def __init__(self, dropout=0.2, logstd1=-1, logstd2=-2, pi=0.5):
super(GestureSpotting, self).__init__()
type = bl.ModelType.MC_DROP
linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
rnn_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
last_linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": 0
}
stats = np.load("train_stats.npy")
#Embedding layers
self.bfc1 = bl.Linear(54, 32, **linear_args)
self.bfc2 = bl.Linear(32, 32, **linear_args)
self.fc1 = nn.Sequential(self.bfc1, nn.ReLU())
self.fc2 = nn.Sequential(self.bfc2, nn.ReLU())
#attributs
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.hidden_state = None
self.mean, self.std = stats[0], stats[1] + 1e-18
#model parameters
self.input_size = 32
self.output_size = 2
self.hidden_dim = 64
self.n_layers = 1
#Recurrent Bayesian Layer
self.lstm = bl.LSTM( input_size = self.input_size,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
# dropout layer
self.fc_combine = bl.Linear(self.input_size, self.input_size,
**linear_args)
self.combine = nn.Sequential(self.fc_combine, nn.ReLU())
# Classifier layer
self.fc = bl.Linear(self.hidden_dim, self.output_size,
**last_linear_args)
def __init__(self, output_size, hidden_dim,n_layers, mode = "mc", dropout = 0.5, logstd1 = -1, logstd2 = -2, pi = 0.5 ):
super(IterStarRGBHandModel, self).__init__()
#type = bl.ModelType.VAR_DROP_B_ADAP
mode = mode.lower()
if mode == "mc": self.type = bl.ModelType.MC_DROP
elif mode == "vd": self.type = bl.ModelType.VAR_DROP_B_ADAP
elif mode == "bbb": self.type = bl.ModelType.BBB
else: self.type = bl.ModelType.DET
rnn_args = {
"mu":0,
"logstd1":logstd1,
"logstd2":logstd2,
"pi":pi,
"type":self.type,
"dropout":dropout
}
last_linear_args = {
"mu":0,
"logstd1":logstd1,
"logstd2":logstd2,
"pi":pi,
"type":bl.ModelType.DET,
"dropout":0
}
resnet_mov = models.resnet50(pretrained=True)
self.mov= nn.Sequential(*(list(resnet_mov.children())[:-1]),SqueezeExtractor(), nn.Conv1d(1, 1, 3, stride=2),nn.ReLU())
resnet_hand = models.resnet50(pretrained=True)
self.hand= nn.Sequential(*(list(resnet_hand.children())[:-1]), SqueezeExtractor(), nn.Conv1d(1, 1, 3, stride=2),nn.ReLU())
#Embedding
self.embeding_size = 1023
self.soft = SoftAttention(1023,128)
# nn.MaxPool1d(2, stride=2)
weights = [1./1.5, 1.0]
class_weights = torch.FloatTensor(weights).cuda()
self.loss_fn = nn.NLLLoss(weight = class_weights) if output_size == 2 else nn.NLLLoss()
self.output_size = output_size #12
self.hidden_dim = hidden_dim #256
self.n_layers = n_layers #2
self.sharpen = False
self.lstm = bl.LSTM( input_size = self.embeding_size,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.dropout = dropout
self.fc = bl.Linear(self.hidden_dim, self.output_size, **last_linear_args)
self.dp_training = True
self._baysian_layers = []
self.hidden = None
self.td = TimeDistributed()
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# self.device = torch.device( 'cpu')
self.frozen = False
def __init__(self, input_size = 512, hidden_size = 128, type = "det" ):
super(SoftAttention, self).__init__()
if type == "mc": self.type = bl.ModelType.MC_DROP
elif type == "vd": self.type = bl.ModelType.VAR_DROP_B_ADAP
elif type == "bbb": self.type = bl.ModelType.BBB
else: self.type = bl.ModelType.DET
args = {
"mu":0,
"logstd1":1.0,
"logstd2":1.0,
"pi":0.5,
"type":self.type,
"dropout":0
}
self.softatt = nn.Sequential(bl.Linear(input_size, hidden_size, **args),
nn.ReLU(),
bl.Linear(hidden_size, 1, **args),
