def __init__(self,
mname,
outtoken,
hidden=128,
enc_layers=1,
dec_layers=1,
nhead=1,
dropout=0.1,
pretrained=False):
super(TransformerModel, self).__init__()
self.backbone = models.__getattribute__(mname)(pretrained=pretrained)
self.backbone.fc = nn.Conv2d(2048, hidden // 4, 1)
self.pos_encoder = PositionalEncoding(hidden, dropout)
self.decoder = nn.Embedding(outtoken, hidden)
self.pos_decoder = PositionalEncoding(hidden, dropout)
self.transformer = nn.Transformer(d_model=hidden,
nhead=nhead,
num_encoder_layers=enc_layers,
num_decoder_layers=dec_layers,
dim_feedforward=hidden * 4,
dropout=dropout,
activation='relu')
self.fc_out = nn.Linear(hidden, outtoken)
self.src_mask = None
self.trg_mask = None
self.memory_mask = None
def __init__(self,
base_model_name: str,
config: Config,
pretrained=False,
num_classes=264):
super(ResNet, self).__init__()
base_model = models.__getattribute__(base_model_name)(
pretrained=pretrained)
layers = list(base_model.children())[:-2]
layers.append(nn.AdaptiveMaxPool2d(1))
self.encoder = nn.Sequential(*layers)
in_features = base_model.fc.in_features
self.config = config
if config.cbr:
self.cbr1 = CBR(1, 3, 5, 1)
if config.loss_type == 'bce' or config.loss_type == 'focal':
self.classifier = nn.Sequential(nn.Linear(in_features, 1024),
nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(1024, 1024), nn.ReLU(),
nn.Dropout(p=0.2),
nn.Linear(1024, num_classes))
elif config.loss_type == 'arcface':
self.classifier = nn.Sequential(
nn.Linear(in_features, 1024), nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(1024, 1024), nn.ReLU(), nn.Dropout(p=0.2),
ArcMarginProduct(1024, num_classes))
def create_model(arch):
"""
Creates NN model
Parameters:
- arch
Return
- NN model
"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
arch = 'vgg11'
model = models.__getattribute__(arch)(pretrained=True)
# Freeze parameters so we don't backprop through them
for param in model.parameters():
param.requires_grad = False
from collections import OrderedDict
classifier = nn.Sequential(
OrderedDict([('fc1', nn.Linear(25088, 4096)), ('relu1', nn.ReLU()),
('dropout', nn.Dropout(p=0.5)),
('fc2', nn.Linear(4096, 102)),
('output', nn.LogSoftmax(dim=1))]))
model.classifier = classifier
return model
def load_checkpoint(filepath):
checkpoint = torch.load(filepath)
arch = checkpoint['arch']
loaded_model = models.__getattribute__(arch)(pretrained=True)
for param in loaded_model.parameters():
param.requires_grad = False
loaded_model.fc = checkpoint['fc']
loaded_model.load_state_dict(checkpoint['state_dict'])
loaded_model.class_to_idx = checkpoint['class_to_idx']
optimizer = checkpoint['optimizer_dict']
epochs = checkpoint['epochs']
criterion = checkpoint['criterion']
return loaded_model, optimizer, criterion, epochs
def __init__(self, name, num_classes=265):
super(BirdcallNet, self).__init__()
self.model = models.__getattribute__(name)(pretrained=True)
if name in ["resnet50","resnext50_32x4d"]:
self.model.fc = nn.Linear(2048, num_classes)
elif name in ['resnet18','resnet34']:
self.model.fc = nn.Linear(512, num_classes)
elif name =="densenet121":
self.model.classifier = nn.Linear(1024, num_classes)
elif name in ['alexnet','vgg16']:
self.model.classifier[-1] = nn.Linear(4096, num_classes)
elif name =="mobilenet_v2":
self.model.classifier[1] = nn.Linear(1280, num_classes)
def main():
allowed_models = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152']
parser = argparse.ArgumentParser(description='Train NN')
parser.add_argument('data_dir', help='directory containing sub-folders with data')
parser.add_argument('--save_dir', help='directory for saving checkpoint', default='checkpoints')
parser.add_argument('--arch', help='pre-trained model architecture', default='resnet18', choices=allowed_models)
parser.add_argument('--learning_rate', help='learning rate during learning', type=float, default=0.01)
