def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.activation = activation_factory(self.config["activation"])
self.conv1 = nn.Conv2d(self.config["in_channels"],
16,
kernel_size=2,
stride=2)
self.conv2 = nn.Conv2d(16, 32, kernel_size=2, stride=2)
self.conv3 = nn.Conv2d(32, 64, kernel_size=2, stride=2)
# MLP Head
# Number of Linear input connections depends on output of conv2d layers
# and therefore the input image size, so compute it.
def conv2d_size_out(size, kernel_size=2, stride=2):
return (size - (kernel_size - 1) - 1) // stride + 1
convw = conv2d_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_width"])))
convh = conv2d_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_height"])))
assert convh > 0 and convw > 0
self.config["head_mlp"]["in"] = convw * convh * 64
self.config["head_mlp"]["out"] = self.config["out"]
self.head = model_factory(self.config["head_mlp"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.config["base_module"]["in"] = self.config["in"]
self.base_module = model_factory(self.config["base_module"])
self.advantage = nn.Linear(self.config["base_module"]["layers"][-1],
self.config["out"])
self.value = nn.Linear(self.config["base_module"]["layers"][-1], 1)
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
sizes = [self.config["in"]] + self.config["layers"]
self.activation = activation_factory(self.config["activation"])
layers_list = [nn.Linear(sizes[i], sizes[i + 1]) for i in range(len(sizes) - 1)]
self.layers = nn.ModuleList(layers_list)
if self.config.get("out", None):
self.predict = nn.Linear(sizes[-1], self.config["out"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.features_per_head = int(self.config["feature_size"] / self.config["heads"])
self.value_all = nn.Linear(self.config["feature_size"], self.config["feature_size"], bias=False)
self.key_all = nn.Linear(self.config["feature_size"], self.config["feature_size"], bias=False)
self.query_all = nn.Linear(self.config["feature_size"], self.config["feature_size"], bias=False)
self.attention_combine = nn.Linear(self.config["feature_size"], self.config["feature_size"], bias=False)
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
# sizes = [self.config["in"]] + self.config["layers"] + [self.config["out"]]
# self.activation = activation_factory(self.config["activation"])
# layers_list = [GCNConv(sizes[i], sizes[i + 1]) for i in range(len(sizes) - 1)]
# self.layers = nn.ModuleList(layers_list)
self.conv1 = GCNConv(self.config["in"], self.config["layers"])
self.conv2 = GCNConv(self.config["layers"], self.config["out"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.config = config
if not self.config["embedding_layer"]["in"]:
self.config["embedding_layer"]["in"] = self.config["in"]
self.config["output_layer"]["in"] = self.config["attention_layer"]["feature_size"]
self.config["output_layer"]["out"] = self.config["out"]
self.embedding = model_factory(self.config["embedding_layer"])
self.attention_layer = SelfAttention(self.config["attention_layer"])
self.output_layer = model_factory(self.config["output_layer"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.activation = activation_factory(self.config["activation"])
self.conv1 = nn.Conv3d(self.config["in_channels"],
32,
kernel_size=(1, 8, 8),
stride=(1, 4, 4))
self.conv2 = nn.Conv3d(32, 64, kernel_size=(3, 4, 4), stride=(1, 2, 2))
self.conv3 = nn.Conv3d(64, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1))
