def Pnet(pretrained=True,
input_shape=(3, 12, 12),
freeze_features=True,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
pnet = ImageDetectionModel(input_shape=(3, 12, 12), output=p_net())
if pretrained == True:
download_model_from_google_drive('1w9ahipO8D9U1dAXMc2BewuL0UqIBYWSX',
dirname, 'pnet.pth')
recovery_model = fix_layer(load(os.path.join(dirname, 'pnet.pth')))
pnet.model = recovery_model
pnet.model.input_shape = input_shape
pnet.model.to(_device)
return pnet
def EfficientNetB1(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 240, 240),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 240, 240)
effb1 = EfficientNet(1.0,
1.1,
240,
0.2,
model_name='efficientnet-b1',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('1F3BtnAjmDz4G9RS9Q0hqU_K7WWXCni1G',
dirname, 'efficientnet-b1.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b1.pth')))
recovery_model.input_shape = input_shape
recovery_model.eval()
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb1.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb1.class_names = []
effb1.model = recovery_model
return effb1
def EfficientNetB6(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 528, 528),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 528, 528)
effb6 = EfficientNet(1.8,
2.6,
528,
0.5,
model_name='efficientnet-b6',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('1XJrKmcmMObN_nnjP2Z-YH_BQ3img58qF',
dirname, 'efficientnet-b6.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b6.pth')))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb6.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb6.class_names = []
effb6.model = recovery_model
return effb6
def EfficientNetB4(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 380, 380),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 380, 380)
effb4 = EfficientNet(1.4,
1.8,
380,
0.4,
model_name='efficientnet-b4',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('1X4ZOBR_ETRHZJeffJHvCmWTTy9_aW8SP',
dirname, 'efficientnet-b4.pth')
recovery_model = fix_layer(
load(sanitize_path(os.path.join(dirname, 'efficientnet-b4.pth'))))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb4.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb4.class_names = []
effb4.model = recovery_model
return effb4
def YoLoV4(pretrained=True,
freeze_features=True,
input_shape=(3, 608, 608),
classes=80,
**kwargs):
detector = YoloDetectionModel(input_shape=input_shape,
output=yolo4_body(classes, input_shape[-1]))
if pretrained:
download_model_from_google_drive('1CcbyinE8gQFjMjt05arSg2W0LLUwjsdt',
dirname,
'pretrained_yolov4_mscoco.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'pretrained_yolov4_mscoco.pth')))
detector.model = recovery_model
detector.model.input_shape = input_shape
detector.model.to(get_device())
return detector
def EfficientNetB2(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 260, 260),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 260, 260)
effb2 = EfficientNet(1.1,
1.2,
260,
0.3,
model_name='efficientnet-b2',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('1PjqhB7WJasF_hqOwYtSBNSXSGBY-cRLU',
dirname, 'efficientnet-b2.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b2.pth')))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb2.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb2.class_names = []
effb2.model = recovery_model
return effb2
def EfficientNetB3(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 300, 300),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 300, 300)
effb3 = EfficientNet(1.2,
1.4,
300,
0.3,
model_name='efficientnet-b3',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('11tMxdYdFfaEREwnESO4cwjtcoEB42zB_',
dirname, 'efficientnet-b3.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b3.pth')))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb3.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb3.class_names = []
effb3.model = recovery_model
return effb3
def EfficientNetB5(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 456, 456),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 456, 456)
effb5 = EfficientNet(1.6,
2.2,
456,
0.4,
model_name='efficientnet-b5',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('17iTD12G9oW3jYAui84MKtdY4gjd9vpgG',
dirname, 'efficientnet-b5.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b5.pth')))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb5.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb5.class_names = []
effb5.model = recovery_model
return effb5
def EfficientNetB7(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 600, 600),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 600, 600)
effb7 = EfficientNet(2.0,
3.1,
600,
0.5,
model_name='efficientnet-b7',
include_top=include_top,
num_classes=classes)
if pretrained == True:
download_model_from_google_drive('1M2DfvsNPRCWSo_CeXnUCQOR46rvOrhLl',
dirname, 'efficientnet-b7.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b7.pth')))
recovery_model.input_shape = input_shape
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
effb7.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
effb7.class_names = []
effb7.model = recovery_model
return effb7
def MobileNetV2(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 224, 224),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
