def testReluGrad(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(t):
with tape_lib.GradientTape() as tape:
tape.watch(t)
result = unified_nn_ops.relu(t)
grads = tape.gradient(result, t)
return grads
with context_lib.set_default(get_immediate_execution_context()):
positive = TensorCastHelper(constant_op.constant([1.]))
negative = TensorCastHelper(constant_op.constant([-1.]))
model_fn = def_function.function(model)
func_output = model_fn(positive)
self.assertAllEqual(func_output.numpy(), [1.])
func_output = model_fn(negative)
self.assertAllEqual(func_output.numpy(), [0.])
eager_output = model(positive)
self.assertAllEqual(eager_output.numpy(), [1.])
eager_output = model(negative)
self.assertAllEqual(eager_output.numpy(), [0.])
def __call__(self, *args, **kwargs):
# Flatten arguments.
flat_args = nest.flatten(args, expand_composites=True)
flat_kwargs = nest.flatten(kwargs, expand_composites=True)
all_args = flat_args + flat_kwargs
# Trace
outer_ctx = context_lib.get_default()
ctx = NewTracingContext(self.name)
with context_lib.set_default(ctx):
# TODO(srbs): Iterating over list of inputs is a known performance
# bottleneck. Add a pybind API for this.
inputs = [ctx.AddParameter(arg.DataType()) for arg in all_args]
structured_args = nest.pack_sequence_as(args, inputs[:len(flat_args)])
structured_kwargs = nest.pack_sequence_as(kwargs, inputs[len(flat_args):])
structured_outputs = self._python_func(*structured_args,
**structured_kwargs)
py_outputs = nest.flatten(structured_outputs, expand_composites=True)
num_outputs = len(py_outputs)
# TODO(srbs): Drop Nones before calling Finalize.
finalized_f = ctx.Finalize(py_outputs)
outer_ctx.RegisterFunction(finalized_f)
# Build call op
call_op = outer_ctx.CreateOperation(self.name, "")
call_op.SetOpName(self.name)
for arg in all_args:
call_op.AddInput(arg)
call_op_outputs = call_op.Execute(num_outputs)
# Cleanup
outer_ctx.RemoveFunction(self.name)
return nest.pack_sequence_as(structured_outputs, call_op_outputs)
def testLog1p(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a):
return unified_math_ops.log1p(a)
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([1.]))
func_output = def_function.function(model)(a)
self.assertArrayNear(func_output.numpy(), [0.69314], 0.001)
eager_output = model(a)
self.assertArrayNear(eager_output.numpy(), [0.69314], 0.001)
def testNeg(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a):
return unified_math_ops.neg(a)
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([2.]))
func_output = def_function.function(model)(a)
self.assertAllEqual(func_output.numpy(), [-2.])
eager_output = model(a)
self.assertAllEqual(eager_output.numpy(), [-2.])
def testDivNoNan(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
return unified_math_ops.div_no_nan(a, b)
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([2.]))
b = TensorCastHelper(constant_op.constant([4.]))
func_output = def_function.function(model)(a, b)
self.assertArrayNear(func_output.numpy(), [0.5], 0.001)
eager_output = model(a, b)
self.assertArrayNear(eager_output.numpy(), [0.5], 0.001)
def testAdd(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
return math_ops.add(a, b)
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([1., 2.]))
b = TensorCastHelper(constant_op.constant([3., 4.]))
func_output = def_function.function(model)(a, b)
self.assertAllEqual(func_output.numpy(), [4., 6.])
eager_output = model(a, b)
self.assertAllEqual(eager_output.numpy(), [4., 6.])
def testAdd(self, use_tfrt, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
return math_ops.add(a, b)
eager_ctx = NewImmediateExecutionContext(use_tfrt)
with context_lib.set_default(eager_ctx):
a = eager_ctx.CreateFloatScalarHandle(1.)
b = eager_ctx.CreateFloatScalarHandle(2.)
