def build_model(self):
self.p3d = C3D().cuda()
self.load_model()
self.gru = GRU(self.p3d).cuda()
print("MODEL:")
print(self.gru)
def __init__(self, embed_size):
super(EncoderC3D, self).__init__()
base_model = C3D().cuda()
base_model.load_state_dict(torch.load(WEIGHT_PATH))
self.conv1 = base_model.conv1
self.pool1 = base_model.pool1
self.conv2 = base_model.conv2
self.pool2 = base_model.pool2
self.conv3a = base_model.conv3a
self.conv3b = base_model.conv3b
self.pool3 = base_model.pool3
self.conv4a = base_model.conv4a
self.conv4b = base_model.conv4b
self.pool4 = base_model.pool4
self.conv5a = base_model.conv5a
self.conv5b = base_model.conv5b
self.pool5 = base_model.pool5
self.fc6 = base_model.fc6
self.fc7 = base_model.fc7
self.fc8 = nn.Linear(4096, embed_size)
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.5)
self.bn = nn.BatchNorm1d(embed_size, momentum=0.01)
def loading_model(self):
print('Loading %s model' % (self.model_type))
if self.model_type == 'C3D':
self.model = C3D()
elif self.model_type == 'I3D':
self.model = I3D(num_classes=400, modality='rgb')
else:
self.model = P3D199(pretrained=False, num_classes=400, dropout=self.dropout)
# Transfer classes
self.model = transfer_model(model=self.model, model_type=self.model_type, num_classes=self.num_classes)
# Check gpu and run parallel
if check_gpu() > 0:
self.model = torch.nn.DataParallel(self.model).cuda()
# define loss function (criterion) and optimizer
if check_gpu() > 0:
self.criterion = nn.CrossEntropyLoss().cuda()
else:
self.criterion = nn.CrossEntropyLoss()
policies = get_optim_policies(model=self.model, modality=self.modality, enable_pbn=True)
self.optimizer = optim.SGD(policies, lr=self.lr, momentum=self.momentum, weight_decay=self.weight_decay)
file = os.path.join(self.data_folder, 'model_best.pth.tar')
if os.path.isfile(file):
print("=> loading checkpoint '{}'".format('model_best.pth.tar'))
checkpoint = torch.load(file)
self.start_epoch = checkpoint['epoch']
self.best_prec1 = checkpoint['best_prec1']
self.model.load_state_dict(checkpoint['state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded model best ")
else:
print("=> no model best found at ")
exit()
cudnn.benchmark = True
train_phase(train_dataloader, optimizer, criterion, epoch)
test_phase(test_dataloader)
if (epoch +
1) % model_ckpt_interval == 0: # save models every 5 epochs
save_model(model_lstm, 'model_my_lstm', epoch, saving_dir)
# lr updates
if (epoch + 1) % global_lr_stepsize == 0:
learning_rate = learning_rate * global_lr_gamma
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
if __name__ == '__main__':
# loading the altered C3D (ie C3D upto before fc-6)
model_CNN_pretrained_dict = torch.load('c3d.pickle')
model_CNN = C3D()
model_CNN_dict = model_CNN.state_dict()
model_CNN_pretrained_dict = {
k: v
for k, v in model_CNN_pretrained_dict.items() if k in model_CNN_dict
}
model_CNN_dict.update(model_CNN_pretrained_dict)
model_CNN.load_state_dict(model_CNN_dict)
model_CNN = model_CNN.cuda()
# lstm
model_lstm = LSTM_anno()
model_lstm = model_lstm.cuda()
main()
return -dist
# 导入测试的参数
args = parse_args()
# print(args)
# 加载测试数据
test_file = '/opt/data/private/DL_Workspace/fsv-baseline/datafile/{}/few-shot-test.txt'.format(
args.dataset)
data_mgr = SetDataManager(args.n_way, args.k_shot, args.q_query,
args.test_episode)
loader = data_mgr.get_data_loader(test_file)
# 加载模型
if args.model == 'C3D':
model = C3D(
num_classes=31, need_feature=True,
pooling=args.pooling) # 这里 num_classes 是一个没有用的参数,但必须加上, 保证和保存的模型参数匹配
elif args.model == 'FS_ResNet':
model = FS_ResNet(args.k_shot, args.backbone_size)
elif args.model == 'FS_MENet':
model = FS_MENet(args.k_shot, args.backbone_size)
else:
raise NotImplementedError('no such a model!')