)
self.weights = None
def __init__(self,
data_type,
input_size,
output_size,
hidden_dim,
n_layers,
mode="mc",
dropout=0.5,
logstd1=-1,
logstd2=-2,
pi=0.5):
super(RTGR, self).__init__()
#type = bl.ModelType.VAR_DROP_B_ADAP
if mode == "mc": type = bl.ModelType.MC_DROP
elif mode == "vd": type = bl.ModelType.VAR_DROP_B_ADAP
elif mode == "bbb": type = bl.ModelType.BBB
else: type = bl.ModelType.DET
linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
rnn_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
last_linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": 0
}
self.data_type = data_type
#Embedding
self.input_size = input_size
self.embeding_size = 16
self.bfc1 = bl.Linear(self.input_size, 32, **linear_args)
self.bfc2 = bl.Linear(32, self.embeding_size, **linear_args)
self.fc1 = nn.Sequential(self.bfc1, nn.ReLU())
self.fc2 = nn.Sequential(self.bfc2, nn.ReLU())
#weights = [1./21, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0,1.0, 1.0, 1.0, 1.0, 1.0]
weights = [1. / 1.5, 1.0]
class_weights = torch.FloatTensor(weights).cuda()
self.loss_fn = nn.NLLLoss(
weight=class_weights) if output_size == 2 else nn.NLLLoss()
self.output_size = output_size #12
self.hidden_dim = hidden_dim #256
self.n_layers = n_layers #2
self.sharpen = False
self.lstm = bl.LSTM( input_size = self.embeding_size,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.dropout = dropout
self.fc = bl.Linear(self.hidden_dim, self.output_size,
**last_linear_args)
self.dp_training = True
self._baysian_layers = []
def __init__(self, data_type, input_size, output_size, hidden_dim,n_layers, mode = "mc", dropout = 0.5, logstd1 = -1, logstd2 = -2, pi = 0.5 ):
super(RTGR, self).__init__()
#type = bl.ModelType.VAR_DROP_B_ADAP
if mode == "mc": type = bl.ModelType.MC_DROP
elif mode == "vd": type = bl.ModelType.VAR_DROP_B_ADAP
elif mode == "bbb": type = bl.ModelType.BBB
else: type = bl.ModelType.DET
linear_args = {
"mu":0,
"logstd1":logstd1,
"logstd2":logstd2,
"pi":pi,
"type":bl.ModelType.DET,
"dropout":dropout
}
rnn_args = {
"mu":0,
"logstd1":logstd1,
"logstd2":logstd2,
"pi":pi,
"type":type,
"dropout":dropout
}
last_linear_args = {
"mu":0,
"logstd1":logstd1,
"logstd2":logstd2,
"pi":pi,
"type":type,
"dropout":0
}
self.data_type = data_type
#Embedding
self.skl_input_size = input_size
self.skl_emb_size = 256
self.hand_input_size = 512
self.hand_emb_size = 256
bfc1 = bl.Linear(self.skl_input_size, self.skl_emb_size, **linear_args)
bfc2 = bl.Linear(self.skl_emb_size, self.skl_emb_size, **linear_args)
bfc3 = bl.Linear(self.hand_input_size, self.hand_emb_size, **linear_args)
# self.bfc4 = bl.Linear(512, self.hand_emb_size, **linear_args)
self.hand_embedding = nn.Sequential(bfc3,nn.ReLU())
self.skl_embedding = nn.Sequential(bfc1,nn.ReLU(), bfc2,nn.ReLU())
self.soft_att = SoftAttention(input_size = self.hand_input_size)
self.soft_att_info = SoftAttention(input_size = self.hand_emb_size)
self.soft_att_spt = SoftAttention(input_size = self.hand_emb_size)
model = models.resnet34(pretrained=True)
features= list(model.children())
self.hand_features = nn.Sequential(*features[:-1], SqueezeExtractor())
weights = [1.0, 1.0]
class_weights = torch.FloatTensor(weights).cuda()
self.loss_fn = nn.NLLLoss(weight = class_weights) if output_size == 2 else nn.NLLLoss()
self.output_size = output_size
self.hidden_dim = hidden_dim
self.n_layers = n_layers