func_output = def_function.function(model)(a, b)
self.assertAllEqual(func_output.numpy(), 3.0)
eager_output = model(a, b)
self.assertAllEqual(eager_output.numpy(), 3.0)
def _computeMnistMlpGrads(self, math_ops_lib, nn_ops_lib, backprop_lib,
cast, num_iters, hidden_layers, hidden_size,
batch_size):
batch_size = 1
image_size = 28 * 28
num_classes = 10
def model(x, hidden_weights, softmax_weight, labels):
with backprop_lib.GradientTape() as tape:
for weight in hidden_weights + [softmax_weight]:
tape.watch(weight)
for hidden_weight in hidden_weights:
x = math_ops_lib.mat_mul(x, hidden_weight)
x = nn_ops_lib.relu(x)
logits = math_ops_lib.mat_mul(x, softmax_weight)
loss = nn_ops_lib.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=labels)
grads = tape.gradient(loss, hidden_weights + [softmax_weight])
return grads
x = maybe_cast(array_ops.ones([batch_size, image_size]), cast)
hidden_weights = []
for i in range(hidden_layers):
hidden_weights.append(
maybe_cast(
random_ops.random_uniform(
[hidden_size if i else image_size, hidden_size]),
cast))
softmax_weight = maybe_cast(
random_ops.random_uniform([hidden_size, num_classes]), cast)
labels = maybe_cast(array_ops.zeros([batch_size], dtype=dtypes.int32),
cast)
with context_lib.set_default(get_immediate_execution_context()):
# Warm up.
for _ in range(10):
model(x, hidden_weights, softmax_weight, labels)
runtimes = timeit.repeat(
lambda: model(x, hidden_weights, softmax_weight, labels),
repeat=num_iters,
number=10)
return min(runtimes) / 10
def testLog1pGrad(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a):
with tape_lib.GradientTape() as tape:
tape.watch(a)
result = unified_math_ops.log1p(a)
grads = tape.gradient(result, a)
return grads
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([1.]))
func_outputs = def_function.function(model)(a)
self.assertArrayNear(func_outputs.numpy(), [0.5], 0.001)
eager_outputs = model(a)
self.assertArrayNear(eager_outputs.numpy(), [0.5], 0.001)
def testAddGrad(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
with tape_lib.GradientTape() as tape:
tape.watch(a)
tape.watch(b)
result = unified_math_ops.add(a, b)
grads = tape.gradient(result, [a, b])
return grads
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([1., 2.]))
b = TensorCastHelper(constant_op.constant([3., 4.]))
func_outputs = def_function.function(model)(a, b)
self.assertAllEqual(func_outputs[0].numpy(), [1.0, 1.0])
self.assertAllEqual(func_outputs[1].numpy(), [1.0, 1.0])
eager_outputs = model(a, b)
self.assertAllEqual(eager_outputs[0].numpy(), [1.0, 1.0])
self.assertAllEqual(eager_outputs[1].numpy(), [1.0, 1.0])
def testDivNoNanGrad(self, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
with tape_lib.GradientTape() as tape:
tape.watch(a)
tape.watch(b)
result = unified_math_ops.div_no_nan(a, b)
grads = tape.gradient(result, [a, b])
return grads
with context_lib.set_default(get_immediate_execution_context()):
a = TensorCastHelper(constant_op.constant([2.]))
b = TensorCastHelper(constant_op.constant([4.]))
func_outputs = def_function.function(model)(a, b)
self.assertArrayNear(func_outputs[0].numpy(), [0.25], 0.001)
self.assertArrayNear(func_outputs[1].numpy(), [-0.125], 0.001)
eager_outputs = model(a, b)
self.assertArrayNear(eager_outputs[0].numpy(), [0.25], 0.001)
self.assertArrayNear(eager_outputs[1].numpy(), [-0.125], 0.001)
def testAddGrad(self, use_tfrt, use_mlir):
if use_mlir:
SetTracingImplementation("mlir")
def model(a, b):
with tape_lib.GradientTape() as tape:
tape.watch(a)
tape.watch(b)
result = math_ops.add(a, b)
grads = tape.gradient(result, [a, b])
return grads
eager_ctx = NewImmediateExecutionContext(use_tfrt)
with context_lib.set_default(eager_ctx):
a = eager_ctx.CreateFloatScalarHandle(1.)
b = eager_ctx.CreateFloatScalarHandle(2.)
func_outputs = def_function.function(model)(a, b)
self.assertAllEqual(func_outputs[0].numpy(), 1.0)
self.assertAllEqual(func_outputs[1].numpy(), 1.0)
eager_outputs = model(a, b)
self.assertAllEqual(eager_outputs[0].numpy(), 1.0)
self.assertAllEqual(eager_outputs[1].numpy(), 1.0)
def __enter__(self): """Enters a context inside which operations are recorded on this tape.""" self._ctx_manager = context_stack.set_default(self._tape_context) self._ctx_manager.__enter__() return self
