def fix_frame_extract(frame_path, feats_path, frames_num, model, video_name):
# load model
C, H, W = 3, 224, 224
if model == 'resnet152':
model = pretrainedmodels.resnet152(pretrained='imagenet')
elif model == 'vgg16':
model = pretrainedmodels.vgg16(pretrained='imagenet')
elif model == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(pretrained='imagenet')
model.last_linear = utils.Identity()
model = model.to(device)
model.eval()
load_image_fn = utils.LoadTransformImage(model)
# load data
img_list = sorted(frame_path.glob('*.jpg'))
# get index
samples_ix = np.linspace(0, len(img_list) - 1, frames_num).astype(int)
img_list = [img_list[i] for i in samples_ix]
# build tensor
imgs = torch.zeros([len(img_list), C, H, W])
for i in range(len(img_list)):
img = load_image_fn(img_list[i])
imgs[i] = img
imgs = imgs.to(device)
with torch.no_grad():
feats = model(imgs)
feats = feats.cpu().numpy()
# save
np.save(os.path.join(feats_path, video_name + ".npy"), feats)
def prepare_model(model_path, hp, device):
stump = None
if hp['stump'] == 'inception':
stump = ptm.inceptionv4()
num_ftrs = stump.last_linear.in_features
stump.last_linear = nn.Linear(num_ftrs, 2)
for i, child in enumerate(stump.features.children()):
if i < len(stump.features) * hp['freeze']:
for param in child.parameters():
param.requires_grad = False
dfs_freeze(child)
elif hp['stump'] == 'alexnet':
stump = models.alexnet(pretrained=True)
num_ftrs = stump.classifier[-1].in_features
stump.classifier[-1] = nn.Linear(num_ftrs, 2)
if hp['pretraining']:
stump.load_state_dict(torch.load(model_path, map_location=device))
print('Loaded stump: ', len(stump.features))
stump.train()
net = RetinaNet2(frame_stump=stump,
pooling_strategy=hp['pooling'],
stump_type=hp['stump'])
return net
def generate_C2D_model(opt):
if opt.c2d_model_name == 'inception_v3':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv3(num_classes=1000, pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt.c2d_model_name == 'resnet152':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet152(num_classes=1000, pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt.c2d_model_name == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(num_classes=1000, pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt.c2d_model_name == 'inceptionresnetv2':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionresnetv2(num_classes=1000, pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
model.last_linear = utils.Identity()
if not opt.no_cuda:
model = model.to(opt.device)
return load_image_fn, model, (C, H, W)
def prepare_model(model_path, hp, device):
stump = None
if hp['multi_channel']:
stump = my_inceptionv4(pretrained=False)
hp['pretraining'] = False
elif hp['narrow_model']:
stump = NarrowInceptionV1(num_classes=2)
hp['pretraining'] = False
elif hp['network'] == 'inception':
stump = ptm.inceptionv4()
num_ftrs = stump.last_linear.in_features
stump.last_linear = nn.Linear(num_ftrs, 2)
elif hp['network'] == 'alexnet':
stump = models.alexnet(pretrained=True)
stump.classifier[-1] = nn.Linear(stump.classifier[-1].in_features, 2)
if hp['pretraining']:
stump.load_state_dict(torch.load(model_path, map_location=device))
print('Loaded stump: ', len(stump.features))
stump.train()
#for i, child in enumerate(stump.features.children()):
# if i < len(stump.features) * hp['freeze']:
# for param in child.parameters():
# param.requires_grad = False
# dfs_freeze(child)
stump.to(device)
return stump
def extract_feats(args):
params = args
if params['model'] == 'inception_v3':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv3(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif params['model'] == 'resnet152':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet152(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif params['model'] == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(num_classes=1000,
pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
else:
print("doesn't support %s" % (params['model']))
model.last_linear = utils.Identity()
model = nn.DataParallel(model)
model = model.cuda()
prepro_feats.extract_feats(params, model, load_image_fn)
def prepare_model(hp, device):
net = None
if hp['network'] == 'AlexNet':
net = models.alexnet(pretrained=True)
num_ftrs = net.classifier[-1].in_features
net.classifier[-1] = Linear(num_ftrs, 2)
