_name)
model = EfficientNet.from_pretrained(model_name)
np.random.seed(13)
src = (ImageList.from_folder(
path='fastai-data').split_by_rand_pct(0.2).label_from_folder())
object_detection_results_path = os.getcwd(
) + "/preprocessed-images" #output of previous step is stored here.
src.add_test_folder(object_detection_results_path)
tfms = ([rotate(degrees=(-90, 90), p=0.8)], [])
bs = 8
data = (src.transform(tfms,
size=image_size,
resize_method=ResizeMethod.SQUISH).databunch(
bs=bs, num_workers=0).normalize(imagenet_stats))
model.add_module('_fc', nn.Linear(
2048,
data.c)) #Replace the final layer of b5 model with number of classes
loss_func = LabelSmoothingCrossEntropy() #following EfficientNet paper
RMSprop = partial(torch.optim.RMSprop) #Following EfficientNet paper
learn = Learner(data,
model,
loss_func=loss_func,
opt_func=RMSprop,
metrics=[accuracy,
FBeta(beta=1, average='macro')])
learn.split(
[[learn.model._conv_stem, learn.model._bn0, learn.model._blocks[:19]],
[learn.model._blocks[19:], learn.model._conv_head],
[learn.model._bn1, learn.model._fc]]) #for differential learning
learn.load(
'b5-seed-13-round-3'
) #best single model - 85.2 local cv. Try ensemble later. with 83.1 and 85.2 models
########################################################################
# Generate Predictions
########################################################################
preds, _ = learn.TTA(ds_type=DatasetType.Test) # Test time augmentation
probs_seed_13 = np.exp(preds) / np.exp(preds).sum(1)[:, None]
del learn
gc.collect() #garbage collect
model = EfficientNet.from_pretrained(model_name)
model.add_module('_fc', nn.Linear(2048, data.c))
learn = Learner(data,
model,
loss_func=loss_func,
opt_func=RMSprop,
metrics=[accuracy,
FBeta(beta=1, average='macro')]) #mew learner
learn.split(
[[learn.model._conv_stem, learn.model._bn0, learn.model._blocks[:19]],
[learn.model._blocks[19:], learn.model._conv_head],
[learn.model._bn1,
learn.model._fc]]) #not needed, but not takin chances
learn.load('b5-seed-15-round-7') #83.6, 83.1 localcv model
preds, _ = learn.TTA(ds_type=DatasetType.Test)
probs_seed_15 = np.exp(preds) / np.exp(preds).sum(1)[:, None]
probs = (probs_seed_13 + probs_seed_15) / 2
probs_np = probs.numpy()
df_test = pd.read_csv(
'given_submission_sample_file.csv', low_memory=False
) #use the given sample file to replace probabilites with our model predictions, This way, no need to worry about corrupted images
df_classes = pd.read_csv('class.csv', low_memory=False) #class mapping
data_dict = df_classes.set_index(
'class_idx')['original_class'].to_dict() #for look up
probs_np = probs.numpy()
df_testfile_mapping = pd.DataFrame(columns=['filename', 'map'])
df_testfile_mapping['filename'] = df_test['filename']
for i, row in df_testfile_mapping.iterrows():
row['map'] = i
data_dict_filename = df_testfile_mapping.set_index(
'filename'
)['map'].to_dict(
) # for lookup, returns the index where the filename is found in the given submission file
i = 0
for test_file in data.test_ds.items:
filename = (os.path.basename(test_file))
map_val = int(data_dict_filename[filename])
for c in range(0, 45):
df_test.loc[
map_val,
data_dict[int(data.classes[c].split("-")[1])]] = probs_np[i][c]
i += 1
#around 7 predictions causes assertion error, for now submit them as class-204
for i, row in df_test.iterrows():
sum_temp = row[1:46].sum()
low_limit = 1 - 1e-6
high_limit = 1 + 1e-6
if not (sum_temp >= low_limit and sum_temp <= high_limit):
for c in range(1, 46):
df_test.iloc[i, c] = 0.
df_test.loc[i, 'thamnophis_sirtalis'] = 1.