mob = MobileNet(input_shape=(3, 224, 224),
classes=classes,
use_bias=False,
width_mult=1.0,
round_nearest=8,
include_top=include_top,
model_name='mobilenet')
if pretrained == True:
download_model_from_google_drive('1ULenXTjOO5PdT3fHv6N8bPXEfoJAn5yL',
dirname, 'mobilenet_v2.pth')
recovery_model = load(os.path.join(dirname, 'mobilenet_v2.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.eval()
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
mob.class_names = []
else:
if classes != 1000:
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
mob.class_names = []
mob.model = recovery_model
return mob
def ResNet50(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=None,
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
resnet50 = ResNet(bottleneck, [3, 4, 6, 3],
input_shape,
num_classes=classes,
include_top=include_top,
model_name='resnet50')
if pretrained == True:
download_model_from_google_drive(model_urls['resnet50'], dirname,
'resnet50.pth')
recovery_model = load(os.path.join(dirname, 'resnet50.pth'))
recovery_model = fix_layer(recovery_model)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
recovery_model.name = 'resnet50'
recovery_model.eval()
recovery_model.to(_device)
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
resnet50.class_names = []
else:
if classes != 1000:
resnet50.class_names = []
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.add_module(
'fc', Dense(classes, activation=None, name='fc'))
recovery_model.add_module('softmax', SoftMax())
resnet50.model = recovery_model
return resnet50
def SE_ResNet50(include_top=True,
pretrained=True,
freeze_features=True,
input_shape=None,
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
resnet50 = SE_ResNet(se_bottleneck, [3, 4, 6, 3],
input_shape,
num_classes=classes,
include_top=include_top,
model_name='se_resnet50')
if pretrained == True:
download_model_from_google_drive(model_urls['se_resnet50'], dirname,
'se_resnet50.pth')
recovery_model = load(os.path.join(dirname, 'se_resnet50.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.name = 'se_resnet50'
recovery_model = _make_recovery_model_include_top(
recovery_model,
include_top=include_top,
classes=classes,
default_shape=(3, 224, 224),
input_shape=input_shape,
freeze_features=freeze_features)
resnet50.model = recovery_model
else:
resnet50.model = _make_recovery_model_include_top(
resnet50.model,
include_top=include_top,
classes=classes,
default_shape=(3, 224, 224),
input_shape=input_shape,
freeze_features=True)
resnet50.model.input_shape = input_shape
resnet50.model.to(_device)
return resnet50
def MobileNetV2(include_top=True,
pretrained=True,
freeze_features=True,
input_shape=(3, 224, 224),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
mob = MobileNet(input_shape=(3, 224, 224),
classes=classes,
use_bias=False,
width_mult=1.0,
round_nearest=8,
include_top=include_top,
model_name='mobilenet')
if pretrained == True:
download_model_from_google_drive('1ULenXTjOO5PdT3fHv6N8bPXEfoJAn5yL',
dirname, 'mobilenet_v2.pth')
recovery_model = load(os.path.join(dirname, 'mobilenet_v2.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
mob.model = recovery_model
else:
mob.model = _make_recovery_model_include_top(mob.model,
include_top=include_top,
classes=classes,
freeze_features=True)
mob.model.input_shape = input_shape
mob.model.to(_device)
return mob
def SE_ResNet101(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=None,
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
resnet101 = SE_ResNet(se_bottleneck, [3, 4, 23, 3],
input_shape,
num_classes=classes,
include_top=include_top,
model_name='resnet101')
if pretrained == True:
download_model_from_google_drive('1BO92ONsWnLQPWs9JVS-4SdpoPo5M4ABR',
dirname, 'se_resnet101.pth')
recovery_model = load(os.path.join(dirname, 'se_resnet101.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.name = 'se_resnet101'
recovery_model.eval()
recovery_model.to(_device)
if freeze_features:
recovery_model.trainable = False
recovery_model.fc.trainable = True
if include_top == False:
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
recovery_model.remove_at(-1)
resnet101.class_names = []
else:
if classes != 1000:
resnet101.class_names = []
recovery_model.fc = Dense(classes, activation=None, name='fc')
recovery_model.fc.input_shape = recovery_model.avg_pool.output_shape
resnet101.model = recovery_model
return resnet101
def on_batch_end(self, training_context):
# Log the learning rate
lr = training_context['optimizer'].lr
self.lrs.append(lr)
loss = training_context['current_loss'].item()
self.losses.append(loss)
training_context['current_model'].load_state_dict(
load('Models/state_dict.pth'))
# Check whether the loss got too large or NaN
if math.isnan(loss) or loss > self.best_loss * 4:
self.adjust_learning_rate(training_context, self.base_lr)
if loss < self.best_loss:
self.best_loss = loss
# Increase the learning rate for the next batch
lr *= self.lr_mult
self.adjust_learning_rate(
training_context,
lr,
)
def ResNet18(include_top=True,
pretrained=True,
freeze_features=True,
input_shape=None,
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
resnet18 = ResNet(basic_block, [2, 2, 2, 2],
input_shape,
use_bias=False,
include_top=include_top,
model_name='resnet18')
if pretrained == True:
