Python filter_acts.FilterActs類代碼示例

    def lmul(self, x):
        """
        dot(x, A)
        aka, do convolution with input image x

        """

        check_cuda(str(type(self)) + ".lmul")
        # TODO Why is it CPU??
        print "Por que?!?!", type(x)
        cpu = "Cuda" not in str(type(x))
        if cpu:
            x = gpu_from_host(x)

        assert x.ndim == 5
        x_axes = self.input_axes
        assert len(x_axes) == 5

        op_axes = ("c", 0, 1, "t", "b")
        if tuple(x_axes) != op_axes:
            print "ssssssssssssssss"
            x = x.dimshuffle(*[x_axes.index(axis) for axis in op_axes])

        _x_4d_shape = (
            self.signal_shape[0],
            self.signal_shape[1],
            self.signal_shape[2],
            self.signal_shape[3] * self.signal_shape[4],
        )

        x = x.reshape(_x_4d_shape)

        x = gpu_contiguous(x)

        rval = FilterActs(self.pad, self.partial_sum, self.kernel_stride[0])(x, self._filters)

        if cpu:
            rval = host_from_gpu(rval)

        rval = rval.reshape(
            (
                self.filter_shape[3],
                self.filter_shape[4],
                rval.shape[1],
                rval.shape[2],
                self.signal_shape[3],
                self.signal_shape[4],
            )
        )

        rval = diagonal_subtensor(rval, 4, 0).sum(axis=0)

        # Format the output based on the output space
        rval_axes = self.output_axes
        assert len(rval_axes) == 5

        if tuple(rval_axes) != op_axes:
            rval = rval.dimshuffle(*[op_axes.index(axis) for axis in rval_axes])

        return rval

def make_funcs(batch_size, rows, cols, channels, filter_rows, num_filters):
    rng = np.random.RandomState([2012, 10, 9])

    filter_cols = filter_rows

    base_image_value = rng.uniform(-1.0, 1.0, (channels, rows, cols, batch_size)).astype("float32")
    base_filters_value = rng.uniform(-1.0, 1.0, (channels, filter_rows, filter_cols, num_filters)).astype("float32")
    images = shared(base_image_value)
    filters = shared(base_filters_value, name="filters")

    # bench.py should always be run in gpu mode so we should not need a gpu_from_host here
    output = FilterActs()(images, filters)

    output_shared = shared(output.eval())

    cuda_convnet = function([], updates={output_shared: output})
    cuda_convnet.name = "cuda_convnet"

    images_bc01v = base_image_value.transpose(3, 0, 1, 2)
    filters_bc01v = base_filters_value.transpose(3, 0, 1, 2)
    filters_bc01v = filters_bc01v[:, :, ::-1, ::-1]

    images_bc01 = shared(images_bc01v)
    filters_bc01 = shared(filters_bc01v)

    output_conv2d = conv2d(
        images_bc01, filters_bc01, border_mode="valid", image_shape=images_bc01v.shape, filter_shape=filters_bc01v.shape
    )

    output_conv2d_shared = shared(output_conv2d.eval())

    baseline = function([], updates={output_conv2d_shared: output_conv2d})
    baseline.name = "baseline"

    return cuda_convnet, baseline

def test_match_valid_conv_strided():

    # Tests that running FilterActs with stride is the same as running
    # theano's conv2D in valid mode and then downsampling

    rng = np.random.RandomState([2012,10,9])

    batch_size = 5
    rows = 9
    cols = 9
    channels = 3
    filter_rows = 3
    filter_cols = filter_rows
    stride = 3
    num_filters = 16

    images = shared(rng.uniform(-1., 1., (channels, rows, cols,
        batch_size)).astype('float32'), name='images')
    filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
        filter_cols, num_filters)).astype('float32'), name='filters')

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs(stride=stride)(gpu_images, gpu_filters)
    output = host_from_gpu(output)

    images_bc01 = images.dimshuffle(3,0,1,2)
    filters_bc01 = filters.dimshuffle(3,0,1,2)
    filters_bc01 = filters_bc01[:,:,::-1,::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01,
            border_mode='valid', subsample=(stride, stride))

    output_conv2d_orig = output_conv2d.dimshuffle(1,2,3,0)
    output_conv2d = output_conv2d_orig  # [:, ::stride, ::stride, :]
    f = function([], [output, output_conv2d, output_conv2d_orig])

    output, output_conv2d, output_conv2d_orig = f()

