def test_reduce():
sample_num = 200
def test_reduce_inner(numpy_reduce_func, nd_reduce_func):
for i in range(sample_num):
ndim = np.random.randint(1, 8)
shape = np.random.randint(1, 11, size=ndim)
axis_flags = np.random.randint(0, 2, size=ndim)
axes = []
for (axis, flag) in enumerate(axis_flags):
if flag:
axes.append(axis)
keepdims = np.random.randint(0, 2)
dat = np.random.rand(*shape) - 0.5
if 0 == len(axes):
axes = tuple(range(ndim))
else:
axes = tuple(axes)
numpy_ret = numpy_reduce_func(dat, axis=axes, keepdims=keepdims)
ndarray_ret = nd_reduce_func(arr=mx.nd.array(dat), axis=axes, keepdims=keepdims)
if type(ndarray_ret) is mx.ndarray.NDArray:
ndarray_ret = ndarray_ret.asnumpy()
assert ndarray_ret.shape == numpy_ret.shape
err = np.square(ndarray_ret - numpy_ret).mean()
assert err < 1E-4
test_reduce_inner(lambda data, axis, keepdims:_np_reduce(data, axis, keepdims, np.sum),
mx.nd.sum)
test_reduce_inner(lambda data, axis, keepdims:_np_reduce(data, axis, keepdims, np.max),
mx.nd.max)
test_reduce_inner(lambda data, axis, keepdims:_np_reduce(data, axis, keepdims, np.min),
mx.nd.min)
def test_reduce():
sample_num = 200
def test_reduce_inner(numpy_reduce_func, numpy_reduce_grad_func,
mx_reduce_sym):
for i in range(sample_num):
# Generate random data that has ndim between 1-7 and all the shape dims between 1-10
ndim = np.random.randint(1, 8)
shape = np.random.randint(1, 11, size=(ndim, ))
axis_num = np.random.randint(0, ndim, size=1)
axis_flags = np.random.randint(0, 2, size=ndim)
axes = []
for (axis, flag) in enumerate(axis_flags):
if flag:
axes.append(axis)
if 0 == len(axes):
axes = None
elif 1 == len(axes):
axes = axes[0]
else:
axes = tuple(axes)
keepdims = np.random.randint(0, 2)
a = mx.symbol.Variable('a')
if axes is None:
b = mx_reduce_sym(a, keepdims=keepdims)
else:
b = mx_reduce_sym(a, axis=axes, keepdims=keepdims)
dat_npy = np.random.rand(*shape)
sum_groundtruth = np.array(
numpy_reduce_func(dat_npy, axis=axes, keepdims=keepdims))
if sum_groundtruth.shape == ():
sum_groundtruth = np.array([sum_groundtruth])
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.array(np.random.rand(*sum_groundtruth.shape))
grad_groundtruth = numpy_reduce_grad_func(outgrad=outgrad_npy,
data=dat_npy,
axis=axes,
keepdims=keepdims)
net = b.bind(mx.cpu(),
args={'a': mx.nd.array(dat_npy)},
args_grad={'a': grad_nd})
net.forward(is_train=True)
err_forward = reldiff(net.outputs[0].asnumpy(), sum_groundtruth)
assert err_forward < 1E-4
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = reldiff(grad_nd.asnumpy(), grad_groundtruth)
assert err_backward < 1E-4
test_reduce_inner(
lambda data, axis, keepdims: _np_reduce(data, axis, keepdims, np.sum),
lambda outgrad, data, axis, keepdims: outgrad.reshape(
_np_reduce(data, axis, 1, np.sum).shape), mx.symbol.sum)
def test_broadcast_axis():
for i in range(sample_num):
# Generate random data that has ndim between 1-7 and all the shape dims between 1-10
ndim = np.random.randint(1, 8)
target_shape = np.random.randint(1, 11, size=(ndim,))
axis = np.random.randint(0, ndim)
shape = target_shape.copy()
size = shape[axis]
shape[axis] = 1
a = mx.symbol.Variable('a')
b = mx.symbol.broadcast_axis(a, axis=axis, size=size)
dat_npy = np.random.rand(*shape)
groundtruth = dat_npy
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.random.rand(*target_shape)
grad_groundtruth = _np_reduce(outgrad_npy, axis=axis, keepdims=True,
numpy_reduce_func=np.sum)
net = b.bind(mx.cpu(), args={'a': mx.nd.array(dat_npy)},
args_grad={'a': grad_nd})
net.forward(is_train=True)
assert (net.outputs[0].shape == target_shape).all()
err_forward = np.square(net.outputs[0].asnumpy() - groundtruth).mean()
assert err_forward < 1E-8
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = np.square(grad_nd.asnumpy() - grad_groundtruth).mean()
assert err_backward < 1E-8
def test_broadcast_axis():
for i in range(sample_num):
# Generate random data that has ndim between 1-7 and all the shape dims between 1-10
ndim = np.random.randint(1, 8)