elif hp['network'] == 'VGG':
net = models.vgg13_bn(pretrained=True)
num_ftrs = net.classifier[-1].in_features
net.classifier[-1] = Linear(num_ftrs, 2)
elif hp['network'] == 'Efficient':
#net = EfficientNet.from_pretrained('efficientnet-b7')
net = models.efficientnet_b7(pretrained=True)
#num_ftrs = net._fc.in_features
#net._fc = Linear(num_ftrs, 2)
num_ftrs = net.classifier[-1].in_features
net.classifier[-1] = Linear(num_ftrs, 2)
elif hp['network'] == 'Inception':
net = inceptionv4()
num_ftrs = net.last_linear.in_features
net.last_linear = Linear(num_ftrs, 2)
for i, child in enumerate(net.features.children()):
if i < len(net.features) * hp['freeze']:
for param in child.parameters():
param.requires_grad = False
nn_utils.dfs_freeze(child)
if hp['pretraining']:
net.load_state_dict(torch.load(hp['pretraining'], map_location=device))
print('Loaded stump: ', len(net.features))
net.train()
#print(f'Model info: {net.__class__.__name__}, layer: {len(net)}, #frozen layer: {len(net) * hp["freeze"]}')
return net
def prepare_model(model_path, device):
stump = ptm.inceptionv4() # if not hp['multi_channel'] else my_inceptionv4(pretrained=False)
num_ftrs = stump.last_linear.in_features
stump.last_linear = nn.Linear(num_ftrs, 2)
stump.load_state_dict(torch.load(model_path, map_location=device))
return stump
def set_model (model_name, num_class, neurons_reducer_block=0, comb_method=None, comb_config=None, pretrained=True,
freeze_conv=False, p_dropout=0.5):
if pretrained:
pre_ptm = 'imagenet'
pre_torch = True
else:
pre_torch = False
pre_ptm = None
if model_name not in _MODELS:
raise Exception("The model {} is not available!".format(model_name))
model = None
if model_name == 'resnet-50':
model = MyResnet(models.resnet50(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'resnet-101':
model = MyResnet(models.resnet101(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'densenet-121':
model = MyDensenet(models.densenet121(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'vgg-13':
model = MyVGGNet(models.vgg13_bn(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'vgg-16':
model = MyVGGNet(models.vgg16_bn(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'vgg-19':
model = MyVGGNet(models.vgg19_bn(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'mobilenet':
model = MyMobilenet(models.mobilenet_v2(pretrained=pre_torch), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'efficientnet-b4':
if pretrained:
model = MyEffnet(EfficientNet.from_pretrained(model_name), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
else:
model = MyEffnet(EfficientNet.from_name(model_name), num_class, neurons_reducer_block, freeze_conv,
comb_method=comb_method, comb_config=comb_config)
elif model_name == 'inceptionv4':
model = MyInceptionV4(ptm.inceptionv4(num_classes=1000, pretrained=pre_ptm), num_class, neurons_reducer_block,
freeze_conv, comb_method=comb_method, comb_config=comb_config)
elif model_name == 'senet':
model = MySenet(ptm.senet154(num_classes=1000, pretrained=pre_ptm), num_class, neurons_reducer_block,
freeze_conv, comb_method=comb_method, comb_config=comb_config)
return model
def __init__(self, num_classes):
super(InceptionV4, self).__init__()
self.name = "InceptionV4"
self.pretrained_model = model_zoo.inceptionv4()
cfg.mean = self.pretrained_model.mean
cfg.std = self.pretrained_model.std
self.fc = nn.Linear(1536, num_classes)
nn.init.xavier_uniform(self.fc.weight, gain=2)
def __init__(self):
self.model = inceptionv4(pretrained='imagenet')
def base_arch(pretrained=True):
all_layers = list(self.model.children())
return nn.Sequential(*all_layers[0], *all_layers[1:])
self.base_arch = base_arch
def __init__(self):
super().__init__()
self.conv_lrg = nn.Sequential(create_body(models.resnet50(True), -2), nn.AdaptiveAvgPool2d(1))
self.conv_small = create_body(pretrainedmodels.inceptionv4(), -2)
self.cnn_lstm = cnn_lstm
self.country_emb = nn.Embedding(codes_count, COUNTRY_EMB_SZ)
head_inp_sz = num_features_model(self.conv_lrg) + \
model_sizes(self.conv_small, (96, 96), False)[-1][1] + self.country_emb.embedding_dim + cnn_lstm[6].rnn.hidden_size
# self.head = create_head(head_inp_sz, 340)[2:]
self.head = create_head(head_inp_sz, 340, ps=0, lin_ftrs=[4000, 2000])[2:]
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 __init__(self,
input_size=128,
num_classes=340,
