def run_gradweight(self, inputs_shape, filters_shape, dCdH_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs, dCdH=dCdH,
WShape=filters_shape,
d=subsample)
img = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
topgrad = gpu_contiguous(dCdH.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(img,
topgrad)
else:
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(img,
topgrad,
shape=filters_shape[1:4])
conv_gemm = conv_gemm.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv)
f = theano.function([], conv_gemm)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def run_gradinput(self, inputs_shape, filters_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bottom_height = (inputs_shape[1] - 1) * subsample[0] + filters_shape[1]
bottom_width = (inputs_shape[2] - 1) * subsample[1] + filters_shape[2]
bottom_depth = (inputs_shape[3] - 1) * subsample[2] + filters_shape[3]
bottom_shape = theano.shared(numpy.array([bottom_height, bottom_width, bottom_depth]))
weight = gpu_contiguous(filters.dimshuffle(0, 4, 1, 2, 3))
top = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_ref = Corr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top)
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top)
else:
conv_ref = Corr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top,
shape=bottom_shape)
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top,
shape=bottom_shape)
f_ref = theano.function([], conv_ref, mode='FAST_RUN')
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def test_opt_convtransp3d_gemm(self):
inputs_shape = (16, 15, 12, 12, 10)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.convTransp3D(W=filters,
b=bias,
d=(1, 1, 1),
H=inputs)
mode = mode_with_gpu.including('convtransp3d_gemm')
f_ref = theano.function([], conv)
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM_gradInputs) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def test_opt_conv3d_fft(self):
inputs_shape = (16, 20, 32, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
conv = theano.tensor.nnet.conv3D(V=inputs,
W=filters,
b=bias,
d=(1, 1, 1))
ref_mode = copy.copy(theano.compile.get_default_mode())
ref_mode.check_py_code = False
mode = mode_with_gpu.including('conv3d_fft')
mode.check_py_code = False
f_ref = theano.function([], conv, mode=ref_mode)
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(
isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft)
def test_opt_nofft_full(self):
inputs_shape = (5, 3, 7, 6)
filters_shape = (2, 3, 3, 3)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.conv.conv2d(inputs,
filters,
border_mode='full',
version='no_fft')
mode = mode_with_gpu.including('conv_fft_full')
f_ref = theano.function([], conv)
f_fft = theano.function([], conv, mode=mode)
# make sure we that no CuFFTOp has been inserted
topo = f_fft.maker.fgraph.toposort()
assert sum(
isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 0
def run_conv_full(self, inputs_shape, filters_shape, pad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
conv_ref = theano.tensor.nnet.convTransp3D(W=filters,
b=bias,
d=(1, 1, 1),
H=inputs)
filters = filters.dimshuffle(4, 0, 1, 2, 3)
inputs = inputs.dimshuffle(0, 4, 1, 2, 3)
conv_fft = theano.sandbox.cuda.fftconv.conv3d_fft(inputs,
filters,
border_mode="full",
pad_last_dim=pad)
conv_fft = conv_fft.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref)
f_fft = theano.function([], conv_fft, mode=mode_with_gpu)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def test_opt_convgrad3d_gemm(self):
inputs_shape = (16, 10, 12, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
dCdH_shape = (16, 5, 1, 13, 10)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs,
dCdH=dCdH,
WShape=filters_shape,
d=(1, 1, 1))
mode = mode_with_gpu.including('convgrad3d_gemm')
f_ref = theano.function([], conv)
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM_gradWeights) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def test_opt_full(self):
inputs_shape = (5, 3, 7, 6)
filters_shape = (2, 3, 3, 3)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.conv.conv2d(inputs, filters,
border_mode='full')
mode = mode_with_gpu.including('conv_fft_full')