download_model_from_google_drive(model_urls['resnet18'], dirname,
'resnet18.pth')
recovery_model = load(os.path.join(dirname, 'resnet18.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
resnet18.model = recovery_model
else:
resnet18.model = _make_recovery_model_include_top(
resnet18.model,
include_top=include_top,
classes=classes,
freeze_features=True)
resnet18.model.input_shape = input_shape
resnet18.model.to(_device)
return resnet18
def EfficientNetB7(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 600, 600),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 600, 600)
effb7 = EfficientNet(2.0,
3.1,
input_shape,
0.5,
model_name='efficientnet-b7',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1M2DfvsNPRCWSo_CeXnUCQOR46rvOrhLl',
dirname, 'efficientnet-b7.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b7.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb7.model = recovery_model
else:
effb7.model = _make_recovery_model_include_top(effb7.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb7.model.input_shape = input_shape
effb7.model.to(get_device())
return effb7
def EfficientNetB5(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 456, 456),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 456, 456)
effb5 = EfficientNet(1.6,
2.2,
input_shape,
0.4,
model_name='efficientnet-b5',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('17iTD12G9oW3jYAui84MKtdY4gjd9vpgG',
dirname, 'efficientnet-b5.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b5.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb5.model = recovery_model
else:
effb5.model = _make_recovery_model_include_top(effb5.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb5.model.input_shape = input_shape
effb5.model.to(get_device())
return effb5
def EfficientNetB6(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 528, 528),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 528, 528)
effb6 = EfficientNet(1.8,
2.6,
input_shape,
0.5,
model_name='efficientnet-b6',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1XJrKmcmMObN_nnjP2Z-YH_BQ3img58qF',
dirname, 'efficientnet-b6.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b6.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb6.model = recovery_model
else:
effb6.model = _make_recovery_model_include_top(effb6.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb6.model.input_shape = input_shape
effb6.model.to(get_device())
return effb6
def EfficientNetB3(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 300, 300),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 300, 300)
effb3 = EfficientNet(1.2,
1.4,
input_shape,
0.3,
model_name='efficientnet-b3',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('11tMxdYdFfaEREwnESO4cwjtcoEB42zB_',
dirname, 'efficientnet-b3.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b3.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb3.model = recovery_model
else:
effb3.model = _make_recovery_model_include_top(effb3.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb3.model.input_shape = input_shape
effb3.model.to(get_device())
return effb3
def EfficientNetB4(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 380, 380),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 380, 380)
effb4 = EfficientNet(1.4,
1.8,
input_shape,
0.4,
model_name='efficientnet-b4',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1X4ZOBR_ETRHZJeffJHvCmWTTy9_aW8SP',
dirname, 'efficientnet-b4.pth')
recovery_model = fix_layer(
load(sanitize_path(os.path.join(dirname, 'efficientnet-b4.pth'))))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb4.model = recovery_model
else:
effb4.model = _make_recovery_model_include_top(effb4.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb4.model.input_shape = input_shape
effb4.model.to(get_device())
return effb4
def EfficientNetB2(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 260, 260),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 260, 260)
effb2 = EfficientNet(1.1,
1.2,
input_shape,
0.3,
model_name='efficientnet-b2',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1PjqhB7WJasF_hqOwYtSBNSXSGBY-cRLU',
dirname, 'efficientnet-b2.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b2.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb2.model = recovery_model
else:
effb2.model = _make_recovery_model_include_top(effb2.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb2.model.input_shape = input_shape
effb2.model.to(get_device())
return effb2
def EfficientNetB1(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 240, 240),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 240, 240)
effb1 = EfficientNet(1.0,
1.1,
input_shape,
0.2,
model_name='efficientnet-b1',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1F3BtnAjmDz4G9RS9Q0hqU_K7WWXCni1G',
dirname, 'efficientnet-b1.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b1.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb1.model = recovery_model
else:
effb1.model = _make_recovery_model_include_top(effb1.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb1.model.input_shape = input_shape
effb1.model.to(get_device())
return effb1
def EfficientNetB0(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3, 224, 224),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
effb0 = EfficientNet(1.0,
1.0,
input_shape,
0.2,
