def test_bprop(self):
r = []
context = Context()
for i in xrange(self.N):
a = self.get_random_array()
a_gpu = Connector(GpuMatrix.from_npa(a, 'float'), bu_device_id=context)
vpooling_block = MeanPoolingBlock(a_gpu, axis=0)
voutput, dL_dvoutput = vpooling_block.output.register_usage(context, context)
_dL_voutput = self.get_random_array((dL_dvoutput.nrows, dL_dvoutput.ncols))
GpuMatrix.from_npa(_dL_voutput, 'float').copy_to(context, dL_dvoutput)
hpooling_block = MeanPoolingBlock(a_gpu, axis=1)
houtput, dL_dhoutput = hpooling_block.output.register_usage(context, context)
_dL_houtput = self.get_random_array((dL_dhoutput.nrows, dL_dhoutput.ncols))
GpuMatrix.from_npa(_dL_houtput, 'float').copy_to(context, dL_dhoutput)
vpooling_block.fprop()
vpooling_block.bprop()
dL_dmatrix = vpooling_block.dL_dmatrix.to_host()
r.append(np.allclose(dL_dmatrix,
np.repeat(_dL_voutput/a.shape[0], a.shape[0], 0),
atol=1e-6))
hpooling_block.fprop()
hpooling_block.bprop()
hpooling_block.dL_dmatrix.to_host()
dL_dmatrix = hpooling_block.dL_dmatrix.to_host()
r.append(np.allclose(dL_dmatrix,
np.repeat(_dL_houtput/a.shape[1], a.shape[1], 1),
atol=1e-6))
self.assertEqual(sum(r), 2 * self.N)
def test_fprop(self):
r = []
for i in xrange(self.N):
a = self.get_random_array()
a_gpu = Connector(GpuMatrix.from_npa(a, 'float'))
vpooling_block = MeanPoolingBlock(a_gpu, axis=0)
hpooling_block = MeanPoolingBlock(a_gpu, axis=1)
vpooling_block.fprop()
r.append(
np.allclose(vpooling_block.output.to_host(),
np.mean(a, axis=0, keepdims=True),
atol=1e-6))
hpooling_block.fprop()
r.append(
np.allclose(hpooling_block.output.to_host(),
np.mean(a, axis=1, keepdims=True),
atol=1e-6))
self.assertEqual(sum(r), 2 * self.N)
def test_fprop(self):
r = []
for i in xrange(self.N):
a = self.get_random_array()
a_gpu = Connector(GpuMatrix.from_npa(a, 'float'))
vpooling_block = MeanPoolingBlock(a_gpu, axis=0)
hpooling_block = MeanPoolingBlock(a_gpu, axis=1)
vpooling_block.fprop()
r.append(np.allclose(vpooling_block.output.to_host(),
np.mean(a, axis=0, keepdims=True),
atol=1e-6))
hpooling_block.fprop()
r.append(np.allclose(hpooling_block.output.to_host(),
np.mean(a, axis=1, keepdims=True),
atol=1e-6))
self.assertEqual(sum(r), 2 * self.N)
def test_bprop(self):
r = []
context = Context()
for i in xrange(self.N):
a = self.get_random_array()
a_gpu = Connector(GpuMatrix.from_npa(a, 'float'),
bu_device_id=context)
vpooling_block = MeanPoolingBlock(a_gpu, axis=0)
voutput, dL_dvoutput = vpooling_block.output.register_usage(
context, context)
_dL_voutput = self.get_random_array(
(dL_dvoutput.nrows, dL_dvoutput.ncols))
GpuMatrix.from_npa(_dL_voutput,
'float').copy_to(context, dL_dvoutput)
hpooling_block = MeanPoolingBlock(a_gpu, axis=1)
houtput, dL_dhoutput = hpooling_block.output.register_usage(
context, context)
_dL_houtput = self.get_random_array(
(dL_dhoutput.nrows, dL_dhoutput.ncols))
GpuMatrix.from_npa(_dL_houtput,
'float').copy_to(context, dL_dhoutput)
vpooling_block.fprop()
vpooling_block.bprop()
dL_dmatrix = vpooling_block.dL_dmatrix.to_host()
r.append(
np.allclose(dL_dmatrix,
np.repeat(_dL_voutput / a.shape[0], a.shape[0], 0),
atol=1e-6))
hpooling_block.fprop()
hpooling_block.bprop()
hpooling_block.dL_dmatrix.to_host()
dL_dmatrix = hpooling_block.dL_dmatrix.to_host()
r.append(
np.allclose(dL_dmatrix,
np.repeat(_dL_houtput / a.shape[1], a.shape[1], 1),
atol=1e-6))
self.assertEqual(sum(r), 2 * self.N)