parser.add_argument('--dropout', help='dropout during learning', type=float, default=0.05)
parser.add_argument('--hidden_units', help='List of number of nodes in hidden layers', nargs='+', type=int, default=[256, 128])
parser.add_argument('--epochs', help='Number of epochs for training', default=3, type=int)
parser.add_argument('--gpu', help='Enable GPU', action='store_true')
args = parser.parse_args()
# Describe directories relative to working directory
data_dir = args.data_dir
train_dir = data_dir + '/train'
valid_dir = data_dir + '/valid'
test_dir = data_dir + '/test'
save_dir = args.save_dir
model_arch = args.arch
model_hidden_units = args.hidden_units
learning_rate = args.learning_rate
drop = args.dropout
print('Data directory: ' + data_dir)
print('hidden units: ' + str(args.hidden_units))
print('Save directory: ' + save_dir)
print('Architecture: ' + args.arch)
fu.create_directory(save_dir)
model = models.__getattribute__(model_arch)(pretrained=True)
for param in model.parameters():
param.requires_grad = False
model.fc = fu.Network(model.fc.in_features, 102, model_hidden_units, drop)
criterion = nn.NLLLoss()
optimizer = optim.Adam(model.fc.parameters(), lr=learning_rate)
device = torch.device('cuda' if torch.cuda.is_available() and args.gpu==True else 'cpu')
print('device: ', device)
epochs = args.epochs
print_every = 50
running_loss = 0
steps = 0
train_loader, test_loader, valid_loader, train_data, test_data, valid_data = load_transform.load_transform(data_dir, train_dir, valid_dir, test_dir)
fu.train(device, model, epochs, criterion, optimizer, print_every, train_loader, test_loader, valid_loader)
fu.save_checkpoint(model, model_arch, epochs, criterion, optimizer, train_data, save_dir)
return model, test_loader, criterion
def __init__(self, base_model_name: str, pretrained=False,
num_classes=6):
super().__init__()
base_model = models.__getattribute__(base_model_name)(
pretrained=pretrained)
layers = list(base_model.children())[:-2]
layers.append(nn.AdaptiveMaxPool2d(1))
self.encoder = nn.Sequential(*layers)
in_features = base_model.fc.in_features
self.classifier = nn.Sequential(
nn.Linear(in_features, 1024), nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(1024, 1024), nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(1024, num_classes))
def __init__(self,
base_model_name: str,
pretrained=False,
num_classes=400,
embed=False):
super().__init__()
base_model = models.__getattribute__(base_model_name)(
pretrained=pretrained)
layers = list(base_model.children())[:-1]
layers.append(nn.AdaptiveMaxPool2d(1))
self.encoder = nn.Sequential(*layers)
self.fc1 = nn.Sequential(nn.Linear(1280, 512), nn.ReLU(),
nn.Dropout(0.2))
self.fc2 = nn.Sequential(nn.Linear(512, num_classes))
self.embed = embed
def load_checkpoint(filename, device):
checkpoint = torch.load(filename)
model_name = checkpoint['model_name']
model = models.__getattribute__(model_name)(pretrained=True)
model.to(device)
for param in model.parameters():
param.requires_grad = False
model.classifier = checkpoint['classifier']
model.class_to_idx = checkpoint['class_to_idx']
model.load_state_dict(checkpoint['state_dict'])
optimizer = checkpoint['optimizer']
model.eval()
return (model)
def __init__(self, params):
super().__init__()
self.__class__.__name__ = 'ResNet50'
base_model_name = params['base_model']
pretrained = params['pretrained']
num_classes = params['n_classes']
base_model = models.__getattribute__(base_model_name)(
pretrained=pretrained)
layers = list(base_model.children())[:-2]
layers.append(nn.AdaptiveMaxPool2d(1))
self.encoder = nn.Sequential(*layers)
in_features = base_model.fc.in_features
self.classifier = nn.Sequential(nn.Linear(in_features, 1024),
nn.ReLU(), nn.Dropout(p=0.2),
nn.Linear(1024, 1024), nn.ReLU(),
nn.Dropout(p=0.2),
nn.Linear(1024, num_classes))
def load_checkpoint(chkpoint_filepath):