pretrained='imagenet',
dropout=0.):
super().__init__()
self.model = inceptionv4(num_classes=1000, pretrained=pretrained)
self.features = self.model.features
self.relu = nn.ReLU()
self.avg_pool = nn.AvgPool2d(input_size // 64, stride=1, padding=0)
self.dropout = nn.Dropout(p=dropout)
self.last_linear = nn.Linear(1536, num_classes)
def nets(model, num_class):
if model == 'inceptionv4':
model = ptm.inceptionv4(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'senet154':
model = ptm.senet154(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'pnasnet':
model = ptm.pnasnet5large(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'xception':
model = ptm.xception(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'incepresv2':
model = ptm.inceptionresnetv2(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'resnet152':
model = models.resnet152(pretrained=True)
model.fc = nn.Linear(2048, num_class)
return model
if model == 'se_resxt101':
model = ptm.se_resnext101_32x4d(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'nasnet':
model = ptm.nasnetalarge(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, num_class)
return model
if model == 'dpn': # 224 input size
model = ptm.dpn107(num_classes=1000, pretrained='imagenet+5k')
model.last_linear = nn.Conv2d(model.last_linear.in_channels, num_class,
kernel_size=1, bias=True)
return model
if model == 'resnext101':# 320 input size
model = torch.hub.load('facebookresearch/WSL-Images', 'resnext101_32x16d_wsl')
model.fc = nn.Linear(2048, num_class)
return model
def __init__(self):
super(Model, self).__init__()
pretrained = pretrainedmodels.inceptionv4(
pretrained='imagenet+background')
# self.features = nn.Sequential(*list(pretrained.children())[:-1])
self.features = pretrained.features
# for layer in list(self.features.children())[:-2]:
# for param in layer.parameters():
# param.requires_grad = False
# for sublayer in list(list(
# self.features.children())[-2].children())[:-1]:
# for param in sublayer.parameters():
# param.requires_grad = False
self.classifier = nn.Linear(1536, num_pigs)
def run():
device = torch.device(GPU_ID if torch.cuda.is_available() else "cpu")
print(f'Using device {device}')
hyperparameter = {
'learning_rate': [1e-2, 1e-3, 3e-4, 1e-4, 3e-5], # 1e-4
'weight_decay': [0, 1e-3, 3e-4, 1e-4], # 1e-4
'num_epochs': 60, # 100
'weights': [0.0, 0.2, 0.4, 0.6, 0.8, 1.0], # 0.6
'optimizer': [optim.Adam, optim.SGD], # Adam
'image_size': 320,
'crop_size': 299,
'freeze': 0.0
}
loaders = prepare_dataset('retina', hyperparameter)
#model: nn.Module = models.resnet50(pretrained=True)
model = ptm.inceptionv4()
for i, child in enumerate(model.features):
if i < hyperparameter['freeze'] * len(model.features):
for param in child.parameters():
param.requires_grad = False
dfs_freeze(child)
num_ftrs = model.last_linear.in_features
model.last_linear = nn.Linear(num_ftrs, 2)
optimizer_ft = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=hyperparameter['learning_rate'][3],
weight_decay=hyperparameter['weight_decay'][3])
criterion = nn.CrossEntropyLoss()
plateu_scheduler = lr_scheduler.ReduceLROnPlateau(optimizer_ft,
mode='min',
factor=0.3,
patience=5,
verbose=True)
description = f"_exp3_processed90000_{model.__class__.__name__}_{hyperparameter['crop_size']}^2_{optimizer_ft.__class__.__name__}_{hyperparameter['freeze']}"
writer = SummaryWriter(comment=description)
#save_batch(next(iter(loaders[0])), '/tmp')
model = train_model(model,
criterion,
optimizer_ft,
plateu_scheduler,
loaders,
device,
writer,
num_epochs=hyperparameter['num_epochs'],
desc=description)
def __init__(self,
classifier=LogisticRegression(),
feature_model=pretrainedmodels.inceptionv4(num_classes=1000, pretrained='imagenet'),
image_transformer=None, image_type=None, padding=0):
"""
Arguments
=========
classifier - Class that implements fit and predict
The classifier that will be used. Is assumed to follow the scikit-learn API
feature_model - Class that implements "eval" and "features"
The feature extraction model. Assumed to follow the pretrainedmodels API
image_transformer - None or callable
The feature model is almost certainly assuming something about the size or pixel
distribution of the data. The image transformer should be callable and convert
an image to the format the feature_model is expecting. This is IMPORTANT!
