def test_inceptionv4_model(input_var, pool, assert_equal_outputs):
original_model = pretrainedmodels.inceptionv4(pretrained='imagenet',
num_classes=1000)
finetune_model = make_model(
'inception_v4',
num_classes=1000,
pool=pool,
pretrained=True,
)
copy_module_weights(original_model.last_linear, finetune_model._classifier)
assert_equal_outputs(input_var, original_model, finetune_model)
def test_nasnetalarge_model(input_var, pool, assert_equal_outputs):
original_model = pretrainedmodels.nasnetalarge(pretrained='imagenet',
num_classes=1000)
finetune_model = make_model(
'nasnetalarge',
num_classes=1000,
pool=pool,
pretrained=True,
)
copy_module_weights(original_model.last_linear, finetune_model._classifier)
assert_equal_outputs(input_var, original_model, finetune_model)
def construct_dcnn(model_name):
global model, optimizer, scheduler, criterion
model = make_model(model_name, num_classes=len(classes),
pretrained=args.pre_trained_model,
dropout_p=args.dropout_p,
input_size = input_size)
model = model.to(device)
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum,
weight_decay=args.weight_decay, nesterov=use_nestrov_moment) # dampening = 0.005
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, args.lr_milestones, gamma= .95)
criterion = nn.CrossEntropyLoss()
def get_net():
net = make_model('se_resnext101_32x4d', num_classes=2, pretrained=True)
# net = EfficientNet.from_pretrained('efficientnet-b6')
# net._fc = nn.Linear(in_features=2304, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=2048, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=1792, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=1536, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=1408, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=1280, out_features=2, bias=True)
#net._fc = nn.Linear(in_features=1280, out_features=2, bias=True)
return net
def test_senet_models(input_var, model_name, pool, assert_equal_outputs):
original_model = getattr(pretrainedmodels,
model_name)(pretrained='imagenet',
num_classes=1000)
finetune_model = make_model(
model_name,
num_classes=1000,
pool=pool,
pretrained=True,
)
copy_module_weights(original_model.last_linear, finetune_model._classifier)
assert_equal_outputs(input_var, original_model, finetune_model)
def test_load_state_dict():
torch.manual_seed(42)
model1 = make_model('resnet18', num_classes=10)
model1_state = model1.state_dict()
torch.manual_seed(84)
model2 = make_model('resnet18', num_classes=10)
model2_state = model2.state_dict()
assert not all(
torch.equal(weights1, weights2) for weights1, weights2
in zip(model1_state.values(), model2_state.values())
)
model2.load_state_dict(model1_state)
model2_state = model2.state_dict()
assert all(
torch.equal(weights1, weights2) for weights1, weights2 in
zip(model1_state.values(), model2_state.values())
)
def test_state_dict_features_and_classifier():
model = make_model('resnet18', num_classes=10)
model_state_keys = model.state_dict().keys()
assert '_classifier.weight' in model_state_keys
assert '_classifier.bias' in model_state_keys
features_keys = [
key for key in model_state_keys
if key.startswith('_features') and key.endswith(('weight', 'bias', 'running_mean', 'running_var'))
]
assert len(features_keys) == 100
def __init__(self, extractor_name):
super(Extractor, self).__init__()
basenet = make_model(model_name=extractor_name,
num_classes=1000,
input_size=(224, 224),
pretrained=True)
# print(basenet)
self.n_features = basenet._classifier.in_features
self.pool = GeM()
self.extractor = basenet._features
def test_models_without_pretrained_weights_dont_have_model_info(model_name):
input_size = None
if (model_name == 'alexnet' or model_name.startswith('vgg')
or model_name.startswith('squeezenet')):
input_size = (224, 224)
model = make_model(
model_name,
pretrained=False,
num_classes=10,
dropout_p=0.5,
input_size=input_size,
)
assert model.original_model_info is None
def __init__(self, encoder, pretrained, num_classes, activation):
super().__init__()
self.net = make_model(model_name=encoder,
pretrained=pretrained,
num_classes=1000)
in_features = self.net._classifier.in_features
if activation == "Mish":
to_Mish(self.net)
print("Mish activation added!")