model = model.cuda()
check_dir = './chechpoints/FS_MENet34_hmdb51_SGD_lr_0.1_epi_True_epoch_100'
print(check_dir)
# best_file = get_best_file(check_dir)
# best_file = get_resume_file(check_dir)
best_file = get_assigned_file(check_dir, num=args.checkpoint)
tmp = torch.load(best_file)
val_dataset = SimpleVideoDataset(val_file, split='val', clip_len=8)
tr_loader = DataLoader(tr_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=4)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
num_workers=4)
# 建立模型
if args.model == 'FS_ResNet':
model = FS_ResNet(args.k_shot, args.backbone_size)
elif args.model == 'FS_MENet':
model = FS_MENet(args.k_shot, args.backbone_size)
elif args.model == 'C3D':
model = C3D(train_classes, pooling=args.pooling)
args.backbone_size = ''
else:
raise NotImplementedError("Not implement such a model")
model = model.cuda()
if args.optim == 'SGD':
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=0.9)
# optimizer = optim.SGD([
# {'params': model.resnet.parameters(), 'lr': args.lr/10.0},
# {'params': model.menet.parameters(), 'lr': args.lr/10.0}
# ])
elif args.optim == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=args.lr)
else:
optimizer = None
raise NotImplementedError("Not support for this optim:{}".format(
DATA_TYPE, MODEL_TYPE, VIDEO_NAME, MODEL_NAME = opt.data_type, opt.model_type, opt.video_name, opt.model_name
clip_len, resize_height, crop_size, = utils.CLIP_LEN, utils.RESIZE_HEIGHT, utils.CROP_SIZE
class_names = utils.get_labels(DATA_TYPE)
DEVICE = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
if '{}_{}.pth'.format(DATA_TYPE, MODEL_TYPE) != MODEL_NAME:
raise NotImplementedError(
'the model name must be the same model type and same data type')
if MODEL_TYPE == 'r2plus1d':
model = R2Plus1D(len(class_names), (2, 2, 2, 2))
else:
model = C3D(len(class_names))
checkpoint = torch.load('epochs/{}'.format(MODEL_NAME),
map_location=lambda storage, loc: storage)
model = model.load_state_dict(checkpoint).to(DEVICE).eval()
# read video
cap, retaining, clips = cv2.VideoCapture(VIDEO_NAME), True, []
frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
while retaining:
retaining, frame = cap.read()
if not retaining and frame is None:
continue
resize_width = math.floor(frame.shape[1] / frame.shape[0] *
resize_height)
# make sure it can be cropped correctly
def loading_model(self):
print('Loading %s model' % (self.model_type))
if self.model_type == 'C3D':
self.model = C3D()
if self.pretrained:
self.model.load_state_dict(torch.load('c3d.pickle'))
elif self.model_type == 'I3D':
if self.pretrained:
self.model = I3D(num_classes=400, modality='rgb')
self.model.load_state_dict(
torch.load('kinetics_i3d_model_rgb.pth'))
else:
self.model = I3D(num_classes=self.num_classes, modality='rgb')
else:
if self.pretrained:
print("=> using pre-trained model")
self.model = P3D199(pretrained=True,
num_classes=400,
dropout=self.dropout)
else:
print("=> creating model P3D")
self.model = P3D199(pretrained=False,
num_classes=400,
dropout=self.dropout)
# Transfer classes
self.model = transfer_model(model=self.model,
model_type=self.model_type,
num_classes=self.num_classes)
# Check gpu and run parallel
if check_gpu() > 0:
self.model = torch.nn.DataParallel(self.model).cuda()
# define loss function (criterion) and optimizer
self.criterion = nn.CrossEntropyLoss()
if check_gpu() > 0:
self.criterion = nn.CrossEntropyLoss().cuda()
params = self.model.parameters()
if self.model_type == 'P3D':
params = get_optim_policies(model=self.model,
modality=self.modality,
enable_pbn=True)
self.optimizer = optim.SGD(params=params,
lr=self.lr,
momentum=self.momentum,
weight_decay=self.weight_decay)
# self.scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer=self.optimizer, mode='min', patience=10, verbose=True)
# optionally resume from a checkpoint
if self.resume:
if os.path.isfile(self.resume):
print("=> loading checkpoint '{}'".format(self.resume))
checkpoint = torch.load(self.resume)
self.start_epoch = checkpoint['epoch']