f_ref = theano.function([], conv)
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2, topo
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft)
def test_opt_conv3d_gemm(self):
inputs_shape = (16, 20, 32, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
conv = theano.tensor.nnet.conv3D(V=inputs, W=filters,
b=bias, d=(1, 1, 1))
mode = mode_with_gpu.including('conv3d_gemm')
mode.check_py_code = False
f_ref = theano.function([], conv, mode="FAST_RUN")
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def test_opt_full(self):
inputs_shape = (5, 3, 7, 6)
filters_shape = (2, 3, 3, 3)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.conv.conv2d(inputs, filters,
border_mode='full')
mode = mode_with_gpu.including('conv_fft_full')
f_ref = theano.function([], conv)
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2, topo
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft)
def test_opt_convgrad3d_fft(self):
inputs_shape = (2, 17, 15, 16, 1)
filters_shape = (10, 6, 7, 4, 1)
dCdH_shape = (inputs_shape[0],
inputs_shape[1] - filters_shape[1] + 1,
inputs_shape[2] - filters_shape[2] + 1,
inputs_shape[3] - filters_shape[3] + 1,
filters_shape[0])
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs, dCdH=dCdH,
WShape=filters_shape,
d=(1, 1, 1))
mode = mode_with_gpu.including('convgrad3d_fft')
mode.check_py_code = False
f_ref = theano.function([], conv, mode="FAST_RUN")
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def test_opt_convtransp3d_fft(self):
inputs_shape = (2, 9, 16, 12, 10)
filters_shape = (10, 3, 8, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.convTransp3D(W=filters, b=bias, d=(1, 1, 1),
H=inputs)
mode = mode_with_gpu.including('convtransp3d_fft')
f_ref = theano.function([], conv)
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def run_conv_valid(self, inputs_shape, filters_shape, pad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Flip filter as conv3D compute correlation
filters_flip = filters[:, ::-1, ::-1, ::-1, :]
# filters_flip = filters
conv_ref = theano.tensor.nnet.conv3D(V=inputs, W=filters_flip,
b=bias, d=(1, 1, 1))
conv_fft = theano.sandbox.cuda.fftconv.conv3d_fft(
inputs.dimshuffle(0, 4, 1, 2, 3),
filters.dimshuffle(0, 4, 1, 2, 3),
border_mode="valid",
pad_last_dim=pad)
conv_fft = conv_fft.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref, mode="FAST_RUN")
mode = mode_with_gpu
mode.check_py_code = False
f_fft = theano.function([], conv_fft, mode=mode)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-05, atol=1e-05)
def run_conv_full(self, inputs_shape, filters_shape, pad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
conv_ref = theano.tensor.nnet.convTransp3D(
W=filters, b=bias, d=(1, 1, 1),
H=inputs)
filters = filters.dimshuffle(4, 0, 1, 2, 3)
inputs = inputs.dimshuffle(0, 4, 1, 2, 3)
conv_fft = theano.sandbox.cuda.fftconv.conv3d_fft(inputs, filters,
border_mode="full",
pad_last_dim=pad)
conv_fft = conv_fft.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref)
f_fft = theano.function([], conv_fft, mode=mode_with_gpu)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def test_opt_conv3d_gemm(self):
inputs_shape = (16, 20, 32, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
conv = theano.tensor.nnet.conv3D(V=inputs,
W=filters,
b=bias,
d=(1, 1, 1))
mode = mode_with_gpu.including('conv3d_gemm')
mode.check_py_code = False
f_ref = theano.function([], conv, mode="FAST_RUN")
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def test_opt_convtransp3d_gemm(self):
inputs_shape = (16, 15, 12, 12, 10)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.convTransp3D(W=filters, b=bias, d=(1, 1, 1),
H=inputs)
mode = mode_with_gpu.including('convtransp3d_gemm')
f_ref = theano.function([], conv)
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM_gradInputs) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def run_gradinput(self, inputs_shape, filters_shape, subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
conv = theano.tensor.nnet.convTransp3D(W=filters,
b=bias,
d=subsample,
H=inputs)
f_ref = theano.function([], conv)
res_ref = f_ref()
# Get bottom shape using convTransp3D
bottom_shape = res_ref.shape