# Loads deep learning model checkpoint.
# Load the saved file, on gpu or else cpu
if torch.cuda.is_available(): # Move tensors to GPU if available
checkpoint = torch.load(chkpoint_filepath)
print('model on gpu')
else:
# Needed to add this map_location parameters when on CPU
print('model on cpu')
checkpoint = torch.load(chkpoint_filepath,
map_location=lambda storage, loc: storage)
# checkpoint = torch.load(chkpoint_filepath)
# model, input_size = get_model(arch)
arch = checkpoint['arch']
model = models.__getattribute__(arch)(pretrained=True)
for param in model.parameters():
param.requires_grad = False
model.classifier = checkpoint['classifier']
# get the hyperparameters, etc, layer feature sizes from the classifier-
input_size = model.classifier.fc1.in_features
output_size = model.classifier.fc2.out_features
try: # added try for chkpts saved w/o optimzizer key
optimizer = checkpoint['optimizer']
except KeyError:
print("optizmer key check")
learning_rate = .001
optimizer = optim.Adam(model.classifier.parameters(), lr=learning_rate)
#end try
criterion = checkpoint['criterion']
# load the class names from the idx
model.class_to_idx = checkpoint['class_to_idx']
model.load_state_dict(checkpoint['state_dict'])
epochs = checkpoint['epochs']
return model, arch, criterion, epochs, optimizer
def __init__(self,
name,
outtoken,
hidden=128,
enc_layers=1,
dec_layers=1,
nhead=1,
dropout=0.1,
pretrained=False):
super(TransformerModelDnet, self).__init__()
self.backbone = models.__getattribute__(name)(pretrained=pretrained)
self.backbone.fc = nn.Conv2d(2048, hidden // 4, 1)
self.backbone.features.denseblock4.denselayer1.norm1 = nn.BatchNorm2d(
1056,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer1.conv1 = nn.Conv2d(
1056, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer1.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer2.norm1 = nn.BatchNorm2d(
1097,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer2.conv1 = nn.Conv2d(
1097, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer2.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer3.norm1 = nn.BatchNorm2d(
1138,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer3.conv1 = nn.Conv2d(
1138, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer3.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer4.norm1 = nn.BatchNorm2d(
1179,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer4.conv1 = nn.Conv2d(
1179, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer4.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer5.norm1 = nn.BatchNorm2d(
1220,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer5.conv1 = nn.Conv2d(
1220, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer5.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer6.norm1 = nn.BatchNorm2d(
1261,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer6.conv1 = nn.Conv2d(
1261, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer6.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer7.norm1 = nn.BatchNorm2d(
1302,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer7.conv1 = nn.Conv2d(
1302, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer7.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer8.norm1 = nn.BatchNorm2d(
1343,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer8.conv1 = nn.Conv2d(
1343, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer8.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer9.norm1 = nn.BatchNorm2d(
1384,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer9.conv1 = nn.Conv2d(
1384, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer9.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer10.norm1 = nn.BatchNorm2d(
1425,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer10.conv1 = nn.Conv2d(
1425, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer10.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer11.norm1 = nn.BatchNorm2d(
1466,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer11.conv1 = nn.Conv2d(
1466, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer11.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer12.norm1 = nn.BatchNorm2d(
1507,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer12.conv1 = nn.Conv2d(
1507, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer12.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer13.norm1 = nn.BatchNorm2d(
1548,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer13.conv1 = nn.Conv2d(
1548, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer13.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer14.norm1 = nn.BatchNorm2d(
1589,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer14.conv1 = nn.Conv2d(
1589, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer14.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer15.norm1 = nn.BatchNorm2d(
1630,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer15.conv1 = nn.Conv2d(
1630, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer15.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer16.norm1 = nn.BatchNorm2d(
1671,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer16.conv1 = nn.Conv2d(
1671, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer16.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer17.norm1 = nn.BatchNorm2d(
1712,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer17.conv1 = nn.Conv2d(
1712, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer17.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer18.norm1 = nn.BatchNorm2d(
1753,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer18.conv1 = nn.Conv2d(
1753, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer18.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer19.norm1 = nn.BatchNorm2d(
1794,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer19.conv1 = nn.Conv2d(
1794, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer19.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer20.norm1 = nn.BatchNorm2d(
1835,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer20.conv1 = nn.Conv2d(
1835, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer20.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer21.norm1 = nn.BatchNorm2d(
1876,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer21.conv1 = nn.Conv2d(
1876, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer21.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer22.norm1 = nn.BatchNorm2d(
1917,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer22.conv1 = nn.Conv2d(
1917, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer22.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer23.norm1 = nn.BatchNorm2d(
1958,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer23.conv1 = nn.Conv2d(
1958, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer23.conv2 = nn.Conv2d(
192,
41,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features.denseblock4.denselayer24.norm1 = nn.BatchNorm2d(
1999,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.features.denseblock4.denselayer24.conv1 = nn.Conv2d(
1999, 192, kernel_size=(1, 1), stride=(1, 1), bias=False)
self.backbone.features.denseblock4.denselayer24.conv2 = nn.Conv2d(
192,
49,
kernel_size=(3, 3),
stride=(1, 1),
padding=(1, 1),
bias=False)
self.backbone.features._modules['norm5'] = nn.BatchNorm2d(
2048,
eps=1e-05,
momentum=0.1,
affine=True,
track_running_stats=True)
self.backbone.classifier = nn.Conv2d(2048, hidden // 4, 1)
self.pos_encoder = PositionalEncoding(hidden, dropout)
self.decoder = nn.Embedding(outtoken, hidden)
self.pos_decoder = PositionalEncoding(hidden, dropout)
self.transformer = nn.Transformer(d_model=hidden,
nhead=nhead,
num_encoder_layers=enc_layers,
num_decoder_layers=dec_layers,
dim_feedforward=hidden * 4,
dropout=dropout,
activation='relu')
self.fc_out = nn.Linear(hidden, outtoken)
self.src_mask = None
self.trg_mask = None
self.memory_mask = None