df_test.to_csv('generated-submission-3.csv', index=False)
copyfile('generated-submission-3.csv',
predictions_output_path) #save to output path
########################################################################
# Register Prediction Complete
########################################################################
aicrowd_helpers.execution_success(
{"predictions_output_path": predictions_output_path})
def main_worker(gpu, ngpus_per_node, args):
global best_acc1
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# create model
if 'efficientnet' in args.arch: # NEW efficientnet-b4
if args.pretrained:
model = EfficientNet.from_pretrained(args.arch, num_classes=29)
print("=> using pre-trained model '{}'".format(args.arch))
else:
print("=> creating model '{}'".format(args.arch))
model = EfficientNet.from_name(args.arch)
else:
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
model = models.__dict__[args.arch](pretrained=True)
else:
print("=> creating model '{}'".format(args.arch))
model = models.__dict__[args.arch]()
if args.distributed:
# For multiprocessing distributed, DistributedDataParallel constructor
# should always set the single device scope, otherwise,
# DistributedDataParallel will use all available devices.
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.workers = int(args.workers / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
# DataParallel will divide and allocate batch_size to all available GPUs
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda(args.gpu)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
if args.gpu is not None:
# best_acc1 may be from a checkpoint from a different GPU
best_acc1 = best_acc1.to(args.gpu)
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# Data loading code
# traindir = os.path.join(args.data, 'train')
# valdir = os.path.join(args.data, 'val')
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# train_dataset = datasets.ImageFolder(
# traindir,
# transforms.Compose([
# transforms.Resize(256)，
# transforms.RandomResizedCrop(224),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# normalize,
# ]))
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset)
else:
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)
train_transform = transforms.Compose([
transforms.Resize(256), ##输入统一256大小，
transforms.RandomResizedCrop(
224), #从原图像随机切割一张（224， 224）的图像，再统一resize 224
transforms.RandomHorizontalFlip(), ##水平翻转
transforms.RandomVerticalFlip(p=0.5), #垂直翻转
transforms.RandomRotation(30), #在（-30， 30）范围内旋转
transforms.ToTensor(),
normalize # transforms.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]) #
])
train_loader = torch.utils.data.DataLoader(
dataset=imgdata("crop//dataset/yun/train.txt",
train_transform), # return img and label
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler)
if 'efficientnet' in args.arch:
image_size = EfficientNet.get_image_size(args.arch)
val_transforms = transforms.Compose([
transforms.Resize(image_size, interpolation=PIL.Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
normalize,
])
print('Using image size', image_size)
else:
val_transforms = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
])
print('Using image size', 224)
val_loader = torch.utils.data.DataLoader(dataset=test_imgdata(
"crop//dataset/yun/train.txt", val_transforms),
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
# if args.evaluate:
# res = validate(val_loader, model, criterion, args)
# with open('res.txt', 'w') as f:
# print(res, file=f)
# print("validation")
# return
for epoch in range(args.start_epoch, args.epochs):
print("%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%")
print(args.start_epoch, args.epochs)
if args.distributed:
train_sampler.set_epoch(epoch)
adjust_learning_rate(optimizer, epoch, args)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch, args)
# evaluate on validation set
acc1 = validate(val_loader, model, criterion, args)
# remember best acc@1 and save checkpoint
is_best = acc1 >= best_acc1
best_acc1 = max(acc1, best_acc1)
if not args.multiprocessing_distributed or (
args.multiprocessing_distributed
and args.rank % ngpus_per_node == 0):
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_acc1': best_acc1,
'optimizer': optimizer.state_dict(),
}, is_best)
warnings.filterwarnings('ignore')
np.set_printoptions(linewidth=np.inf)
np.set_printoptions(threshold=np.inf)
# Arguments
networkname = "efficientnet-b1"
test_interval = 100
save_interval = 1000
batch_size = 64
max_epochs = 105
learning_rate = 3e-5