bottom_val = numpy.random.random(bottom_shape).astype('float32')
bottom = shared(bottom_val)
weight = gpu_contiguous(filters.dimshuffle(0, 4, 1, 2, 3))
top = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top)
else:
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top, shape=bottom.shape[1:4])
conv_gemm = conv_gemm.dimshuffle(0, 2, 3, 4, 1)
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res = f()
utt.assert_allclose(res_ref, res)
def run_gradweight(self,
inputs_shape,
filters_shape,
dCdH_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
img = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
topgrad = gpu_contiguous(dCdH.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_ref = Corr3dMM_gradWeights(subsample=subsample)(img, topgrad)
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(img,
topgrad)
else:
conv_ref = GpuCorr3dMM_gradWeights(subsample=subsample)(
img, topgrad, shape=filters_shape[1:4])
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(
img, topgrad, shape=filters_shape[1:4])
f_ref = theano.function([], conv_ref)
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def test_opt_convtransp3d_fft(self):
inputs_shape = (16, 15, 21, 12, 10)
filters_shape = (10, 6, 12, 4, 1)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.convTransp3D(W=filters, b=bias, d=(1,1,1),
H=inputs)
mode = mode_with_gpu.including('convtransp3d_fft')
f_ref = theano.function([], conv)
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def test_opt_convgrad3d_gemm(self):
inputs_shape = (16, 10, 12, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
dCdH_shape = (16, 5, 1, 13, 10)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs, dCdH=dCdH,
WShape=filters_shape,
d=(1, 1, 1))
mode = mode_with_gpu.including('convgrad3d_gemm')
f_ref = theano.function([], conv)
f_gemm = theano.function([], conv, mode=mode)
# make sure we inserted the gemm trickery
topo = f_gemm.maker.fgraph.toposort()
assert sum(isinstance(n.op, GpuCorr3dMM_gradWeights) for n in topo) > 0
res_ref = f_ref()
res_gemm = f_gemm()
utt.assert_allclose(res_ref, res_gemm)
def run_gradinput(self, inputs_shape, filters_shape, subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bottom_height = (inputs_shape[1] - 1) * subsample[0] + filters_shape[1]
bottom_width = (inputs_shape[2] - 1) * subsample[1] + filters_shape[2]
bottom_depth = (inputs_shape[3] - 1) * subsample[2] + filters_shape[3]
bottom_shape = theano.shared(
numpy.array([bottom_height, bottom_width, bottom_depth]))
weight = gpu_contiguous(filters.dimshuffle(0, 4, 1, 2, 3))
top = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_ref = Corr3dMM_gradInputs(subsample=subsample)(kern=weight,
topgrad=top)
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top)
else:
conv_ref = Corr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top, shape=bottom_shape)
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top, shape=bottom_shape)
f_ref = theano.function([], conv_ref)
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def test_opt_convgrad3d_fft(self):
inputs_shape = (16, 20, 32, 16, 1)
filters_shape = (10, 6, 12, 4, 1)
dCdH_shape = (16, 15, 21, 13, 10)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs, dCdH=dCdH,
WShape=filters_shape,
d=(1,1,1))
mode = mode_with_gpu.including('convgrad3d_fft')
mode.check_py_code = False
f_ref = theano.function([], conv, mode="FAST_RUN")
f_fft = theano.function([], conv, mode=mode)
# make sure we inserted the fft trickery
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 2
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-04, atol=1e-04)
def run_conv_valid(self, inputs_shape, filters_shape, pad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Flip filter as conv3D compute correlation
filters_flip = filters[:,::-1,::-1,::-1,:]
#filters_flip = filters
conv_ref = theano.tensor.nnet.conv3D(V=inputs, W=filters_flip,
b=bias, d=(1,1,1))
conv_fft = theano.sandbox.cuda.fftconv.conv3d_fft(inputs.dimshuffle(0, 4, 1, 2, 3),
filters.dimshuffle(0, 4, 1, 2, 3),
border_mode = "valid",
pad_last_dim = pad)
conv_fft = conv_fft.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref, mode="FAST_RUN")
mode = mode_with_gpu
