def test_dropout_bprop_single_comp(transformer_factory):
nin, batch_size = 32, 2
# set inputs
N = ng.make_axis(batch_size, name='N')
F = ng.make_axis(nin).named('F')
mul_factor = ng.placeholder(())
inp = ng.placeholder([F, N])
layer = Dropout(keep=0.5)
fprop = layer.train_outputs(inp * mul_factor)
out_graph = ng.sum(fprop, out_axes=())
# create data
x = np.random.uniform(size=(nin, batch_size))
bprop = ng.deriv(out_graph, mul_factor)
# evaluate
trans = ngt.make_transformer()
comp = trans.computation([fprop, bprop, layer.mask], inp, mul_factor)
fout, bout, mask = comp(x, 2)
# Calculate derivative by hand and compare
np.testing.assert_allclose(bout, (x * mask[:, None]).sum(), rtol=1e-6)
trans.close()
def test_dropout_train(transformer_factory):
nin, batch_size = 32, 2
# set inputs
N = ng.make_axis(batch_size, name='N')
F = ng.make_axis(nin).named('F')
inp = ng.placeholder([F, N])
layer = Dropout(keep=0.5)
fprop = layer.train_outputs(inp)
# create data
x = np.random.uniform(size=(nin, batch_size))
# evaluate
with executor([fprop, layer.mask], inp) as comp:
out, mask = comp(x)
numpy_out = x * mask[:, None]
np.testing.assert_allclose(out, numpy_out, rtol=1e-6)
out1, mask1 = out.copy(), mask.copy()
out2, mask2 = comp(x)
assert (out1 != out2).any()
assert (mask1 != mask2).any()