self.head_grapheme_root = Head(in_features, num_classes[0])
self.head_vowel_diacritic = Head(in_features, num_classes[1])
self.head_consonant_diacritic = Head(in_features, num_classes[2])
print(self.net)
def cnn(image):
global warray
PATH = 'direct2642_noisy_resnet2.pt' #direct2642_noisy_resnet2.pt, direct2642_noisy.pt
# If RoS is being used
# listener()
# model,saber,imgtopic = initialize()
model = make_model('resnet18', num_classes=1, pretrained=True, input_size=(227, 227))
if torch.cuda.is_available():
device = torch.device("cuda")
use_gpu=1
print ("CUDA Available")
else:
device = torch.device("cpu")
if use_gpu == 1:
model = nn.DataParallel(model).cuda()
if path.exists(PATH):
model.load_state_dict(torch.load(PATH))
else:
print("Did not find Model!")
model = model.to(device)
# print ("\nInitialization complete")
# print ("q value is {}".format(q.get()))
# transform = transforms.Compose([
# transforms.ToPILImage(),
# transforms.Resize((227,227)),
# transforms.ToTensor(),
# ])
if image.shape[2]>3:
image = image[:,:,:3]
image = image_loader(image)
model.eval()
with torch.no_grad():
w = model(image)
# print ("Before {}".format(w))
w = ((w.cpu().numpy().flatten()))
w = w*0.7897025648 - 0.0967433495
return w
def build_model(args, gpus):
# Build the model to run
print("Building a model...")
if args.task == 1:
#from se_resnet import se_resnet_custom
#model = nn.DataParallel(se_resnet_custom(size=args.model_num_blocks,
# dropout_p=args.dropout_p, num_classes=23),
# device_ids=gpus)
pass # TODO make model here
elif args.task == 2:
# TODO make this use MSE and have 3 heads, one for X,Y,Z
model = make_model(args.model,
num_classes=3,
dropout_p=args.dropout_p,
pretrained=True)
elif args.task == 3 or args.task == 4:
model = make_model(args.model,
num_classes=23,
dropout_p=args.dropout_p,
pretrained=True)
#model = make_model('resnet18', num_classes=23, dropout_p=args.dropout_p, pretrained=True)
#model = make_model('resnext101_32x4d', num_classes=23, dropout_p=args.dropout_p, pretrained=True)
return model
def initialize():
global saber
PATH = 'direct2642_noisy_resnet2.pt' #direct2642_noisy_resnet2.pt, direct2642_noisy.pt
quality = "qhd"
print("\n Initializing Kinect ...")
# command="roslaunch kinect2_bridge kinect2_bridge.launch"
# os.system("gnome-terminal -e 'bash -c \"roslaunch kinect2_bridge kinect2_bridge.launch; exec bash\"'")
imgtopic = "/kinect2/{}/image_color".format(quality)
print("\nDetecting sabertooth....\n")
pl = list(port.comports())
print(pl)
address = ''
for p in pl:
print(p)
if 'Sabertooth' in str(p):
address = str(p).split(" ")
print("\nAddress found @")
print(address[0])
saber = Sabertooth(address[0], baudrate=9600, address=128, timeout=0.1)
if torch.cuda.is_available():
device = torch.device("cuda")
use_gpu = 1
print("CUDA Available")
else:
device = torch.device("cpu")
model = cnn_finetune.make_model('resnet18',
num_classes=1,
pretrained=True,
input_size=(227, 227))
if use_gpu == 1:
model = nn.DataParallel(model).cuda()
if path.exists(PATH):
model.load_state_dict(torch.load(PATH))
model = model.to(device)
print("\n Initialization complete")
return model, saber, imgtopic
def test_dpn_models(input_var, model_name, pool, assert_equal_outputs):
if model_name in {'dpn68b', 'dpn92', 'dpn107'}:
pretrained = 'imagenet+5k'
else:
pretrained = 'imagenet'
original_model = getattr(pretrainedmodels,
model_name)(pretrained=pretrained,
num_classes=1000)
finetune_model = make_model(
model_name,
num_classes=1000,
pool=pool,
pretrained=True,
)
copy_module_weights(original_model.classifier, finetune_model._classifier)