mode.check_py_code = False
f_fft = theano.function([], conv_fft, mode=mode)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft, rtol=1e-05, atol=1e-05)
def run_gradinput(self, inputs_shape, filters_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[4]).astype('float32'))
conv = theano.tensor.nnet.convTransp3D(W=filters,
b=bias,
d=subsample,
H=inputs)
f_ref = theano.function([], conv)
res_ref = f_ref()
# Get bottom shape using convTransp3D
bottom_shape = res_ref.shape
bottom_val = numpy.random.random(bottom_shape).astype('float32')
bottom = shared(bottom_val)
weight = gpu_contiguous(filters.dimshuffle(0, 4, 1, 2, 3))
top = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top)
else:
conv_gemm = GpuCorr3dMM_gradInputs(subsample=subsample)(
kern=weight, topgrad=top,
shape=bottom.shape[1:4])
conv_gemm = conv_gemm.dimshuffle(0, 2, 3, 4, 1)
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res = f()
utt.assert_allclose(res_ref, res)
def run_gradweight(self, inputs_shape, filters_shape, dCdH_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
dCdH_val = numpy.random.random(dCdH_shape).astype('float32')
inputs = shared(inputs_val)
dCdH = shared(dCdH_val)
conv = theano.tensor.nnet.convGrad3D(V=inputs, dCdH=dCdH,
WShape=filters_shape,
d=subsample)
img = gpu_contiguous(inputs.dimshuffle(0, 4, 1, 2, 3))
topgrad = gpu_contiguous(dCdH.dimshuffle(0, 4, 1, 2, 3))
if (subsample == (1, 1, 1)):
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(img,
topgrad)
else:
conv_gemm = GpuCorr3dMM_gradWeights(subsample=subsample)(
img, topgrad, shape=filters_shape[1:4])
conv_gemm = conv_gemm.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv)
f = theano.function([], conv_gemm, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def run_conv3d_fwd(inputs_shape, filters_shape, subsample,
border_mode, conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
# Scale down the input values to prevent very large absolute errors
# due to float rounding
inputs_val /= 10
filters_val /= 10
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Compile a theano function for the CuDNN implementation
conv = dnn.dnn_conv3d(img=inputs, kerns=filters,
border_mode=border_mode, subsample=subsample,
conv_mode=conv_mode)
f = theano.function([], conv, mode=mode_with_gpu)
# If conv_mode is 'conv' the reference implementation should use
# filters filpped according to the width, height and time axis
if conv_mode == 'conv':
flipped_filters = filters[:, :, ::-1, ::-1, ::-1]
else:
flipped_filters = filters
# If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2, 5)]
else:
if isinstance(border_mode, int):
pad_per_dim = [border_mode] * 3
else:
pad_per_dim = border_mode
pad_before_after = ([(0, 0), (0, 0)] +
[(p, p) for p in pad_per_dim])
padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
'constant')
padded_inputs = shared(padded_inputs_val)
# Compile a theano function for the reference implementation
conv_ref = theano.tensor.nnet.conv3D(
V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias, d=subsample)
f_ref = theano.function([], conv_ref.dimshuffle(0, 4, 1, 2, 3))
# Compare the results of the two implementations
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def run_conv3d_fwd(inputs_shape, filters_shape, subsample,
border_mode, conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
# Scale down the input values to prevent very large absolute errors
# due to float rounding
inputs_val /= 10
filters_val /= 10
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Compile a theano function for the CuDNN implementation
conv = dnn.dnn_conv3d(img=inputs, kerns=filters,
border_mode=border_mode, subsample=subsample,
conv_mode=conv_mode)
f = theano.function([], conv, mode=mode_with_gpu)
# If conv_mode is 'conv' the reference implementation should use
# filters filpped according to the width, height and time axis
if conv_mode == 'conv':
flipped_filters = filters[:, :, ::-1, ::-1, ::-1]
else:
flipped_filters = filters
# If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2, 5)]
else:
if isinstance(border_mode, int):
pad_per_dim = [border_mode] * 3
else:
pad_per_dim = border_mode
pad_before_after = ([(0, 0), (0, 0)] +