target_shape = np.random.randint(1, 11, size=(ndim, ))
axis = np.random.randint(0, ndim)
shape = target_shape.copy()
size = shape[axis]
shape[axis] = 1
a = mx.symbol.Variable('a')
b = mx.symbol.broadcast_axis(a, axis=axis, size=size)
dat_npy = np.random.rand(*shape)
groundtruth = dat_npy
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.random.rand(*target_shape)
grad_groundtruth = _np_reduce(outgrad_npy,
axis=axis,
keepdims=True,
numpy_reduce_func=np.sum)
net = b.bind(mx.cpu(),
args={'a': mx.nd.array(dat_npy)},
args_grad={'a': grad_nd})
net.forward(is_train=True)
assert (net.outputs[0].shape == target_shape).all()
err_forward = np.square(net.outputs[0].asnumpy() -
groundtruth).mean()
assert err_forward < 1E-8
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = np.square(grad_nd.asnumpy() -
grad_groundtruth).mean()
assert err_backward < 1E-8
def test_reduce():
sample_num = 200
def test_reduce_inner(numpy_reduce_func, numpy_reduce_grad_func, mx_reduce_sym):
for i in range(sample_num):
# Generate random data that has ndim between 1-7 and all the shape dims between 1-10
ndim = np.random.randint(1, 8)
shape = np.random.randint(1, 11, size=(ndim,))
axis_num = np.random.randint(0, ndim, size=1)
axis_flags = np.random.randint(0, 2, size=ndim)
axes = []
for (axis, flag) in enumerate(axis_flags):
if flag:
axes.append(axis)
if 0 == len(axes):
axes = None
elif 1 == len(axes):
axes = axes[0]
else:
axes = tuple(axes)
keepdims = np.random.randint(0, 2)
a = mx.symbol.Variable("a")
if axes is None:
b = mx_reduce_sym(a, keepdims=keepdims)
else:
b = mx_reduce_sym(a, axis=axes, keepdims=keepdims)
dat_npy = np.random.rand(*shape)
sum_groundtruth = np.array(numpy_reduce_func(dat_npy, axis=axes, keepdims=keepdims))
if sum_groundtruth.shape == ():
sum_groundtruth = np.array([sum_groundtruth])
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.array(np.random.rand(*sum_groundtruth.shape))
grad_groundtruth = numpy_reduce_grad_func(outgrad=outgrad_npy, data=dat_npy, axis=axes, keepdims=keepdims)
net = b.bind(mx.cpu(), args={"a": mx.nd.array(dat_npy)}, args_grad={"a": grad_nd})
net.forward(is_train=True)
err_forward = np.square(net.outputs[0].asnumpy() - sum_groundtruth).sum() / np.prod(shape)
assert err_forward < 1e-6
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = np.square(grad_nd.asnumpy() - grad_groundtruth).sum()
assert err_backward < 1e-6
test_reduce_inner(
lambda data, axis, keepdims: _np_reduce(data, axis, keepdims, np.sum),
lambda outgrad, data, axis, keepdims: outgrad.reshape(_np_reduce(data, axis, 1, np.sum).shape),
mx.symbol.sum,
)
def test_broadcasting_ele(sym_bcast):
dat_npy = np.random.rand(*shape)
groundtruth = dat_npy
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.random.rand(*target_shape)
grad_groundtruth = _np_reduce(outgrad_npy, axis=axis, keepdims=True,
numpy_reduce_func=np.sum)
net = sym_bcast.bind(mx.cpu(), args={'a': mx.nd.array(dat_npy)},
args_grad={'a': grad_nd})
net.forward(is_train=True)
assert (net.outputs[0].shape == target_shape).all()
err_forward = reldiff(net.outputs[0].asnumpy(), groundtruth)
assert err_forward < 1E-4
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = reldiff(grad_nd.asnumpy(), grad_groundtruth)
assert err_backward < 1E-4
def test_broadcasting_ele(sym_bcast):
dat_npy = np.random.rand(*shape)
groundtruth = dat_npy
grad_nd = mx.nd.empty(shape)
outgrad_npy = np.random.rand(*target_shape)
grad_groundtruth = _np_reduce(outgrad_npy, axis=axis, keepdims=True,
numpy_reduce_func=np.sum)
net = sym_bcast.bind(mx.cpu(), args={'a': mx.nd.array(dat_npy)},
args_grad={'a': grad_nd})
net.forward(is_train=True)
assert (net.outputs[0].shape == target_shape).all()
err_forward = reldiff(net.outputs[0].asnumpy(), groundtruth)
assert err_forward < 1E-4
net.backward(out_grads=mx.nd.array(outgrad_npy))
err_backward = reldiff(grad_nd.asnumpy(), grad_groundtruth)
assert err_backward < 1E-4