assert_equal_outputs(input_var, original_model, finetune_model)
def __init__(self, cnn_finetune, num_country, country_emb_dim=10):
super().__init__()
num_classes = cnn_finetune['num_classes']
cnn_finetune = make_model(**cnn_finetune)
self.features = cnn_finetune._features
self.pool = cnn_finetune.pool
self.dropout = cnn_finetune.dropout
img_in_features = cnn_finetune._classifier.in_features
self.country_emb = nn.Embedding(num_country, country_emb_dim)
self.fc_1 = nn.Linear(img_in_features + country_emb_dim, 512)
self.fc_2 = nn.Linear(512, num_classes)
self.relu = nn.ReLU(inplace=True)
def load_model(self, modelname="resnet50", pretrained=True):
"""load model return model"""
if not modelname: modelname = "resnet50"
model = make_model(modelname,
num_classes=self.num_classes,
pretrained=pretrained)
#set model mode
model.train()
model = model.to(self.device)
if "cuda" in str(self.device):
model = torch.nn.DataParallel(model)
self.model = model
print(modelname)
return model
def get_model(model: str,
num_classes: int,
pretrained: bool,
input_size: int,
use_cuda: bool = None) -> nn.Module:
if model.startswith('efficientnet'):
model = EfficientNet.from_name(model)
model._fc = nn.Linear(1536, num_classes)
else:
model = make_model(model,
num_classes,
pretrained,
input_size=(input_size, input_size),
pool=AvgPool())
if use_cuda:
model = model.cuda()
return model
def __init__(self, cnn_finetune, arcface, embedding_size):
super().__init__()
num_classes = cnn_finetune['num_classes']
cnn_finetune = make_model(**cnn_finetune)
self.features = cnn_finetune._features
self.pool = cnn_finetune.pool
self.dropout = cnn_finetune.dropout
in_features = cnn_finetune._classifier.in_features
self.fc_1 = nn.Linear(in_features, 1024)
self.fc_2 = nn.Linear(1024, embedding_size)
self.relu = nn.ReLU(inplace=True)
self.bn = nn.BatchNorm1d(embedding_size)
self.arcface = ArcMarginProduct(in_features=embedding_size,
out_features=num_classes,
**arcface)
def initialize():
global rdeftrain, rdeftest, tta, toa, model_ft_best, model_ft
rdeftrain = 0
rdeftest = 0
tta = []
toa = []
model_ft = make_model('resnet18',
num_classes=1,
pretrained=True,
input_size=(224, 224)) #,dropout_p=0.5)
model_ft_best = copy.deepcopy(model_ft)
model_ft = model_ft.to(device)
model_ft_best = model_ft_best.to(device)
if use_gpu == 1:
model_ft = nn.DataParallel(model_ft).cuda()
model_ft_best = nn.DataParallel(model_ft_best).cuda()
def test_every_pretrained_model_has_model_info(model_name):
input_size = get_default_input_size_for_model(model_name)
model = make_model(
model_name,
pretrained=True,
num_classes=10,
dropout_p=0.5,
input_size=input_size,
)
original_model_info = model.original_model_info
assert original_model_info
assert_iterable_length_and_type(original_model_info.input_space, 3, str)
assert_iterable_length_and_type(original_model_info.input_size, 3, int)
assert_iterable_length_and_type(original_model_info.input_range, 2,
(int, float))
assert_iterable_length_and_type(original_model_info.mean, 3, (int, float))
assert_iterable_length_and_type(original_model_info.std, 3, (int, float))
def __init__(
self,
arch="se_resnet50",
n_class=6,
pretrained=True,
image_size=256,
):
super(Finetune, self).__init__()
self.model = make_model(
model_name=arch,
num_classes=n_class,
pretrained=pretrained,
input_size=(image_size, image_size),
)
self.extract_feature = False
self.model_name = arch
self.n_class = n_class
def train():
# GPU
device = torch.device("cuda" if GPU else "cpu")
# model
model = make_model('vgg16', num_classes=num_classes, pretrained=True, input_size=(img_height, img_width))
