Add benchmark for Adam

benchmarks/optimizers/__init__.py

@@ -0,0 +1,67 @@
+import numpy
+import six
+
+import chainer
+import chainer.functions as F
+from chainer import initializers
+import chainer.links as L
+
+from benchmarks import BenchmarkBase
+
+
+class LinearModel(object):
+
+    def __init__(self, xp, optimizer, unit_num, dtype):
+        self.dtype = dtype
+        weight = initializers.HeNormal(1 / numpy.sqrt(2), dtype)
+        bias = initializers.Constant(0, dtype)
+        self.model = L.Linear(unit_num, 2, initialW=weight, initial_bias=bias)
+        self.optimizer = optimizer
+        optimizer.setup(self.model)
+
+        if xp != numpy:
+            self.model.to_gpu()
+
+
+class OptimizerBenchmark(BenchmarkBase):
+
+    """The base class for benchmark of optimizers."""
+
+    # Call `test_*` methods only once as `backward()` has a side-effect.
+    number = 1
+
+    # Repeat the test for 10 times instead of 3 (`timeit.default_repeat`).
+    repeat = 10
+
+    def _make_dataset(self, batch_size, unit_num, dtype):
+        xp = self.xp
+        x = xp.random.uniform(-1, 1, (batch_size, unit_num)).astype(dtype)
+        w = xp.random.uniform(-1, 1, (unit_num, 1)).astype(dtype)
+        b = xp.random.uniform(-1, 1, (1, )).astype(dtype)
+        a = (xp.dot(x, w) + b).reshape((batch_size, ))
+        t = (a < 0).astype(numpy.int32)
+        return chainer.Variable(x), chainer.Variable(t)
+
+    def setup_benchmark(self, optimizer, batch_size, unit_num, dtype):
+        """Performs setup of benchmark for optimizers.
+
+        Call this in `setup` method of your benchmark class.
+        Note that this function performs forward computation.
+        """
+
+        self.optimizer = optimizer
+
+        model = LinearModel(self.xp, optimizer, unit_num, dtype)
+        model.model.cleargrads()
+        x, t = self._make_dataset(batch_size, unit_num, dtype)
+        y = model.model(x)
+        loss = F.softmax_cross_entropy(y, t)
+        loss.backward()
+
+    def update(self, n_times):
+        """Runs optimizer.update()."""
+
+        optimizer = self.optimizer
+
+        for i in six.moves.range(n_times):
+            optimizer.update()

benchmarks/optimizers/adam.py

@@ -0,0 +1,23 @@
+import numpy
+
+from chainer import optimizers
+
+from benchmarks.optimizers import OptimizerBenchmark
+from benchmarks.utils import backends
+from benchmarks.utils import parameterize
+
+
+@backends('gpu', 'cpu')
+@parameterize(
+    [('dtype', [numpy.float32, numpy.float64]),
+     ('amsgrad', [True, False])])
+class Adam(OptimizerBenchmark):
+    def setup(self, dtype, amsgrad):
+        unit_num = 100000
+        batch_size = 32
+        optimizer = optimizers.Adam(0.05, amsgrad=amsgrad)
+
+        self.setup_benchmark(optimizer, batch_size, unit_num, dtype)
+
+    def time_update(self, dtype, amsgrad):
+        self.update(1000)