self.rnn = bl.LSTM( input_size = self.skl_emb_size ,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.bfc_output = bl.Linear(self.hidden_dim, 15, **last_linear_args)
self.rnn_spt = bl.LSTM( input_size = self.skl_emb_size ,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.bfc_output_spt = bl.Linear(self.hidden_dim, 2, **last_linear_args)
self.dropout = dropout
self._baysian_layers = []
self.hidden = None
def __init__(self,
data_type,
input_size,
output_size,
hidden_dim,
n_layers,
mode="mc",
dropout=0.5,
logstd1=-1,
logstd2=-2,
pi=0.5):
super(RTGR, self).__init__()
#type = bl.ModelType.VAR_DROP_B_ADAP
if mode == "mc": type = bl.ModelType.MC_DROP
elif mode == "vd": type = bl.ModelType.VAR_DROP_B_ADAP
elif mode == "bbb": type = bl.ModelType.BBB
else: type = bl.ModelType.DET
linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
rnn_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": dropout
}
last_linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": type,
"dropout": 0
}
self.data_type = data_type
#self.norm = nn.InstanceNorm1d(input_size)
#Embedding
model = models.resnet18(pretrained=True)
features = list(model.children())
conv = features[0]
self.features = nn.Sequential(*features[1:-1])
self.conv_left = inflate_conv(conv,
time_dim=4,
time_padding=0,
time_stride=1,
time_dilation=1,
center=False)
self.conv_right = inflate_conv(conv,
time_dim=4,
time_padding=0,
time_stride=1,
time_dilation=1,
center=False)
self.loss_fn = nn.NLLLoss(
weight=class_weights) if output_size == 2 else nn.NLLLoss()
self.output_size = 15 #12
self.hidden_dim = hidden_dim #256
self.n_layers = n_layers #2
self.sharpen = False
self.lstm = bl.LSTM( input_size = 512,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.dropout = dropout
self.fc = bl.Linear(self.hidden_dim, self.output_size,
**last_linear_args)
self.dp_training = True
self._baysian_layers = []
def __init__(self,
output_size,
hidden_dim,
n_layers,
mode="mc",
dropout=0.5,
logstd1=-1,
logstd2=-2,
pi=0.5):
super(IterStarRGBModel, self).__init__()
#type = bl.ModelType.VAR_DROP_B_ADAP
if mode == "mc": self.type = bl.ModelType.MC_DROP
elif mode == "vd": self.type = bl.ModelType.VAR_DROP_B_ADAP
elif mode == "bbb": self.type = bl.ModelType.BBB
else: self.type = bl.ModelType.DET
linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": self.type,
"dropout": dropout
}
rnn_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": self.type,
"dropout": dropout
}
last_linear_args = {
"mu": 0,
"logstd1": logstd1,
"logstd2": logstd2,
"pi": pi,
"type": self.type,
"dropout": 0
}
resnet = models.resnet50(pretrained=True)
self.resnet = nn.Sequential(*(list(resnet.children())[:-1]))
#self.norm = nn.InstanceNorm1d(input_size)
#Embedding
self.conv1d = nn.Conv1d(1, 1, 3, stride=2)
# nn.MaxPool1d(2, stride=2)
self.embeding_size = 1023
weights = [1. / 1.5, 1.0]
class_weights = torch.FloatTensor(weights).cuda()
self.loss_fn = nn.NLLLoss(
weight=class_weights) if output_size == 2 else nn.NLLLoss()
self.output_size = output_size #12
self.hidden_dim = hidden_dim #256
self.n_layers = n_layers #2
self.sharpen = False
self.lstm = bl.LSTM( input_size = self.embeding_size,\
hidden_size = self.hidden_dim, num_layers = self.n_layers, \
batch_first=True,**rnn_args)
self.dropout = dropout
# linear and sigmoid layers
self.fc = bl.Linear(self.hidden_dim, self.output_size,
**last_linear_args)
self.dp_training = True
self._baysian_layers = []
self.hidden = None
# self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.device = torch.device('cpu')