def __init__(self, args):
self.args = args
self.criterion = nn.CrossEntropyLoss()
self.loss_diff = DiffLoss()
self.loss_recon = MSE()
self.loss_cmd = CMD()
self.metrics = MetricsTop().getMetics(args.datasetName)
def __init__(self, args):
self.args = args
self.origin_dimensions = args.input_feature_dimensions
self.modal_lengths = args.input_time_length # the sequence is <text, audio, video>
self.criterion = nn.L1Loss()
self.metrics = MetricsTop().getMetrics(args.datasetName)
self.ctc_settings = {}
def __init__(self, args):
assert args.datasetName == 'sims'
self.args = args
self.args.tasks = "MTAV"
self.criterion = nn.L1Loss() if args.train_mode == 'regression' else nn.CrossEntropyLoss()
self.metrics = MetricsTop(args.train_mode).getMetics(args.datasetName)
def __init__(self, args):
assert args.tasks in ['M']
self.args = args
# self.regression_criterion = nn.L1Loss()
self.regression_criterion = nn.MSELoss()
self.classification_criterion = nn.CrossEntropyLoss()
self.metrics = MetricsTop().getMetics(args.datasetName)
def __init__(self, args):
self.args = args
self.criterion = nn.MSELoss(
) if args.train_mode == 'regression' else nn.CrossEntropyLoss()
self.loss_diff = DiffLoss()
self.loss_recon = MSE()
self.loss_cmd = CMD()
self.metrics = MetricsTop(args.train_mode).getMetics(args.datasetName)
def __init__(self, args):
assert args.datasetName == 'SIMS'
self.args = args
self.criterion = nn.CrossEntropyLoss()
self.metrics = MetricsTop().getMetics(args.datasetName)
def __init__(self, args):
self.args = args
self.criterion = nn.L1Loss()
self.metrics = MetricsTop().getMetics(args.datasetName)
def __init__(self, args):
assert args.train_mode == 'regression'
self.args = args
self.args.tasks = "MTAV"
self.metrics = MetricsTop(args.train_mode).getMetics(args.datasetName)
self.feature_map = {
'fusion':
torch.zeros(args.train_samples,
args.post_fusion_dim,
requires_grad=False).to(args.device),
'text':
torch.zeros(args.train_samples,
args.post_text_dim,
requires_grad=False).to(args.device),
'audio':
torch.zeros(args.train_samples,
args.post_audio_dim,
requires_grad=False).to(args.device),
'vision':
torch.zeros(args.train_samples,
args.post_video_dim,
requires_grad=False).to(args.device),
}
self.center_map = {
'fusion': {
'pos':
torch.zeros(args.post_fusion_dim,
requires_grad=False).to(args.device),
'neg':
torch.zeros(args.post_fusion_dim,
requires_grad=False).to(args.device),
},
'text': {
'pos':
torch.zeros(args.post_text_dim,
requires_grad=False).to(args.device),
'neg':
torch.zeros(args.post_text_dim,
requires_grad=False).to(args.device),
},
'audio': {
'pos':
torch.zeros(args.post_audio_dim,
requires_grad=False).to(args.device),
'neg':
torch.zeros(args.post_audio_dim,
requires_grad=False).to(args.device),
},
'vision': {
'pos':
torch.zeros(args.post_video_dim,
requires_grad=False).to(args.device),
'neg':
torch.zeros(args.post_video_dim,
requires_grad=False).to(args.device),
}
}
self.dim_map = {
'fusion': torch.tensor(args.post_fusion_dim).float(),
'text': torch.tensor(args.post_text_dim).float(),
'audio': torch.tensor(args.post_audio_dim).float(),
'vision': torch.tensor(args.post_video_dim).float(),
}
# new labels
self.label_map = {
'fusion':
torch.zeros(args.train_samples,
requires_grad=False).to(args.device),
'text':
torch.zeros(args.train_samples,
requires_grad=False).to(args.device),
'audio':
torch.zeros(args.train_samples,
requires_grad=False).to(args.device),
'vision':
torch.zeros(args.train_samples,
requires_grad=False).to(args.device)
}
self.name_map = {
'M': 'fusion',
'T': 'text',
'A': 'audio',
'V': 'vision'
}