image_type - string
The format the image should be in. RGB or something else
padding - int
The amount of padding to put around the bounding box (i.e., extra pixels to retain)
"""
self.classifier = classifier
self.feature_model = feature_model
self.feature_model.eval() # Setup the model
# If it is none, then use pretrained models to get it. Otherwise, you gotta define this
if image_transformer is None:
transformer = utils.TransformImage(self.feature_model)
self.image_transformer = lambda x: torch.autograd.Variable(transformer(x).unsqueeze(0), requires_grad=False)
else:
self.image_transformer = image_transformer
if image_type is None:
self.image_type = self.feature_model.input_space
else:
self.image_type = image_type
self.padding = padding
# Map strings to integers
self.label_mapping = []
def load_models(conf_file):
out = {}
with open(conf_file) as f:
doc = yaml.full_load(f)
conf = doc['config']
for k, stat in conf.items():
if stat['net'] == 'inception_v4':
out[k] = (pretrainedmodels.inceptionv4(pretrained=None),
stat['sparisity'])
elif stat['net'] == 'resnet18':
out[k] = (pretrainedmodels.resnet18(pretrained=None),
stat['sparisity'])
elif stat['net'] == 'mobilenet_v2':
out[k] = (models.mobilenet_v2(), stat['sparisity'])
else:
raise "Unknown net"
return out
def generate_2D_model(opt):
if opt['model'] == 'inception_v3':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv3(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'vgg16':
C, H, W = 3, 224, 224
model = pretrainedmodels.vgg16(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'vgg19':
C, H, W = 3, 224, 224
model = pretrainedmodels.vgg19(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'resnet50':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet50(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'resnet101':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet101(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'resnet152':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet152(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(num_classes=1000,
pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif opt['model'] == 'nasnet':
C, H, W = 3, 331, 331
model = pretrainedmodels.nasnetalarge(num_classes=1001,
pretrained='imagenet+background')
load_image_fn = utils.LoadTransformImage(model)
else:
print("doesn't support %s" % (opt['model']))
model.last_linear = utils.Identity()
model = nn.DataParallel(model)
# if opt['saved_model'] != '':
# model.load_state_dict(torch.load(opt['saved_model']), strict=False)
model = model.cuda()
return model
def extract_feats(frame_path, feats_path, interval, model, video_name):
"""
extract feature from frames of one video
:param video_name:
:param model: name of model
:param frame_path: path of frames
:param feats_path: path to store results
:param interval: (str) The interval when extract frames from videos
:return: None
"""
# load model
C, H, W = 3, 224, 224
if model == 'resnet152':
model = pretrainedmodels.resnet152(pretrained='imagenet')
elif model == 'vgg16':
model = pretrainedmodels.vgg16(pretrained='imagenet')
elif model == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(pretrained='imagenet')
model.last_linear = utils.Identity()
model = model.to(device)
model.eval()
load_image_fn = utils.LoadTransformImage(model)
# load data
img_list = sorted(frame_path.glob('*.jpg'))
# get index
samples_ix = np.arange(0, len(img_list), interval)
img_list = [img_list[int(i)] for i in samples_ix]
# build tensor
imgs = torch.zeros([len(img_list), C, H, W])
for i in range(len(img_list)):
img = load_image_fn(img_list[i])
imgs[i] = img
imgs = imgs.to(device)
with torch.no_grad():
feats = model(imgs)
feats = feats.cpu().numpy()
# save
np.save(os.path.join(feats_path, video_name + ".npy"), feats)
def run():
device = torch.device(GPU_ID if torch.cuda.is_available() else "cpu")
print(f'Using device {device}')
hyperparameter = {
'learning_rate': [1e-2, 1e-3, 3e-4, 1e-4, 3e-5, 1e-7], # 1e-4
