def testEvalExpr(self):
x = symbolic.Symbol('x')
y = symbolic.Symbol('y')
xy = x * y
# Without symbol-to-value map.
self.assertEqual(xy, symbolic.ToStatic(xy))
self.assertEqual(xy, symbolic.ToTensor(xy))
with symbolic.SymbolToValueMap(symbolic.STATIC_VALUES, {x: 2, y: 3}):
self.assertEqual(symbolic.ToStatic(xy), 6)
# The inner map overrides the outer map.
with symbolic.SymbolToValueMap(symbolic.STATIC_VALUES, {x: 5, y: 6}):
self.assertEqual(symbolic.ToStatic(xy), 30)
# Back to the outer map.
self.assertEqual(symbolic.ToStatic(xy), 6)
# EvalExpr can also evaluate a symbolic expression to a
# Tensor.
a = tf.placeholder(tf.float32)
b = tf.placeholder(tf.float32)
with symbolic.SymbolToValueMap(symbolic.TENSOR_VALUES, {x: a, y: b}):
with symbolic.SymbolToValueMap(symbolic.STATIC_VALUES, {x: 2, y: 3}):
# Value maps of different types do not affect each other.
self.assertEqual(symbolic.ToStatic(xy), 6)
ab = symbolic.ToTensor(xy)
self.assertIsInstance(ab, tf.Tensor)
with self.session() as sess:
self.assertEqual(12, sess.run(ab, {a: 3, b: 4}))
# EvalExpr supports partial evaluation.
with symbolic.SymbolToValueMap(symbolic.STATIC_VALUES, {y: 3}):
x3 = symbolic.ToStatic(xy)
with symbolic.SymbolToValueMap(symbolic.STATIC_VALUES, {x: 9}):
self.assertEqual(27, symbolic.ToStatic(x3))
def FProp(self, theta, inputs):
"""Apply projection to inputs.
Args:
theta: A NestedMap object containing weights' values of this layer and its
children layers.
inputs: The inputs tensor. Shaped [..., input_dims].
Returns:
Projected inputs.
"""
p = self.params
with tf.name_scope(p.name):
computation_cost.Add(
self, 'flops',
tf.reduce_prod(tf.cast(tf.shape(inputs)[:-1], tf.int64)) *
tf.cast(symbolic.ToTensor(p.input_dims * p.output_dims), tf.int64) *
2)
return py_utils.ProjectLastDim(inputs, theta.w, p.input_dims,
p.output_dims)