def __init__(self, in_shape, normalize):
super().__init__()
bias = not normalize
self._nb_output_channel = 3200
self.conv1 = Conv2d(in_shape[0], 32, 7, stride=2, padding=1, bias=bias)
self.conv2 = Conv2d(32, 64, 3, stride=2, padding=1, bias=bias)
self.conv3 = Conv2d(64, 64, 3, stride=2, padding=1, bias=bias)
self.conv4 = Conv2d(64, 128, 3, stride=2, padding=1, bias=bias)
if normalize:
self.bn1 = BatchNorm2d(32)
self.bn2 = BatchNorm2d(64)
self.bn3 = BatchNorm2d(64)
self.bn4 = BatchNorm2d(128)
else:
self.bn1 = Identity()
self.bn2 = Identity()
self.bn3 = Identity()
self.bn4 = Identity()
relu_gain = init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
self.conv4.weight.data.mul_(relu_gain)
def __init__(self, nb_in_chan, output_shape_dict, normalize):
self.embedding_size = 512
super(RMC, self).__init__(self.embedding_size, output_shape_dict)
bias = not normalize
self.conv1 = Conv2d(
nb_in_chan, 32, kernel_size=3, stride=2, padding=1, bias=bias
)
self.conv2 = Conv2d(
32, 32, kernel_size=3, stride=2, padding=1, bias=bias
)
self.conv3 = Conv2d(
32, 32, kernel_size=3, stride=2, padding=1, bias=bias
)
self.attention = RMCCell(100, 100, 34)
self.conv4 = Conv2d(
34, 8, kernel_size=3, stride=1, padding=1, bias=bias
)
# BATCH x 8 x 10 x 10
self.linear = Linear(800, 512, bias=bias)
if normalize:
self.bn1 = BatchNorm2d(32)
self.bn2 = BatchNorm2d(32)
self.bn3 = BatchNorm2d(32)
self.bn4 = BatchNorm2d(8)
self.bn_linear = BatchNorm1d(512)
else:
self.bn1 = Identity()
self.bn2 = Identity()
self.bn3 = Identity()
self.bn_linear = Identity()
def __init__(self, in_shape, id, normalize):
super().__init__(in_shape, id)
bias = not normalize
self._in_shape = in_shape
self._out_shape = None
self.conv1 = Conv2d(in_shape[0], 32, 7, stride=2, padding=1, bias=bias)
self.conv2 = Conv2d(32, 32, 3, stride=2, padding=1, bias=bias)
self.conv3 = Conv2d(32, 32, 3, stride=2, padding=1, bias=bias)
self.conv4 = Conv2d(32, 32, 3, stride=2, padding=1, bias=bias)
if normalize == "bn":
self.bn1 = BatchNorm2d(32)
self.bn2 = BatchNorm2d(32)
self.bn3 = BatchNorm2d(32)
self.bn4 = BatchNorm2d(32)
elif normalize == "gn":
self.bn1 = GroupNorm(8, 32)
self.bn2 = GroupNorm(8, 32)
self.bn3 = GroupNorm(8, 32)
self.bn4 = GroupNorm(8, 32)
else:
self.bn1 = Identity()
self.bn2 = Identity()
self.bn3 = Identity()
self.bn4 = Identity()
relu_gain = init.calculate_gain("relu")
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
self.conv4.weight.data.mul_(relu_gain)
def __init__(self, in_shape, nb_head, normalize):
self._nb_output_channel = 800
super().__init__()
self.normalize = normalize
bias = not normalize
self.conv1 = Conv2d(in_shape[0],
32,
kernel_size=3,
stride=2,
padding=1,
bias=bias)
self.conv2 = Conv2d(32,
32,
kernel_size=3,
stride=2,
padding=1,
bias=bias)
self.attention = MultiHeadSelfAttention(20 * 20, 34, 34, nb_head)
self.mlp = Linear(34, 34)
self.conv3 = Conv2d(34,
32,
kernel_size=3,
stride=2,
padding=1,
bias=bias)
self.conv4 = Conv2d(32,
32,
kernel_size=3,
stride=2,
padding=1,
bias=bias)
if normalize:
self.bn1 = BatchNorm2d(32)
self.bn2 = BatchNorm2d(32)
self.bn3 = BatchNorm2d(32)
self.bn4 = BatchNorm2d(32)