'weight_decay': [0, 1e-3, 5e-4, 1e-4, 1e-5], # 1e-4
'num_epochs': 70, # 100
'weights': [0.0, 0.2, 0.4, 0.6, 0.8, 1.0], # 0.6
'optimizer': [optim.Adam, optim.SGD], # Adam
'image_size': 300,
'crop_size': 299
}
loaders = prepare_dataset('retina', hyperparameter)
#model: nn.Module = models.resnet50(pretrained=True)
#num_ftrs = model.fc.in_features
#model.fc = nn.Linear(num_ftrs, 2)
model = ptm.inceptionv4(num_classes=1000, pretrained='imagenet')
num_ft = model.last_linear.in_features
model.last_linear = nn.Linear(num_ft, 2)
children = model.features.children()
for i, child in enumerate(children):
if i < 0.0 * len(list(children)):
for param in child.parameters():
param.require_grad = False
optimizer_ft = optim.Adam(filter(lambda p: p.requires_grad,
model.parameters()),
lr=1e-7,
weight_decay=0)
criterion = nn.CrossEntropyLoss()
lr_finder = LRFinder(model, optimizer_ft, criterion, device=device)
lr_finder.range_test(loaders[0], end_lr=0.1, num_iter=100, step_mode='exp')
lr_finder.plot()
lr_finder.reset()
return 0
def inceptionv4(input_size=(3, 299, 299), num_classes=1000, pretrained=None):
# Minimum Input size = (96,96) one is must above than 97
model = models.inceptionv4(num_classes=1000, pretrained=pretrained)
if input_size != (3, 299, 299):
model.features[0].conv = nn.Conv2d(input_size[0],
32,
kernel_size=3,
stride=2,
bias=False)
model.input_size = input_size
# calculate last avgpool_1a kernel size
test_tensor = torch.randn(1, input_size[0], input_size[1], input_size[2])
x = model.features(test_tensor)
print(x.shape)
model.avg_pool = nn.AvgPool2d(kernel_size=(x.shape[2], x.shape[3]),
padding=0)
x = model.avg_pool(x)
x = x.view(x.size(0), -1).shape[1]
model.last_linear = nn.Linear(in_features=x,
out_features=num_classes,
bias=True)
return model
def main(args):
global C, H, W
coco_labels = json.load(open(args.coco_labels))
num_classes = coco_labels['num_classes']
if args.model == 'inception_v3':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv3(pretrained='imagenet')
elif args.model == 'resnet152':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet152(pretrained='imagenet')
elif args.model == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(num_classes=1000,
pretrained='imagenet')
else:
print("doesn't support %s" % (args['model']))
load_image_fn = utils.LoadTransformImage(model)
dim_feats = model.last_linear.in_features
model = MILModel(model, dim_feats, num_classes)
model = model.cuda()
dataset = CocoDataset(coco_labels)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True)
optimizer = optim.Adam(model.parameters(),
lr=args.learning_rate,
weight_decay=args.weight_decay)
exp_lr_scheduler = optim.lr_scheduler.StepLR(
optimizer,
step_size=args.learning_rate_decay_every,
gamma=args.learning_rate_decay_rate)
crit = nn.MultiLabelSoftMarginLoss()
if not os.path.isdir(args.checkpoint_path):
os.mkdir(args.checkpoint_path)
train(dataloader, model, crit, optimizer, exp_lr_scheduler, load_image_fn,
args)
def get_backbone(backbone, pretrained=True):
if backbone in ['resnext50_32x4d_ssl', 'resnet18_ssl', 'resnet50_ssl', 'resnext101_32x4d_ssl']:
if pretrained:
model = torch.hub.load(panda_config.ARCH_TO_PRETRAINED[backbone], backbone)
else:
model = getattr(_models, backbone.split('_ssl')[0])(pretrained=pretrained)
encoder = nn.Sequential(*list(model.children())[:-2])
in_features = model.fc.in_features
elif backbone in ['resnet18', 'resnet34', 'resnet50']:
pretrained = 'imagenet' if pretrained else None
model = getattr(_models, backbone)(pretrained=pretrained)
in_features = model.fc.in_features
encoder = nn.Sequential(*list(model.children())[:-2])
elif backbone in ['se_resnext50_32x4d', 'se_resnext101_32x4d', 'se_resnet50', 'se_resnet101', 'se_resnet152']:
pretrained = 'imagenet' if pretrained else None
model = getattr(pretrainedmodels, backbone)(pretrained=pretrained)
encoder = nn.Sequential(*list(model.children())[:-2])
in_features = model.last_linear.in_features