self.best_prec1 = checkpoint['best_prec1']
self.model.load_state_dict(checkpoint['state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
self.evaluate, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(self.resume))
if self.evaluate:
file_model_best = os.path.join(self.data_folder,
'model_best.pth.tar')
if os.path.isfile(file_model_best):
print(
"=> loading checkpoint '{}'".format('model_best.pth.tar'))
checkpoint = torch.load(file_model_best)
self.start_epoch = checkpoint['epoch']
self.best_prec1 = checkpoint['best_prec1']
self.model.load_state_dict(checkpoint['state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
self.evaluate, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(self.resume))
cudnn.benchmark = True
# from models.BabyC3D import Crazy
model = BabyC3D()
# model = Crazy()
elif model_choice == 'small':
from models.SmallC3D import SmallC3D
model = SmallC3D()
elif model_choice == 'se3cnn':
from models.Se3cnn import Se3cnn
model = Se3cnn()
elif model_choice == 'c3d':
from models.C3D import C3D
model = C3D()
elif model_choice == 'small_siamese':
from models.Siamese import SmallSiamese
model = SmallSiamese()
elif model_choice == 'baby_siamese':
from models.Siamese import BabySiamese
model = BabySiamese()
elif model_choice == 'babyse3cnn':
from models.BabySe3cnn import BabySe3cnn
model = BabySe3cnn()
else:
# Not possible because of argparse
raise ValueError('Not a possible model')
'val_top5_accuracy': [],
'test_loss': [],
'test_top1_accuracy': [],
'test_top5_accuracy': []
}
# record best val accuracy
best_accuracy = 0
train_loader, val_loader, test_loader = utils.load_data(
DATA_TYPE, BATCH_SIZE)
NUM_CLASS = len(train_loader.dataset.label2index)
if MODEL_TYPE == 'r2plus1d':
model = R2Plus1D(NUM_CLASS, (2, 2, 2, 2))
else:
model = C3D(NUM_CLASS)
if PRE_TRAIN is not None:
checkpoint = torch.load('epochs/{}'.format(PRE_TRAIN),
map_location=lambda storage, loc: storage)
# load pre-trained model which trained on the same dataset
if DATA_TYPE in PRE_TRAIN:
# load same type pre-trained model
if PRE_TRAIN.split('.')[0].split('_')[1] == MODEL_TYPE:
model.load_state_dict(checkpoint)
else:
raise NotImplementedError(
'the pre-trained model must be the same model type')
# warm starting model by loading weights from a model which trained on other dataset, then fine tuning
else:
if PRE_TRAIN.split('.')[0].split('_')[1] == MODEL_TYPE:
def make_predictions(model_choice, model_name, loader):
"""
Make the prediction for a class of model, trained in a file named model_name
Saves a dict 'path like 1a0g_PMP_0.pdb.npy : predicted 128 embedding' in predictions/model_name
:param model_choice:
:param model_name:
:return:
"""
torch.multiprocessing.set_sharing_strategy('file_system')
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# I made a mistake in the saving script
model_path = os.path.join('../trained_models', model_name,
model_name + '.pth')
if model_choice == 'baby':
from models.BabyC3D import BabyC3D
# from models.BabyC3D import Crazy
model = BabyC3D()
# model = Crazy()
elif model_choice == 'small':
from models.SmallC3D import SmallC3D
model = SmallC3D()
elif model_choice == 'se3cnn':
from models.Se3cnn import Se3cnn
model = Se3cnn()
elif model_choice == 'c3d':
from models.C3D import C3D
model = C3D()
elif model_choice == 'small_siamese':
from models.Siamese import SmallSiamese
model = SmallSiamese()
elif model_choice == 'baby_siamese':
from models.Siamese import BabySiamese
model = BabySiamese()
elif model_choice == 'babyse3cnn':
from models.BabySe3cnn import BabySe3cnn
model = BabySe3cnn()
else:
# Not possible because of argparse
raise ValueError('Not a possible model')
model.to(device)
model = torch.nn.DataParallel(model)
# import torch.optim as optim
# optimizer = optim.Adam(None)
# print(model, model_path)
dict_results = run_model(loader, model, model_path)
pickle.dump(
dict_results,
open(f'../data/post_processing/predictions/{model_name}.p', 'wb'))
return dict_results