else:
self.bn1 = Identity()
self.bn2 = Identity()
self.bn3 = Identity()
self.bn4 = Identity()
relu_gain = init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
self.conv4.weight.data.mul_(relu_gain)
def __init__(self, input_shape, id, normalize, nb_hidden, nb_layer):
super().__init__(input_shape, id)
self._nb_hidden = nb_hidden
nb_input_channel = input_shape[0]
bias = not normalize
self.linears = nn.ModuleList([
nn.Linear(
nb_input_channel if i == 0 else nb_hidden,
nb_hidden,
bias
)
for i in range(nb_layer)
])
if normalize == 'bn':
self.norms = nn.ModuleList([
nn.BatchNorm1d(nb_hidden) for _ in range(nb_layer)
])
elif normalize == 'gn':
if nb_hidden % 16 != 0:
raise Exception('linear_nb_hidden must be divisible by 16 for Group Norm')
self.norms = nn.ModuleList([
nn.GroupNorm(nb_hidden // 16, nb_hidden) for _ in range(nb_layer)
])
else:
self.norms = nn.ModuleList([
Identity() for _ in range(nb_layer)
])
def __init__(self, in_shape, normalize):
super().__init__()
bias = not normalize
self._nb_output_channel = 2592
self.conv1 = Conv2d(in_shape[0], 16, 8, stride=4, padding=0, bias=bias)
self.conv2 = Conv2d(16, 32, 4, stride=2, padding=0, bias=bias)
if normalize:
self.bn1 = BatchNorm2d(16)
self.bn2 = BatchNorm2d(32)
else:
self.bn1 = Identity()
self.bn2 = Identity()
relu_gain = init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
def __init__(self, in_shape, normalize):
super().__init__()
bias = not normalize
self.conv1 = Conv2d(in_shape[0], 64, 7, stride=2, padding=1,
bias=bias) # 40x40
relu_gain = init.calculate_gain('relu')
self.conv1.weight.data.mul_(relu_gain)
if normalize:
self.bn1 = BatchNorm2d(64)
else:
self.bn1 = Identity()
def __init__(self, in_shape, normalize):
super().__init__()
bias = not normalize
self._nb_output_channel = 3136
self.conv1 = Conv2d(in_shape[0], 32, 8, stride=4, padding=0, bias=bias)
self.conv2 = Conv2d(32, 64, 4, stride=2, padding=0, bias=bias)
self.conv3 = Conv2d(64, 64, 3, stride=1, padding=0, bias=bias)
if normalize:
self.bn1 = BatchNorm2d(32)
self.bn2 = BatchNorm2d(64)
self.bn3 = BatchNorm2d(64)
else:
self.bn1 = Identity()
self.bn2 = Identity()
self.bn3 = Identity()
relu_gain = init.calculate_gain("relu")
self.conv1.weight.data.mul_(relu_gain)
self.conv2.weight.data.mul_(relu_gain)
self.conv3.weight.data.mul_(relu_gain)
def __init__(self, nb_input_channel, _, normalize):
super().__init__()
self._nb_output_channel = 256
bias = not normalize
self.linear = Linear(nb_input_channel,
self._nb_output_channel,
bias=bias)
if normalize:
self.bn_linear = BatchNorm1d(self._nb_output_channel)
else:
self.bn_linear = Identity()
def __init__(self, nb_input_channel, nb_out_channel, normalize):
super().__init__()
self._nb_output_channel = 256
self.linear = Linear(2592, self._nb_output_channel)
if normalize:
self.lstm = LSTMCellLayerNorm(
self._nb_output_channel,
self._nb_output_channel) # hack for experiment
self.bn_linear = BatchNorm1d(self._nb_output_channel)
else:
self.bn_linear = Identity()
self.lstm = LSTMCell(self._nb_output_channel,
self._nb_output_channel)
self.lstm.bias_ih.data.fill_(0)
self.lstm.bias_hh.data.fill_(0)