    warnings.warn("""test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?
                     One possibility is that theano is numerically unstable and Alex's code is better.
                     Probably theano CPU 64 bit is OK but it's worth checking the others.""")
    if np.abs(output - output_conv2d).max() > 2.4e-6:
        assert type(output) == type(output_conv2d)
        assert output.dtype == output_conv2d.dtype
        if output.shape != output_conv2d.shape:
            print 'cuda-convnet shape: ',output.shape
            print 'theano shape: ',output_conv2d.shape
            assert False
        err = np.abs(output - output_conv2d)
        print 'absolute error range: ', (err.min(), err.max())
        print 'mean absolute error: ', err.mean()
        print 'cuda-convnet value range: ', (output.min(), output.max())
        print 'theano value range: ', (output_conv2d.min(), output_conv2d.max())
        assert False

def test_grad():

    rng = np.random.RandomState([2012, 10, 9])

    batch_size = 5
    rows = 10
    cols = 9
    channels = 3
    filter_rows = 4
    filter_cols = filter_rows
    num_filters = 16

    images = shared(rng.uniform(-1.0, 1.0, (channels, rows, cols, batch_size)).astype("float32"), name="images")
    filters = shared(
        rng.uniform(-1.0, 1.0, (channels, filter_rows, filter_cols, num_filters)).astype("float32"), name="filters"
    )

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs()(gpu_images, gpu_filters)
    output = host_from_gpu(output)
    # XXX: use verify_grad
    output_grad = grad(output.sum(), images)

    images_bc01 = images.dimshuffle(3, 0, 1, 2)
    filters_bc01 = filters.dimshuffle(3, 0, 1, 2)
    filters_bc01 = filters_bc01[:, :, ::-1, ::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01, border_mode="valid")

    output_conv2d = output_conv2d.dimshuffle(1, 2, 3, 0)
    # XXX: use verify_grad
    output_conv2d_grad = grad(output_conv2d.sum(), images)
    f = function([], [output_grad, output_conv2d_grad])

    output_grad, output_conv2d_grad = f()

    warnings.warn(
        """test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?
                     One possibility is that theano is numerically unstable and Alex's code is better.
                     Probably theano CPU 64 bit is OK but it's worth checking the others."""
    )
    if np.abs(output_grad - output_conv2d_grad).max() > 7.7e-6:
        assert type(output_grad) == type(output_conv2d_grad)
        assert output_grad.dtype == output_conv2d_grad.dtype
        if output_grad.shape != output_conv2d_grad.shape:
            print "cuda-convnet shape: ", output_grad.shape
            print "theano shape: ", output_conv2d_grad.shape
            assert False
        err = np.abs(output_grad - output_conv2d_grad)
        print "absolute error range: ", (err.min(), err.max())
        print "mean absolute error: ", err.mean()
        print "cuda-convnet value range: ", (output_grad.min(), output_grad.max())
        print "theano value range: ", (output_conv2d_grad.min(), output_conv2d_grad.max())
        assert False

def test_match_valid_conv():

    # Tests that running FilterActs with no padding is the same as running
    # theano's conv2D in valid mode

    rng = np.random.RandomState([2012,10,9])

    batch_size = 5
    rows = 10
    cols = 9
    channels = 3
    filter_rows = 4
    filter_cols = filter_rows
    num_filters = 16

    images = shared(rng.uniform(-1., 1., (channels, rows, cols,
        batch_size)).astype('float32'), name='images')
    filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
        filter_cols, num_filters)).astype('float32'), name='filters')

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    output = FilterActs()(gpu_images, gpu_filters)
    output = host_from_gpu(output)

    images_bc01 = images.dimshuffle(3,0,1,2)
    filters_bc01 = filters.dimshuffle(3,0,1,2)
    filters_bc01 = filters_bc01[:,:,::-1,::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01,
            border_mode='valid')

    output_conv2d = output_conv2d.dimshuffle(1,2,3,0)

    try:
        f = function([], [output, output_conv2d])
    except:
        raise KnownFailureTest("cuda-convnet code depends on an unmerged theano feature.")

    output, output_conv2d = f()

    warnings.warn("test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?")
    if np.abs(output - output_conv2d).max() > 2.4e-6:
        assert type(output) == type(output_conv2d)
        assert output.dtype == output_conv2d.dtype
        if output.shape != output_conv2d.shape:
            print 'cuda-convnet shape: ',output.shape
            print 'theano shape: ',output_conv2d.shape
            assert False
        err = np.abs(output - output_conv2d)
        print 'absolute error range: ', (err.min(), err.max())
        print 'mean absolute error: ', err.mean()
        print 'cuda-convnet value range: ', (output.min(), output.max())
        print 'theano value range: ', (output_conv2d.min(), output_conv2d.max())
        assert False

    def lmul(self, x):
        """
        .. todo::