model = model.to(device)
opt = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
model.train()
xs, ts, paths = data_load('../Dataset/train/images/', hf=True, vf=True)
# training
mb = 16
mbi = 0
train_ind = np.arange(len(xs))
np.random.seed(0)
np.random.shuffle(train_ind)
for i in range(100):
if mbi + mb > len(xs):
mb_ind = train_ind[mbi:]
np.random.shuffle(train_ind)
mb_ind = np.hstack((mb_ind, train_ind[:(mb-(len(xs)-mbi))]))
else:
mb_ind = train_ind[mbi: mbi+mb]
mbi += mb
x = torch.tensor(xs[mb_ind], dtype=torch.float).to(device)
t = torch.tensor(ts[mb_ind], dtype=torch.long).to(device)
opt.zero_grad()
y = model(x)
y = F.log_softmax(y, dim=1)
loss = torch.nn.CrossEntropyLoss()(y, t)
loss.backward()
opt.step()
pred = y.argmax(dim=1, keepdim=True)
acc = pred.eq(t.view_as(pred)).sum().item() / mb
print("iter >>", i+1, ',loss >>', loss.item(), ',accuracy >>', acc)
torch.save(model.state_dict(), 'cnn.pt')
def __init__(
self,
model_name='resnet50',
num_classes=6,
pretrained=True,
n_channels=4,
):
super(DSResnet, self).__init__()
self.model = make_model(model_name=model_name,
num_classes=num_classes,
pretrained=True,
dropout_p=0.3)
# print(self.model)
conv1 = self.model._features[0]
self.model._features[0] = nn.Conv2d(in_channels=n_channels,
out_channels=conv1.out_channels,
kernel_size=conv1.kernel_size,
stride=conv1.stride,
padding=conv1.padding,
bias=conv1.bias)
# copy pretrained weights
self.model._features[0].weight.data[:, :3, :, :] = conv1.weight.data
self.model._features[
0].weight.data[:, 3:n_channels, :, :] = conv1.weight.data[:, :int(
n_channels - 3), :, :]
# self.deepsuper_1 = nn.Sequential(
# nn.AdaptiveAvgPool2d(),
# Flatten(),
# nn.BatchNorm1d(256),
# nn.Linear(256, num_classes)
# )
self.deepsuper_2 = nn.Sequential(nn.AdaptiveAvgPool2d(1), Flatten(),
nn.BatchNorm1d(512),
nn.Linear(512, num_classes))
self.deepsuper_3 = nn.Sequential(nn.AdaptiveAvgPool2d(1), Flatten(),
nn.BatchNorm1d(1024),
nn.Linear(1024, num_classes))
self.is_infer = False
def __init__(self,
model_name="seresnext26_32x4d",
num_classes=1108,
n_channels=6,
num_samples=4,
weight=None):
super(SENetCellMultipleDropout, self).__init__()
self.model = make_model(model_name=model_name,
num_classes=31,
pretrained=True)
print("*" * 100)
print("SENetGrouplevel")
conv1 = self.model._features[0].conv1
self.model._features[0].conv1 = nn.Conv2d(
in_channels=n_channels,
out_channels=conv1.out_channels,
kernel_size=conv1.kernel_size,
stride=conv1.stride,
padding=conv1.padding,
bias=conv1.bias)
# copy pretrained weights
self.model._features[
0].conv1.weight.data[:, :3, :, :] = conv1.weight.data
self.model._features[
0].conv1.weight.data[:, 3:
n_channels, :, :] = conv1.weight.data[:, :int(
n_channels - 3), :, :]
if weight:
model_state_dict = torch.load(weight)['model_state_dict']
self.model.load_state_dict(model_state_dict)
print(
f"\n\n******************************* Loaded checkpoint {weight}"
)
in_features = self.model._classifier.in_features
self.num_samples = num_samples
self.classifier = nn.Linear(in_features, num_classes)
def test():
device = torch.device("cuda" if GPU else "cpu")
model = make_model('vgg16', num_classes=num_classes, pretrained=True, input_size=(img_height, img_width))
model = model.to(device)
model.eval()
model.load_state_dict(torch.load('cnn.pt'))
xs, ts, paths = data_load('../Dataset/test/images/')
for i in range(len(paths)):
x = xs[i]
t = ts[i]
path = paths[i]
x = np.expand_dims(x, axis=0)
x = torch.tensor(x, dtype=torch.float).to(device)
pred = model(x)