[(p, p) for p in pad_per_dim])
padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
'constant')
padded_inputs = shared(padded_inputs_val)
# Compile a theano function for the reference implementation
conv_ref = theano.tensor.nnet.conv3D(
V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias, d=subsample)
f_ref = theano.function([], conv_ref.dimshuffle(0, 4, 1, 2, 3))
# Compare the results of the two implementations
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def run_conv(self, inputs_shape, filters_shape, pad=False, **other_args):
inputs_val = numpy.random.random(inputs_shape).astype("float32")
filters_val = numpy.random.random(filters_shape).astype("float32")
inputs = shared(inputs_val)
filters = shared(filters_val)
conv_ref = theano.tensor.nnet.conv.conv2d(inputs, filters, **other_args)
conv_fft = theano.sandbox.cuda.fftconv.conv2d_fft(inputs, filters, pad_last_dim=pad, **other_args)
f_ref = theano.function([], conv_ref)
f_fft = theano.function([], conv_fft, mode=mode_with_gpu)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft)
def run_conv_valid(self,
inputs_shape,
filters_shape,
border_mode='valid',
filter_dilation=(1, 1, 1),
subsample=(1, 1, 1),
verify_grad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv_ref = Corr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample)(inputs.dimshuffle(
0, 4, 1, 2,
3), filters.dimshuffle(0, 4, 1, 2, 3))
conv_ref = conv_ref.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref, mode='FAST_RUN')
conv = GpuCorr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample)(inputs.dimshuffle(
0, 4, 1, 2,
3), filters.dimshuffle(0, 4, 1, 2, 3))
conv = conv.dimshuffle(0, 2, 3, 4, 1)
f = theano.function([], conv, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
if verify_grad:
utt.verify_grad(GpuCorr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample),
[
inputs_val.transpose(0, 4, 1, 2, 3),
filters_val.transpose(0, 4, 1, 2, 3)
],
mode=mode_with_gpu)
def run_conv(self, inputs_shape, filters_shape, pad=False, **other_args):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv_ref = theano.tensor.nnet.conv.conv2d(inputs, filters,
**other_args)
conv_fft = theano.sandbox.cuda.fftconv.conv2d_fft(inputs, filters,
pad_last_dim=pad,
**other_args)
f_ref = theano.function([], conv_ref)
f_fft = theano.function([], conv_fft, mode=mode_with_gpu)
res_ref = f_ref()
res_fft = f_fft()
utt.assert_allclose(res_ref, res_fft)
def test_opt_nofft_valid(self):
inputs_shape = (5, 3, 7, 6)
filters_shape = (2, 3, 3, 3)
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv = theano.tensor.nnet.conv.conv2d(inputs, filters,
version='no_fft')
mode = mode_with_gpu.including('conv_fft_valid')
f_fft = theano.function([], conv, mode=mode)
# make sure we that no CuFFTOp has been inserted
topo = f_fft.maker.fgraph.toposort()
assert sum(isinstance(n.op, theano.sandbox.cuda.fftconv.CuFFTOp)
for n in topo) == 0
def run_conv_valid(self, inputs_shape, filters_shape,
subsample=(1, 1, 1)):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
conv_ref = theano.tensor.nnet.conv3D(V=inputs, W=filters,
b=bias, d=subsample)
conv = GpuCorr3dMM(border_mode = "valid",
subsample=subsample)(inputs.dimshuffle(0, 4, 1, 2, 3),
filters.dimshuffle(0, 4, 1, 2, 3))
conv = conv.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref)
f = theano.function([], conv, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
def run_conv_valid(self, inputs_shape, filters_shape,
border_mode='valid',
filter_dilation=(1, 1, 1),
subsample=(1, 1, 1),
verify_grad=False):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
conv_ref = Corr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample)(
inputs.dimshuffle(0, 4, 1, 2, 3),
filters.dimshuffle(0, 4, 1, 2, 3))
conv_ref = conv_ref.dimshuffle(0, 2, 3, 4, 1)
f_ref = theano.function([], conv_ref, mode='FAST_RUN')
conv = GpuCorr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample)(
inputs.dimshuffle(0, 4, 1, 2, 3),
filters.dimshuffle(0, 4, 1, 2, 3))
conv = conv.dimshuffle(0, 2, 3, 4, 1)
f = theano.function([], conv, mode=mode_with_gpu)
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref, res)
if verify_grad:
utt.verify_grad(GpuCorr3dMM(border_mode=border_mode,