# MLP Head
# Number of Linear input connections depends on output of conv2d layers
# and therefore the input image size, so compute it.
def conv2d_size_out(size, kernel_size=3, stride=1):
return (size - (kernel_size - 1) - 1) // stride + 1
convw = conv2d_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_width"],
kernel_size=8,
stride=4),
kernel_size=4,
stride=2))
convh = conv2d_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_height"],
kernel_size=8,
stride=4),
kernel_size=4,
stride=2))
convd = conv2d_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_depth"],
kernel_size=3,
stride=1),
kernel_size=1,
stride=1))
assert convh > 0 and convw > 0
self.config["head_mlp"]["in"] = convw * convh * convd * 64
self.config["head_mlp"]["out"] = self.config["out"]
self.fc1 = nn.Linear(self.config["head_mlp"]["in"], 512)
self.fc2 = nn.Linear(512, self.config["head_mlp"]["out"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.activation = activation_factory(self.config["activation"])
self.conv1 = nn.Conv2d(self.config["in_channels"],
32,
kernel_size=3,
stride=1)
self.conv2 = nn.Conv2d(32, 32, kernel_size=3, stride=1)
self.bn1 = nn.BatchNorm2d(32)
self.conv3 = nn.Conv2d(32, 32, kernel_size=3, stride=1)
self.conv4 = nn.Conv2d(32, 32, kernel_size=3, stride=1)
self.bn2 = nn.BatchNorm2d(32)
self.conv5 = nn.Conv2d(32, 32, kernel_size=3, stride=1)
self.maxpool1 = nn.MaxPool2d(3, stride=2)
# MLP Head
# Number of Linear input connections depends on output of conv2d layers
# and therefore the input image size, so compute it.
def conv2d_size_out(size, kernel_size=3, stride=1):
return (size - (kernel_size - 1) - 1) // stride + 1
def maxpool_size_out(size, kernel_size=3, stride=2, padding=0):
return (size + 2 * padding - (kernel_size - 1) - 1) // stride + 1
convw = conv2d_size_out(
conv2d_size_out(
maxpool_size_out(
conv2d_size_out(
conv2d_size_out(
conv2d_size_out(self.config["in_width"]))))))
convh = conv2d_size_out(
conv2d_size_out(
maxpool_size_out(
conv2d_size_out(
conv2d_size_out(
conv2d_size_out(self.config["in_height"]))))))
assert convh > 0 and convw > 0
self.config["head_mlp"]["in"] = convw * convh * 32
self.config["head_mlp"]["out"] = self.config["out"]
self.head = model_factory(self.config["head_mlp"])
def __init__(self, config):
super().__init__()
Configurable.__init__(self, config)
self.activation = activation_factory(self.config["activation"])
self.conv1 = nn.Conv3d(self.config["in_channels"],
32,
kernel_size=(1, 7, 7),
stride=(1, 1, 1),
padding=(0, 3, 3))
self.conv2 = nn.Conv3d(32,
32,
kernel_size=(3, 5, 5),
stride=(1, 1, 1),
padding=(1, 2, 2))
self.conv3 = nn.Conv3d(32, 64, kernel_size=(3, 5, 5), stride=(1, 1, 1))
self.conv4 = nn.Conv3d(64, 64, kernel_size=(1, 5, 5), stride=(1, 1, 1))
self.maxpool1 = nn.MaxPool3d(kernel_size=(3, 5, 5), stride=(1, 2, 2))
self.bn1 = nn.BatchNorm3d(64)
# MLP Head
# Number of Linear input connections depends on output of conv2d layers
# and therefore the input image size, so compute it.
def conv2d_size_out(size, kernel_size=5, stride=1, padding=0):
return (size + 2 * padding - (kernel_size - 1) - 1) // stride + 1
def maxpool_size_out(size, kernel_size=5, stride=2, padding=0):
return (size + 2 * padding - (kernel_size - 1) - 1) // stride + 1
convw = conv2d_size_out(
conv2d_size_out(
maxpool_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_width"],
kernel_size=7,
stride=1,
padding=3),
kernel_size=5,
stride=1,
padding=2))))
convh = conv2d_size_out(
conv2d_size_out(
maxpool_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_height"],
kernel_size=7,
stride=1,
padding=3),
kernel_size=5,
stride=1,
padding=2))))
convd = conv2d_size_out(conv2d_size_out(maxpool_size_out(
conv2d_size_out(conv2d_size_out(self.config["in_depth"],
kernel_size=1,
stride=1,
padding=0),
kernel_size=3,
stride=1,
padding=1),
kernel_size=3,
stride=1),
kernel_size=3,
stride=1),
kernel_size=1,
stride=1)
assert convh > 0 and convw > 0
self.config["head_mlp"]["in"] = convw * convh * convd * 64
self.config["head_mlp"]["out"] = self.config["out"]
self.fc1 = nn.Linear(self.config["head_mlp"]["in"], 256)
self.fc2 = nn.Linear(256, self.config["head_mlp"]["out"])