            WRITEME properly

        dot(x, A)
        aka, do convolution with input image x
        """

        check_cuda(str(type(self)) + ".lmul")

        cpu = 'Cuda' not in str(type(x))

        if cpu:
            x = gpu_from_host(x)

        # x must be formatted as channel, topo dim 0, topo dim 1, batch_index
        # for use with FilterActs
        assert x.ndim == 4
        x_axes = self.input_axes
        assert len(x_axes) == 4

        op_axes = ('c', 0, 1, 'b')

        if tuple(x_axes) != op_axes:
            x = x.dimshuffle(*[x_axes.index(axis) for axis in op_axes])

        x = gpu_contiguous(x)

        # Patch old pickle files.
        if not hasattr(self, 'kernel_stride'):
            self.kernel_stride = (1, 1)
        rval = FilterActs(self.pad, self.partial_sum, self.kernel_stride[0])(
            x,
            self._filters
        )

        # Format the output based on the output space
        rval_axes = self.output_axes
        assert len(rval_axes) == 4

        if cpu:
            rval = host_from_gpu(rval)

        if tuple(rval_axes) != op_axes:
            rval = rval.dimshuffle(*[op_axes.index(axis)
                                     for axis in rval_axes])

        return rval

    def lmul(self, x):
        """
        dot(x, A)
        aka, do convolution with input image x

        """

        cpu = 'Cuda' not in str(type(x))

        if cpu:
            x = gpu_from_host(x)

        # x must be formatted as channel, topo dim 0, topo dim 1, batch_index
        # for use with FilterActs
        assert x.ndim == 4
        x_axes = self.input_axes
        assert len(x_axes) == 4

        op_axes = ('c', 0, 1, 'b')

        if tuple(x_axes) != op_axes:
            x = x.dimshuffle(*[x_axes.index(axis) for axis in op_axes])

        x = gpu_contiguous(x)

        rval = FilterActs(self.pad, self.partial_sum)(x, self._filters)

        # Format the output based on the output space
        rval_axes = self.output_axes
        assert len(rval_axes) == 4

        if tuple(rval_axes) != op_axes:
            rval = rval.dimshuffle(*[op_axes.index(axis) for axis in rval_axes])

        if cpu:
            rval = host_from_gpu(rval)

        return rval

def test_match_grad_valid_conv():

    # Tests that weightActs is the gradient of FilterActs
    # with respect to the weights.

    for partial_sum in [0, 1, 4]:
        rng = np.random.RandomState([2012, 10, 9])

        batch_size = 3
        rows = 7
        cols = 9
        channels = 8
        filter_rows = 4
        filter_cols = filter_rows
        num_filters = 16

        images = shared(rng.uniform(-1., 1., (channels, rows, cols,
                                              batch_size)).astype('float32'),
                        name='images')
        filters = rng.uniform(-1., 1.,
                              (channels, filter_rows,
                               filter_cols, num_filters)).astype('float32')
        filters = shared(filters, name='filters')

        gpu_images = gpu_from_host(images)
        gpu_filters = gpu_from_host(filters)

        output = FilterActs(partial_sum=partial_sum)(gpu_images, gpu_filters)
        output = host_from_gpu(output)

        images_bc01 = images.dimshuffle(3, 0, 1, 2)
        filters_bc01 = filters.dimshuffle(3, 0, 1, 2)
        filters_bc01 = filters_bc01[:, :, ::-1, ::-1]

        output_conv2d = conv2d(images_bc01, filters_bc01,
                               border_mode='valid')

        output_conv2d = output_conv2d.dimshuffle(1, 2, 3, 0)

        theano_rng = MRG_RandomStreams(2013 + 1 + 31)

        coeffs = theano_rng.normal(avg=0., std=1.,
                                   size=output_conv2d.shape, dtype='float32')

        cost_conv2d = (coeffs * output_conv2d).sum()

        weights_grad_conv2d = T.grad(cost_conv2d, filters)

        cost = (coeffs * output).sum()
        hid_acts_grad = T.grad(cost, output)

        weights_grad = WeightActs(partial_sum=partial_sum)(
            gpu_images,
            gpu_from_host(hid_acts_grad),
            as_tensor_variable((4, 4))
        )[0]
        weights_grad = host_from_gpu(weights_grad)