max_cells = 31
dir = "b1-3e-5-long"
datadir = "/data/128p_pc/"
img_size = EfficientNet.get_image_size(networkname)
os.makedirs("../logs", exist_ok=True)
os.makedirs("../logs/" + dir, exist_ok=True)
train_mae_log = []
test_mae_log = []
train_dev_log = []
test_dev_log = []
train_cor_log = []
test_cor_log = []
class CustomDatasetFromImages(Dataset):
def __init__(self, csv_path):
# Transforms
self.to_tensor = transforms.Compose([
return result
def post_transforms():
# we use ImageNet image normalization
# and convert it to torch.Tensor
return [A.Normalize(p=1.0), ToTensorV2(p=1.0), ]
if __name__ == "__main__":
warnings.simplefilter("ignore", UserWarning)
warnings.simplefilter("ignore", DeprecationWarning)
warnings.filterwarnings('ignore')
os.environ["PYTHONWARNINGS"] = "ignore"
config = ConfigExperiment()
config.size = EfficientNet.get_image_size(config.model_name)
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
utils.set_global_seed(config.seed)
utils.prepare_cudnn(deterministic=True)
train_transforms = plant.compose([
pre_transforms(config.size),
hard_transforms(),
post_transforms()
])
valid_transforms = plant.compose([
pre_transforms(config.size),
post_transforms()
])
def task(gpu, ngpus_per_node, args):
args.gpu = gpu
if args.gpu is not None:
print("Use GPU: {} for training".format(args.gpu))
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
# For multiprocessing distributed training, rank needs to be the
# global rank among all the processes
args.rank = args.rank * ngpus_per_node + gpu
#dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
# world_size=args.world_size, rank=args.rank)
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
# create model
if 'efficientnet' in args.arch:
model = EfficientNet.from_pretrained(args.arch)
else:
model = models.__dict__[args.arch](pretrained=True)
if args.distributed:
if args.gpu is not None:
torch.cuda.set_device(args.gpu)
model.cuda(args.gpu)
# When using a single GPU per process and per
# DistributedDataParallel, we need to divide the batch size
# ourselves based on the total number of GPUs we have
args.batch_size = int(args.batch_size / ngpus_per_node)
args.workers = int(args.workers / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.gpu])
else:
model.cuda()
# DistributedDataParallel will divide and allocate batch_size to all
# available GPUs if device_ids are not set
model = torch.nn.parallel.DistributedDataParallel(model)
elif args.gpu is not None:
torch.cuda.set_device(args.gpu)
model = model.cuda(args.gpu)
else:
# DataParallel will divide and allocate batch_size to all available GPUs
if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
model.features = torch.nn.DataParallel(model.features)
model.cuda()
else:
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
# data
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if 'efficientnet' in args.arch:
image_size = EfficientNet.get_image_size(args.arch)
dataset = ImageFileList(
args.data_root, args.data_list,
transforms.Compose([
transforms.Resize(image_size, interpolation=Image.BICUBIC),
transforms.CenterCrop(image_size),
transforms.ToTensor(),
normalize,
]))
else:
dataset = ImageFileList(
args.data_root, args.data_list,
transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
normalize,
]))
if args.distributed:
sampler = DistributedSequentialSampler(dataset, args.world_size,
args.rank)
else:
sampler = None
loader = torch.utils.data.DataLoader(dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=sampler)
# extract
batch_time = AverageMeter('Time', ':6.3f')
data_time = AverageMeter('Data', ':6.3f')
progress = ProgressMeter(len(loader), batch_time, prefix='Test: ')
model.eval()
end = time.time()
features = []
for i, images in enumerate(loader):
data_time.update(time.time() - end)
if args.gpu is not None:
images = images.cuda(args.gpu, non_blocking=True)
with torch.no_grad():
if args.action == 'extract':
if args.distributed:
feat = F.adaptive_avg_pool2d(
model.module.extract_features(images),
1).squeeze(-1).squeeze(-1)
else:
feat = F.adaptive_avg_pool2d(
model.extract_features(images),
1).squeeze(-1).squeeze(-1)
else:
feat = torch.softmax(model(images), dim=1)
features.append(feat.cpu().numpy())
batch_time.update(time.time() - end)
end = time.time()
if i % args.print_freq == 0:
progress.print(i)
features = np.concatenate(features, axis=0)
if args.distributed:
all_features = gather_tensors_batch(features, 100)
if args.rank == 0:
np.save(args.output, all_features)
else:
np.save(args.output, features)