def __init__(self, num_classes, head_ps=0.5, pretrained=True, in_channels=3, **kwargs):
self.effnet = EfficientNet.from_pretrained(self.arch)
if in_channels == 1:
w = self.effnet._conv_stem.weight.data[:,0].unsqueeze(1)
out_channels = self.effnet._conv_stem.out_channels
self.effnet._conv_stem = Conv2dStaticSamePadding(1, out_channels, kernel_size=(3,3), stride=(2,2),
bias=False, image_size=(128,128))
self.effnet._conv_stem.weight.data = w
num_features = self.effnet._bn1.num_features
self.head = create_head(num_features*2, num_classes, ps=head_ps)
def __init__(self, in_channels, out_channels, image_size=256):
super().__init__()
self.net = EfficientNetOri.from_pretrained('efficientnet-b0',
num_classes=out_channels)
if (in_channels is not 3):
self.net._conv_stem = Conv2dStaticSamePadding(
in_channels,
32,
image_size=image_size,
kernel_size=(3, 3),
stride=(2, 2),
bias=False)
def __init__(self, in_channels, out_channels, image_size=512):
super().__init__()
self.net = EfficientNetOri.from_name('efficientnet-b0')
if (in_channels != 3):
self.net._conv_stem = Conv2dStaticSamePadding(
in_channels,
32,
image_size=image_size,
kernel_size=(3, 3),
stride=(2, 2),
bias=False)
feature = self.net._fc.in_features
self.net._fc = nn.Linear(in_features=feature,
out_features=out_channels,
bias=True)
def __init__(self, coeff=0):
super(AttnCNN, self).__init__()
self.model = EfficientNet.from_pretrained(
model_name='efficientnet-b%d' % coeff)
self.model._conv_stem = Conv2dStaticSamePadding(
1,
self.model._conv_stem.out_channels,
kernel_size=(3, 3),
stride=(2, 2),
bias=False,
image_size=128)
self.out = AttnBiLSTM(num_classes=7,
input_size=self.model._fc.in_features,
hidden_size=512)
self.feature_dim = self.out.feature_dim
del self.model._fc
def __init__(self, pretrain: str, n_classes: int = 1):
"""
Args:
pretrain (str): one of:
- 'efficientnet-b0'
- 'efficientnet-b1'
- 'efficientnet-b2'
- 'efficientnet-b3'
- 'efficientnet-b4'
- 'efficientnet-b5'
- 'efficientnet-b6'
- 'efficientnet-b7'
"""
super().__init__(pretrain, n_classes)
self.backbone._conv_stem = Conv2dStaticSamePadding(
self.backbone._conv_stem.in_channels,
self.backbone._conv_stem.out_channels,
kernel_size=3,
stride=2, # lower stride ?
image_size=224,
dilation=2,
bias=False)
elif MODEL == "EFFICIENT":
if PRETRAIN:
model = EfficientNet.from_pretrained('efficientnet-b6', num_classes=555)
checkpoint = torch.load("experiment/efficient_pretrain_nabirds_model_1.pth")
model.load_state_dict(checkpoint)
model._fc = nn.Linear(2048, args.num_classes)
else:
# model = EFFICIENT(args.num_classes)
model = EfficientNet.from_pretrained('efficientnet-b6', num_classes=args.num_classes)
if args.model:
print("loading pretrained model")
checkpoint = torch.load(args.model)
model.load_state_dict(checkpoint)
if CHANNELS == "DOUBLE":
model._conv_stem = Conv2dStaticSamePadding(in_channels=3*2, out_channels=56, kernel_size=(3, 3), stride=2, image_size=(456, 456))
if CHANNELS == "TRIPLE":
model._conv_stem = Conv2dStaticSamePadding(in_channels=3*3, out_channels=56, kernel_size=(3, 3), stride=2, image_size=(456, 456))
elif MODEL == "RESNEXT":
model = torchvision.models.resnext50_32x4d(pretrained=True)
model.fc = nn.Sequential(
nn.Linear(2048, 256),
nn.ReLU(),
nn.Linear(256, 128),
nn.ReLU(),
nn.Linear(128, args.num_classes),
)
if FREEZE:
for name, param in model.named_parameters():