model_name='efficientnet-b0',
include_top=include_top,
num_classes=classes)
if pretrained:
download_model_from_google_drive('1bxnoDerzoNfiZZLft4ocD3DAgx4v6aTN',
dirname, 'efficientnet-b0.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'efficientnet-b0.pth')))
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
effb0.model = recovery_model
else:
effb0.model = _make_recovery_model_include_top(effb0.model,
include_top=include_top,
classes=classes,
freeze_features=False)
effb0.model.input_shape = input_shape
effb0.model.to(get_device())
return effb0
def SEResNet_IR_50_512(include_top=True,
pretrained=True,
freeze_features=False,
input_shape=(3,112,112),
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape)==3:
input_shape=tuple(input_shape)
else:
input_shape=(3, 112, 112)
seresnet = SEResNet_IR(include_top=include_top,num_layers=50,Bottleneck=BottleNeck_IR_SE,drop_ratio=0.4,feature_dim=512,input_shape=input_shape)
if pretrained == True:
download_model_from_google_drive('1aLYbFvtvsV2gQ16D_vwzrKdbCgij7IoZ', dirname, 'arcface_se_50_512.pth')
recovery_model = load(os.path.join(dirname, 'arcface_se_50_512.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.name = 'arcface_se_50_512'
recovery_model = _make_recovery_model_include_top(recovery_model, include_top=include_top, classes=classes, freeze_features=freeze_features)
seresnet.model = recovery_model
else:
seresnet.model = _make_recovery_model_include_top(seresnet.model, include_top=include_top, classes=classes, freeze_features=False)
seresnet.model.input_shape = input_shape
seresnet.model.to(_device)
return seresnet
def SE_ResNet101(include_top=True,
pretrained=True,
freeze_features=True,
input_shape=None,
classes=1000,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 224, 224)
resnet101 = SE_ResNet(se_bottleneck, [3, 4, 23, 3],
input_shape,
num_classes=classes,
include_top=include_top,
model_name='resnet101')
if pretrained == True:
download_model_from_google_drive('1BO92ONsWnLQPWs9JVS-4SdpoPo5M4ABR',
dirname, 'se_resnet101.pth')
recovery_model = load(os.path.join(dirname, 'se_resnet101.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.name = 'se_resnet101'
recovery_model = _make_recovery_model_include_top(
recovery_model,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
resnet101.model = recovery_model
else:
resnet101.model = _make_recovery_model_include_top(
resnet101.model,
include_top=include_top,
classes=classes,
freeze_features=True)
resnet101.model.input_shape = input_shape
resnet101.model.to(_device)
return resnet101
def RfbNet(include_top=True,
pretrained=True,
input_shape=None,
base_filters=16,
num_classes=1,
num_regressors=4,
**kwargs):
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
else:
input_shape = (3, 480, 640)
rfbnet = SsdDetectionModel(input_shape=(3, 480, 640),
output=RFBnet(base_filters=base_filters,
num_classes=num_classes,
num_regressors=num_regressors))
rfbnet.detection_threshold = 0.7
rfbnet.nms_threshold = 0.7
rfbnet.palette[0] = (128, 255, 128)
rfbnet.palette[1] = (128, 255, 128)
rfbnet.preprocess_flow = [
Resize((480, 640), True),
Normalize(127.5, 127.5)
]
if pretrained == True:
download_model_from_google_drive('1T_0VYOHaxoyuG1fAxY-6g0C7pfXiujns',
dirname, 'version-RFB-640.pth')
recovery_model = fix_layer(
load(os.path.join(dirname, 'version-RFB-640.pth')))
priors = recovery_model.priors.clone()
recovery_model.__delattr__("priors")
recovery_model.register_buffer("priors", priors)
recovery_model.name = 'rfb640'
recovery_model.eval()
recovery_model.to(_device)
rfbnet.model = recovery_model
return rfbnet
def DenseNet201(include_top=True,
pretrained=True,
freeze_features=True,
input_shape=(3, 224, 224),
classes=1000,
**kwargs):
"""
Constructor the image classicication model with DenseNet201 as backbond
Args:
freeze_features ():
include_top ():
pretrained (bool): If True, returns a model pre-trained on ImageNet.
input_shape (tuple or list): the default input image size in CHW order (C, H, W)
classes (int): number of classes
References
Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf
Returns:
the image classicication model with DenseNet201
Examples:
>>> dense201 = DenseNet201(include_top=True,pretrained=True,input_shape=(3,224,224),classes=1000)
>>> 'n02124075' in dense201.infer_single_image(get_image_from_google_drive('1SwablQsZO8mBuB84xnr1IoOisE3pm03l'),1).key_list[0]
True
"""
if input_shape is not None and len(input_shape) == 3:
input_shape = tuple(input_shape)
densenet201 = DenseNet([6, 12, 48, 32],
32,
64,
include_top=include_top,
pretrained=True,
input_shape=input_shape,
num_classes=classes,
name='densenet201')
if pretrained == True:
download_model_from_google_drive('1V2JazzdnrU64lDfE-O4bVIgFNQJ38q3J',
dirname, 'densenet201.pth')
recovery_model = load(os.path.join(dirname, 'densenet201.pth'))
recovery_model = fix_layer(recovery_model)
recovery_model.name = 'densenet201'
recovery_model = _make_recovery_model_include_top(
recovery_model,
input_shape=input_shape,
include_top=include_top,
classes=classes,
freeze_features=freeze_features)
densenet201.model = recovery_model
else:
densenet201.model = _make_recovery_model_include_top(
densenet201.model,
include_top=include_top,
classes=classes,
freeze_features=True)
densenet201.model.input_shape = input_shape
densenet201.model.to(_device)
return densenet201
