def __init__(self, in_channels, out_channels, debug=False):
"""
Args:
in_channels (int): number of input channels
out_channels (int): number of output channels
vgg16_bn (torch.model): pretrained VGG-16 (with Batch Normalization) model
"""
super(Segnet, self).__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.debug = debug
# Encoder (VGG16 without Classifier)
self.enc_block00 = segnetEncoderBlock2(
self.encoder_dims('block00', 'in'),
self.encoder_dims('block00', 'out'))
self.enc_block01 = segnetEncoderBlock2(
self.encoder_dims('block01', 'in'),
self.encoder_dims('block01', 'out'))
self.enc_block02 = segnetEncoderBlock3(
self.encoder_dims('block02', 'in'),
self.encoder_dims('block02', 'out'))
self.enc_block03 = segnetEncoderBlock3(
self.encoder_dims('block03', 'in'),
self.encoder_dims('block03', 'out'))
self.enc_block04 = segnetEncoderBlock3(
self.encoder_dims('block04', 'in'),
self.encoder_dims('block04', 'out'))
# Initialize the encoder's weights
self._load_encoder_weights()
# Decoder
self.dec_block04 = segnetDecoderBlock3(
self.decoder_dims('block04', 'in'),
self.decoder_dims('block04', 'out'))
self.dec_block03 = segnetDecoderBlock3(
self.decoder_dims('block03', 'in'),
self.decoder_dims('block03', 'out'))
self.dec_block02 = segnetDecoderBlock3(
self.decoder_dims('block02', 'in'),
self.decoder_dims('block02', 'out'))
self.dec_block01 = segnetDecoderBlock2(
self.decoder_dims('block01', 'in'),
self.decoder_dims('block01', 'out'))
self.dec_block00 = segnetDecoderBlock2(
self.decoder_dims('block00', 'in'),
self.decoder_dims('block00', 'out'))
# Initialize the decoder's weights
normal_init(self.dec_block04)
normal_init(self.dec_block03)
normal_init(self.dec_block02)
normal_init(self.dec_block01)
normal_init(self.dec_block00)
# Softmax
self.softmax = nn.Softmax(dim=1)
def __init__(self, n_inputs, hidden_size, n_outputs, dropout=0.0):
super(SimpleRNN, self).__init__()
self.hidden_size = hidden_size
# self.inp = nn.Linear(n_inputs, hidden_size)
self.rnn = nn.LSTM(n_inputs,
hidden_size,
num_layers=1,
dropout=dropout)
self.out = nn.Linear(hidden_size, n_outputs)
self.softmax = nn.LogSoftmax()
for w in self.parameters():
w.data = normal_init(w.data.size())
def weight_init(self, mean=0, std=0.01):
for m in self._modules:
utils.normal_init(self._modules[m], mean, std)