def __init__(self, configs):
super(MMNIST_ConvLSTM, self).__init__()
_KEYS = ['encoder_configs', 'reconstruct_configs', 'predict_configs']
en_conf, rec_conf, pred_conf = unpack(configs, _KEYS)
self.encoder = ConvLSTM(en_conf)
self.reconstructor = Generator(rec_conf)
self.predictor = Generator(pred_conf)
def __init__(self, num_classes=61, mem_size=512, no_cam=False):
super(NewAttentionModelBi, self).__init__()
self.num_classes = num_classes
self.noCam = no_cam
self.mem_size = mem_size
self.resnet = resnet34(pretrained=True, noBN=True)
self.attention_rgb = Variable(
(torch.FloatTensor(512).normal_(0, .05)).unsqueeze(0).cuda())
self.attention_flow = Variable(
(torch.FloatTensor(512).normal_(0, .05)).unsqueeze(0).cuda())
self.lstm_cell = ConvLSTM(1024, mem_size)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
self._custom_train_mode = True
def __init__(self,
num_classes=61,
mem_size=512,
no_cam=False,
enable_motion_segmentation=False):
super(AttentionModel, self).__init__()
self.num_classes = num_classes
self.noCam = no_cam
self.mem_size = mem_size
self.enable_motion_segmentation = enable_motion_segmentation
self.resnet = resnet34(pretrained=True, noBN=True)
self.weight_softmax = self.resnet.fc.weight
self.lstm_cell = ConvLSTM(512, mem_size)
self.avgpool = nn.AvgPool2d(7)
self.dropout = nn.Dropout(0.7)
self.fc = nn.Linear(mem_size, self.num_classes)
self.classifier = nn.Sequential(self.dropout, self.fc)
self.motion_segmentation = MotionSegmentationBlock()
self._custom_train_mode = True
def __init__(self, in_dim, ae_en_h_dims, ae_de_h_dims, conv_lstm_in_size,
conv_lstm_in_dim, conv_lstm_h_dim, conv_lstm_kernel_sizes,
conv_lstm_n_layers, fc_in_dim, fc_h_dims, fc_out_dim,
**kwargs):
super(DeepAP, self).__init__()
self.device = kwargs.get('device', 'cpu')
################
# masked layer #
################
mask = [[i for i in range(in_dim)], [i for i in range(in_dim)]]
self.mask_layer = MaskNet(in_dim, in_dim, mask, device=self.device)
self.mask_thre = kwargs.get('mask_thre', 0.0001)
######################
# auto_encoder layer #
######################
self.ae = AutoEncoder(in_dim=in_dim,
en_h_dims=ae_en_h_dims,
de_h_dims=ae_de_h_dims)
if kwargs.get('ae_pretrain_weight') is not None:
self.ae.load_state_dict(kwargs['ae_pretrain_weight'])
else:
raise ValueError('AutoEncoder not pretrained.')
if kwargs.get('if_trainable'):
for p in self.ae.parameters():
p.requires_grad = kwargs['if_trainable']
else:
self.ae.weight.requires_grad = False
####################
# conv_lstm layers #
####################
self.conv_lstm_list = nn.ModuleList()
for i in conv_lstm_kernel_sizes:
i_kernel_size = (i, i)
conv_lstm = ConvLSTM(
in_size=conv_lstm_in_size,
in_dim=conv_lstm_in_dim,
h_dim=conv_lstm_h_dim,
kernel_size=i_kernel_size,
num_layers=conv_lstm_n_layers,
batch_first=kwargs.get('conv_lstm_batch_first', True),
bias=kwargs.get('conv_lstm_bias', True),
only_last_state=kwargs.get('only_last_state', True),
device=self.device)
self.conv_lstm_list.append(conv_lstm)
#########################
# fully-connected layer #
#########################
self.fc = FC(
in_dim=fc_in_dim, # assert in_size == n_conv_lstm * conv_lstm_h_dim
h_dims=fc_h_dims,
out_dim=fc_out_dim,
p_dropout=kwargs.get('fc_p_dropout', 0.1))