elif backbone.startswith('efficientnet'):
encoder = enet.EfficientNet.from_name(backbone)
if pretrained:
encoder.load_state_dict(torch.load(panda_config.ARCH_TO_PRETRAINED[backbone]))
in_features = encoder._fc.in_features
encoder._fc = nn.Identity()
elif backbone == 'inception_resnet_v2':
pretrained = 'imagenet' if pretrained else None
encoder = pretrainedmodels.inceptionresnetv2(pretrained=pretrained)
in_features = encoder.last_linear.in_features
encoder.last_linear = nn.Identity()
elif backbone == 'inception_v4':
pretrained = 'imagenet' if pretrained else None
encoder = pretrainedmodels.inceptionv4(pretrained=pretrained)
in_features = encoder.last_linear.in_features
encoder.last_linear = nn.Identity()
else:
raise ValueError(f'Unrecognized backbone {backbone}')
return encoder, in_features
def __init__(self, num_classes=1001):
super(InceptionV4, self).__init__()
# Modules
self.features = nn.Sequential(
BasicConv2d(3, 32, kernel_size=3, stride=2),
BasicConv2d(32, 32, kernel_size=3, stride=1),
BasicConv2d(32, 64, kernel_size=3, stride=1, padding=1),
Mixed_3a(),
Mixed_4a(),
Mixed_5a(),
Inception_A(),
Inception_A(),
Inception_A(),
Inception_A(),
Reduction_A(), # Mixed_6a
Inception_B(),
Inception_B(),
Inception_B(),
Inception_B(),
Inception_B(),
Inception_B(),
Inception_B(),
Reduction_B(), # Mixed_7a
Inception_C(),
Inception_C(),
Inception_C())
# load pre_train model
pre_model = pretrainedmodels.inceptionv4(num_classes=1000,
pretrained='imagenet')
self.last_linear = pre_model.last_linear
self.load_state_dict(pre_model.state_dict())
self.last_linear = nn.Linear(1536, num_classes)
init_params(self.last_linear)
def get_original_model(self):
if self.pretrained:
model_kwargs = {'pretrained': 'imagenet', 'num_classes': 1000}
else:
model_kwargs = {'pretrained': None}
return pretrainedmodels.inceptionv4(**model_kwargs)
def main(train_root, train_csv, train_split, val_root, val_csv, val_split,
epochs, aug, model_name, batch_size, num_workers, val_samples,
early_stopping_patience, n_classes, weighted_loss, balanced_loader,
_run):
assert (model_name
in ('inceptionv4', 'resnet152', 'densenet161', 'senet154'))
AUGMENTED_IMAGES_DIR = os.path.join(fs_observer.dir, 'images')
CHECKPOINTS_DIR = os.path.join(fs_observer.dir, 'checkpoints')
BEST_MODEL_PATH = os.path.join(CHECKPOINTS_DIR, 'model_best.pth')
LAST_MODEL_PATH = os.path.join(CHECKPOINTS_DIR, 'model_last.pth')
for directory in (AUGMENTED_IMAGES_DIR, CHECKPOINTS_DIR):
os.makedirs(directory)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if model_name == 'inceptionv4':
model = ptm.inceptionv4(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, n_classes)
aug['size'] = 299
aug['mean'] = model.mean
aug['std'] = model.std
elif model_name == 'resnet152':
model = models.resnet152(pretrained=True)
model.fc = nn.Linear(model.fc.in_features, n_classes)
aug['size'] = 224
aug['mean'] = [0.485, 0.456, 0.406]
aug['std'] = [0.229, 0.224, 0.225]
elif model_name == 'densenet161':
model = models.densenet161(pretrained=True)
model.classifier = nn.Linear(model.classifier.in_features, n_classes)
aug['size'] = 224
aug['mean'] = [0.485, 0.456, 0.406]
aug['std'] = [0.229, 0.224, 0.225]
elif model_name == 'senet154':
model = ptm.senet154(num_classes=1000, pretrained='imagenet')
model.last_linear = nn.Linear(model.last_linear.in_features, n_classes)
aug['size'] = model.input_size[1]
aug['mean'] = model.mean
aug['std'] = model.std
model.to(device)
augs = Augmentations(**aug)
model.aug_params = aug
train_ds = CSVDatasetWithName(train_root,
train_csv,
'image',
'label',
transform=augs.tf_transform,
add_extension='.jpg',
split=train_split)
val_ds = CSVDatasetWithName(val_root,
val_csv,
'image',
'label',
transform=augs.tf_transform,
add_extension='.jpg',
split=val_split)
datasets = {'train': train_ds, 'val': val_ds}
if balanced_loader:
data_sampler = sampler.WeightedRandomSampler(train_ds.sampler_weights,
len(train_ds))
shuffle = False
else:
data_sampler = None
