Layer

source

Layer(
   inputs, n_units, shape = None, dtype = None, name = 'layer', **kwargs
)

---

Layer Base Class

Passing Attributes

All keyword attributes passed to this constructor will be set as instance attributes so a common case for the implementing class might be:

class CustomLayer(Layer):
    def __init__(layer, n_units, param=1):
        # for the linter
        self.param = param
        # ...
        super().__init__(inputs=layer, n_units=n_units, param=value)

in = tx.Input(10)
y = CustomLayer(in, 4, param=2)
assert y.param == 2

Attributes

• inputs (Sequence[Layer]) : a list of input nodes for the current layer
• n_units : the number of units for the current layer (last dim)
• name (str) : name to be used for the layer scope
• config (LayerConfig) : a layer configuration with the arguments used in the current layer instance
• scoped_name (str) : layer full scope name

Args

• inputs (Sequence[Layer]) : a single layer,a list of input layers, or None if no inputs are required
• n_units (int) : dimension of input vector (dimension of columns in case batch_size != None
• dtype (DType) : dtype for the current layer output
• shape (TensorShape) : output shape. If not None overrides compute_shape
• name (str) : layer name (used to nam the placeholder)
• kwargs (Any) : other keyword args to be set as instance attributes

Methods:

.compute_shape

source

.compute_shape()

---

called before init_state

Returns

• shape (tf.TensorShape) : best guess for the output shape of the layer

.init_state

source

.init_state()

---

init_state meant to be overriden in subclasses

Creates an empty LayerState object

Overriding init_state()

Classes implementing Layer should override this method

def init_state(self):
    state = super().init_state()
    # or state = LayerState()
    state.var1 = var1
    state.var2 = var2
    return state

Layer will take this state object and add var1 and var2 to attributes.

Returns

• state (LayerState) : current layer state object

.compute

source

.compute(
   *args
)

.as_function

source

.as_function(
   name = 'layer_function', compile = False
)

---

returns a python function of a Tensorflow compiled graph as a callable

Note

This returns the entire graph as a function that terminates on this layer. If you want the function for this layer alone just get the tf.function(layer.compute)

Args

• name (str) : function name to be returned
• compile (bool) : if True, returns a Tensorflow compile graph as a callable else, returns a python function.

Returns

• fn (Callable) : either a Tensorflow static graph or a python callable function.

.reuse_with

source

.reuse_with(
   *layers, **kwargs
)