pred = F.softmax(pred, dim=1).detach().cpu().numpy()[0]
print("in {}, predicted probabilities >> {}".format(path, pred))
def main(args):
if args.model_name == 'mobilenetv2':
model = MobileNetV2(num_classes=1000)
elif args.model_name == 'mobilenetv3large':
model = MobileNetV3Large(n_classes=1000)
elif args.model_name == 'mobilenetv3small':
model = MobileNetV3Small(n_classes=1000)
else:
model = make_model(args.model_name, num_classes=1000, pretrained=True)
input_size = (1, ) + tuple(model.original_model_info.input_size)
x_dummy = torch.rand(input_size)
input_names = ['input_0']
output_names = ['output_0']
torch.onnx.export(model,
x_dummy,
args.model_save_path,
export_params=True,
input_names=input_names,
output_names=output_names)
def __init__(self, grad_layer, num_classes):
super(GAIN, self).__init__()
self.model = make_model(model_name='resnet50',
pretrained=True,
num_classes=num_classes)
# print(self.model)
self.grad_layer = grad_layer
self.num_classes = num_classes
# Feed-forward features
self.feed_forward_features = None
# Backward features
self.backward_features = None
# Register hooks
self._register_hooks(grad_layer)
# sigma, omega for making the soft-mask
self.sigma = 0.25
self.omega = 100
def test_every_pretrained_model_has_model_info(model_name):
input_size = None
if (model_name == 'alexnet' or model_name.startswith('vgg')
or model_name.startswith('squeezenet')):
input_size = (224, 224)
model = make_model(
model_name,
pretrained=True,
num_classes=10,
dropout_p=0.5,
input_size=input_size,
)
original_model_info = model.original_model_info
assert original_model_info
assert_iterable_length_and_type(original_model_info.input_space, 3, str)
assert_iterable_length_and_type(original_model_info.input_size, 3, int)
assert_iterable_length_and_type(original_model_info.input_range, 2,
(int, float))
assert_iterable_length_and_type(original_model_info.mean, 3, (int, float))
assert_iterable_length_and_type(original_model_info.std, 3, (int, float))
def __init__(self,
model_name,
in_channels=3,
out_dim=10,
hdim=512,
use_dropout=True,
activation=F.leaky_relu,
use_bn=True,
pretrained=False,
kernel_size=3,
stride=1,
padding=1):
super(PretrainedCNN, self).__init__()
print("Architecture: ", model_name)
is_remote = os.getcwd() != "/kaggle/working"
if model_name in get_official_names(
) or model_name in get_pretrained_names():
self.base_model = cnn_finetune.make_model(model_name,
num_classes=out_dim,
pretrained=is_remote
and pretrained)
elif model_name in get_timm_names():
self.base_model = timm.create_model(model_name,
num_classes=out_dim,
pretrained=is_remote
and pretrained)
else:
print("Not supported architecture")
assert False
if not use_dropout:
dropout_replace(self.base_model)
self.conv0 = nn.Conv2d(in_channels,
3,
kernel_size=kernel_size,
stride=stride,
padding=padding,
bias=True)
def __init__(self,
arch="resnet18",
n_class=10,
pretrained=True,
num_embeddings=750,
embedding_dim=10):
super(BaselineModel, self).__init__()
resnet = make_model(
model_name=arch,
num_classes=n_class,
pretrained=pretrained,
)
print(resnet)
self.encoder1 = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),
nn.BatchNorm2d(64),
nn.ReLU(inplace=True),
)
self.encoder2 = nn.Sequential(
nn.MaxPool2d(kernel_size=2, stride=2),
resnet._features[1],
)
self.encoder3 = resnet._features[2]
self.encoder4 = resnet._features[3]
self.encoder5 = resnet._features[4]
# Embedding layers
self.embedding = nn.Sequential(
nn.Embedding(num_embeddings=num_embeddings + 1,
embedding_dim=embedding_dim),
nn.Dropout(0.25),
GlobalConcatPool1d(),
)
self.logit = nn.Linear(2068, n_class)