shuffle = True
dataloaders = {
'train':
DataLoader(datasets['train'],
batch_size=batch_size,
shuffle=shuffle,
num_workers=num_workers,
sampler=data_sampler,
worker_init_fn=set_seeds),
'val':
DataLoader(datasets['val'],
batch_size=batch_size,
shuffle=False,
num_workers=num_workers,
worker_init_fn=set_seeds),
}
if weighted_loss:
criterion = nn.CrossEntropyLoss(
weight=torch.Tensor(datasets['train'].class_weights_list).cuda())
else:
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
factor=0.1,
min_lr=1e-5,
patience=10)
metrics = {
'train': pd.DataFrame(columns=['epoch', 'loss', 'acc']),
'val': pd.DataFrame(columns=['epoch', 'loss', 'acc'])
}
best_val_loss = 1000.0
epochs_without_improvement = 0
batches_per_epoch = None
for epoch in range(epochs):
print('train epoch {}/{}'.format(epoch + 1, epochs))
epoch_train_result = train_epoch(device, model, dataloaders, criterion,
optimizer, 'train', batches_per_epoch)
metrics['train'] = metrics['train'].append(
{
**epoch_train_result, 'epoch': epoch
}, ignore_index=True)
print('train', epoch_train_result)
epoch_val_result = train_epoch(device, model, dataloaders, criterion,
optimizer, 'val', batches_per_epoch)
metrics['val'] = metrics['val'].append(
{
**epoch_val_result, 'epoch': epoch
}, ignore_index=True)
print('val', epoch_val_result)
scheduler.step(epoch_val_result['loss'])
if epoch_val_result['loss'] < best_val_loss:
best_val_loss = epoch_val_result['loss']
epochs_without_improvement = 0
torch.save(model, BEST_MODEL_PATH)
print('Best loss at epoch {}'.format(epoch))
else:
epochs_without_improvement += 1
print('-' * 40)
if epochs_without_improvement > early_stopping_patience:
torch.save(model, LAST_MODEL_PATH)
break
if epoch == (epochs - 1):
torch.save(model, LAST_MODEL_PATH)
for phase in ['train', 'val']:
metrics[phase].epoch = metrics[phase].epoch.astype(int)
metrics[phase].to_csv(os.path.join(fs_observer.dir, phase + '.csv'),
index=False)
print('Best validation loss: {}'.format(best_val_loss))
# TODO: return more metrics
return {'max_val_acc': metrics['val']['acc'].max()}
help='the CNN model you want to use to extract_feats')
args = parser.parse_args()
params = vars(args)
if params['model'] == 'inception_v3':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv3(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif params['model'] == 'resnet152':
C, H, W = 3, 224, 224
model = pretrainedmodels.resnet152(pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
elif params['model'] == 'inception_v4':
C, H, W = 3, 299, 299
model = pretrainedmodels.inceptionv4(num_classes=1000,
pretrained='imagenet')
load_image_fn = utils.LoadTransformImage(model)
else:
print("doesn't support %s" % (params['model']))
model.last_linear = utils.Identity()
# model = nn.DataParallel(model)
if torch.cuda.is_available():
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
device = torch.device('cuda')
else:
device = torch.device('cpu')
model = model.to(device)
def __init__(self):
super(Model, self).__init__()
self.model = pretrainedmodels.inceptionv4(pretrained='imagenet')
import os
import torch
import torchvision
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import numpy as np
from torch.utils.data import Dataset
from torchvision import transforms, datasets
from torch.autograd import Variable
import matplotlib.pyplot as plt
import time
import pretrainedmodels
model_ft = pretrainedmodels.inceptionv4(num_classes=1000, pretrained='imagenet')
data_transforms = {
'train': transforms.Compose([
transforms.RandomSizedCrop(299),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=model_ft.mean,
std=model_ft.std)
]),
'val': transforms.Compose([
transforms.Scale(341),
transforms.CenterCrop(299),
transforms.ToTensor(),
transforms.Normalize(mean=model_ft.mean,
std=model_ft.std)
def load_inceptionv4(cls, num_classes):
model = inceptionv4(num_classes=1000, pretrained="imagenet")
model.last_linear = nn.Linear(model.last_linear.in_features,
num_classes)
return model