        f = function([], [output, output_conv2d, weights_grad,
                          weights_grad_conv2d])

        output, output_conv2d, weights_grad, weights_grad_conv2d = f()

        if np.abs(output - output_conv2d).max() > 8e-6:
            assert type(output) == type(output_conv2d)
            assert output.dtype == output_conv2d.dtype
            if output.shape != output_conv2d.shape:
                print('cuda-convnet shape: ', output.shape)
                print('theano shape: ', output_conv2d.shape)
                assert False
            err = np.abs(output - output_conv2d)
            print('absolute error range: ', (err.min(), err.max()))
            print('mean absolute error: ', err.mean())
            print('cuda-convnet value range: ', (output.min(), output.max()))
            print('theano value range: ', (output_conv2d.min(),
                                           output_conv2d.max()))
            assert False

        warnings.warn(
            "test_match_grad_valid_conv success criterion is not very strict."
            " Can we verify that this is OK? One possibility is that theano"
            " is numerically unstable and Alex's code is better. Probably"
            " theano CPU 64 bit is OK but it's worth checking the others.")

        if np.abs(weights_grad - weights_grad_conv2d).max() > 8.6e-6:
            if type(weights_grad) != type(weights_grad_conv2d):
                raise AssertionError("weights_grad is of type " +
                                     str(weights_grad))
            assert weights_grad.dtype == weights_grad_conv2d.dtype
            if weights_grad.shape != weights_grad_conv2d.shape:
                print('cuda-convnet shape: ', weights_grad.shape)
                print('theano shape: ', weights_grad_conv2d.shape)
                assert False
            err = np.abs(weights_grad - weights_grad_conv2d)
            print('absolute error range: ', (err.min(), err.max()))
            print('mean absolute error: ', err.mean())
            print('cuda-convnet value range: ', (weights_grad.min(),
                                                 weights_grad.max()))
            print('theano value range: ', (weights_grad_conv2d.min(),
                                           weights_grad_conv2d.max()))
#.........這裏部分代碼省略.........

def test_match_valid_conv_padded():

    # Tests that running FilterActs with no padding is the same as running
    # theano's conv2D in valid mode

    rng = np.random.RandomState([2012,10,9])

    batch_size = 5
    rows = 10
    cols = 9
    channels = 3
    filter_rows = 4
    filter_cols = filter_rows
    num_filters = 16

    images = shared(rng.uniform(-1., 1., (channels, rows, cols,
        batch_size)).astype('float32'), name='images')
    filters = shared(rng.uniform(-1., 1., (channels, filter_rows,
        filter_cols, num_filters)).astype('float32'), name='filters')

    gpu_images = gpu_from_host(images)
    gpu_filters = gpu_from_host(filters)

    PAD = 3

    output = FilterActs(PAD)(gpu_images, gpu_filters)
    output = host_from_gpu(output)

    images_bc01 = T.alloc(0., batch_size, channels, rows + PAD * 2, cols + PAD * 2)

    images_bc01 = T.set_subtensor(images_bc01[:,:,PAD:-PAD,PAD:-PAD], images.dimshuffle(3,0,1,2))


    filters_bc01 = filters.dimshuffle(3,0,1,2)
    filters_bc01 = filters_bc01[:,:,::-1,::-1]

    output_conv2d = conv2d(images_bc01, filters_bc01,
            border_mode='valid')

    output_conv2d = output_conv2d.dimshuffle(1,2,3,0)

    f = function([], [output, output_conv2d])

    output, output_conv2d = f()

    warnings.warn("""test_match_valid_conv success criterion is not very strict. Can we verify that this is OK?
                     One possibility is that theano is numerically unstable and Alex's code is better.
                     Probably theano CPU 64 bit is OK but it's worth checking the others.""")

    assert output.shape == output_conv2d.shape

    if np.abs(output - output_conv2d).max() > 2.4e-6:
        assert type(output) == type(output_conv2d)
        assert output.dtype == output_conv2d.dtype
        if output.shape != output_conv2d.shape:
            print('cuda-convnet shape: ',output.shape)
            print('theano shape: ',output_conv2d.shape)
            assert False
        err = np.abs(output - output_conv2d)
        print('absolute error range: ', (err.min(), err.max()))
        print('mean absolute error: ', err.mean())
        print('cuda-convnet value range: ', (output.min(), output.max()))
        print('theano value range: ', (output_conv2d.min(), output_conv2d.max()))
        assert False