filter_dilation=filter_dilation,
subsample=subsample),
[inputs_val.transpose(0, 4, 1, 2, 3),
filters_val.transpose(0, 4, 1, 2, 3)],
mode=mode_with_gpu)
def run_conv3d_bwd(inputs_shape, filters_shape, subsample, border_mode,
conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Compile a theano function for the CuDNN implementation
conv = dnn.dnn_conv3d(img=inputs,
kerns=filters,
border_mode=border_mode,
subsample=subsample,
conv_mode=conv_mode)
grad_i, grad_w = theano.tensor.grad(conv.sum(), [inputs, filters])
f = theano.function([], [grad_i, grad_w], mode=mode_with_gpu)
# If conv_mode is 'conv' the reference implementation should use
# filters filpped according to the width, height and time axis
if conv_mode == 'conv':
flipped_filters = filters[:, :, ::-1, ::-1, ::-1]
else:
flipped_filters = filters
# If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2, 5)]
else:
if isinstance(border_mode, int):
pad_per_dim = [border_mode] * 3
else:
pad_per_dim = border_mode
pad_before_after = ([(0, 0),
(0, 0)] + [(p, p) for p in pad_per_dim])
padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
'constant')
padded_inputs = shared(padded_inputs_val)
# Compile a theano function for the reference implementation
conv_ref = theano.tensor.nnet.conv3D(
V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias,
d=subsample)
(grad_padded_i_ref,
grad_w_ref) = theano.tensor.grad(conv_ref.sum(),
[padded_inputs, filters])
# Recover grad_i_ref from grad_padded_i_ref
if border_mode == 'valid':
grad_i_ref = grad_padded_i_ref
else:
shp = grad_padded_i_ref.shape
grad_i_ref = grad_padded_i_ref[:, :, pad_per_dim[0]:shp[2] -
pad_per_dim[0],
pad_per_dim[1]:shp[3] -
pad_per_dim[1],
pad_per_dim[2]:shp[4] -
pad_per_dim[2]]
f_ref = theano.function([], [grad_i_ref, grad_w_ref])
# Compare the results of the two implementations
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref[0], res[0])
utt.assert_allclose(res_ref[1], res[1])
def run_conv3d_bwd(inputs_shape, filters_shape, subsample,
border_mode, conv_mode):
inputs_val = numpy.random.random(inputs_shape).astype('float32')
filters_val = numpy.random.random(filters_shape).astype('float32')
inputs = shared(inputs_val)
filters = shared(filters_val)
bias = shared(numpy.zeros(filters_shape[0]).astype('float32'))
# Compile a theano function for the CuDNN implementation
conv = dnn.dnn_conv3d(img=inputs, kerns=filters,
border_mode=border_mode, subsample=subsample,
conv_mode=conv_mode)
grad_i, grad_w = theano.tensor.grad(conv.sum(), [inputs, filters])
f = theano.function([], [grad_i, grad_w], mode=mode_with_gpu)
# If conv_mode is 'conv' the reference implementation should use
# filters filpped according to the width, height and time axis
if conv_mode == 'conv':
flipped_filters = filters[:, :, ::-1, ::-1, ::-1]
else:
flipped_filters = filters
# If border mode is anything but 'valid', the reference implementation
# should operate on padded inputs
if border_mode == 'valid':
padded_inputs = inputs
else:
if border_mode == 'full':
pad_per_dim = [filters_shape[i] - 1 for i in range(2, 5)]
else:
if isinstance(border_mode, int):
pad_per_dim = [border_mode] * 3
else:
pad_per_dim = border_mode
pad_before_after = ([(0, 0), (0, 0)] +
[(p, p) for p in pad_per_dim])
padded_inputs_val = numpy.pad(inputs_val, pad_before_after,
'constant')
padded_inputs = shared(padded_inputs_val)
# Compile a theano function for the reference implementation
conv_ref = theano.tensor.nnet.conv3D(
V=padded_inputs.dimshuffle(0, 2, 3, 4, 1),
W=flipped_filters.dimshuffle(0, 2, 3, 4, 1),
b=bias, d=subsample)
(grad_padded_i_ref,
grad_w_ref) = theano.tensor.grad(conv_ref.sum(),
[padded_inputs, filters])
# Recover grad_i_ref from grad_padded_i_ref
if border_mode == 'valid':
grad_i_ref = grad_padded_i_ref
else:
shp = grad_padded_i_ref.shape
grad_i_ref = grad_padded_i_ref[
:, :,
pad_per_dim[0]:shp[2] - pad_per_dim[0],
pad_per_dim[1]:shp[3] - pad_per_dim[1],
pad_per_dim[2]:shp[4] - pad_per_dim[2]]
f_ref = theano.function([], [grad_i_ref, grad_w_ref])
# Compare the results of the two implementations
res_ref = f_ref()
res = f()
utt.assert_allclose(res_ref[0], res[0])
utt.assert_allclose(res_ref[1], res[1])