def test_convolution_with_non_zero_padding():
element_type = Type.f32
image_shape = Shape([1, 1, 10, 10])
filter_shape = Shape([1, 1, 3, 3])
data = Parameter(element_type, image_shape)
filters = Parameter(element_type, filter_shape)
parameter_list = [data, filters]
image_arr = np.arange(100, dtype=np.float32).reshape(1, 1, 10, 10)
filter_arr = (np.ones(9, dtype=np.float32).reshape(1, 1, 3, 3)) * -1
filter_arr[0][0][1][1] = 1
strides = [1, 1]
dilations = [2, 2]
pads_begin = [2, 1]
pads_end = [1, 2]
model = ov.convolution(data, filters, strides, pads_begin, pads_end,
dilations)
function = Model([model], parameter_list, "test")
runtime = get_runtime()
computation = runtime.computation(function, *parameter_list)
result = computation(image_arr, filter_arr)[0]
expected = convolution2d(image_arr[0][0], filter_arr[0][0], strides,
dilations, pads_begin,
pads_end).reshape([1, 1, 9, 9])
assert np.allclose(result, expected)
def test_convolution_simple():
element_type = Type.f32
image_shape = Shape([1, 1, 16, 16])
filter_shape = Shape([1, 1, 3, 3])
data = Parameter(element_type, image_shape)
filters = Parameter(element_type, filter_shape)
parameter_list = [data, filters]
image_arr = np.arange(-128, 128, 1, dtype=np.float32).reshape(1, 1, 16, 16)
filter_arr = np.ones(9, dtype=np.float32).reshape(1, 1, 3, 3)
filter_arr[0][0][0][0] = -1
filter_arr[0][0][1][1] = -1
filter_arr[0][0][2][2] = -1
filter_arr[0][0][0][2] = -1
filter_arr[0][0][2][0] = -1
strides = [1, 1]
pads_begin = [0, 0]
pads_end = [0, 0]
dilations = [1, 1]
model = ov.convolution(data, filters, strides, pads_begin, pads_end,
dilations)
function = Model([model], parameter_list, "test")
runtime = get_runtime()
computation = runtime.computation(function, *parameter_list)
result = computation(image_arr, filter_arr)[0]
expected = convolution2d(image_arr[0][0],
filter_arr[0][0]).reshape(1, 1, 14, 14)
assert np.allclose(result, expected)
def create_ngraph_function(model_path: str) -> Model:
"""Create a model on the fly from the source code using ngraph"""
def shape_and_length(shape: list) -> typing.Tuple[list, int]:
length = reduce(lambda x, y: x * y, shape)
return shape, length
weights = np.fromfile(model_path, dtype=np.float32)
weights_offset = 0
padding_begin = padding_end = [0, 0]
# input
input_shape = [64, 1, 28, 28]
param_node = op.Parameter(Type.f32, Shape(input_shape))
# convolution 1
conv_1_kernel_shape, conv_1_kernel_length = shape_and_length([20, 1, 5, 5])
conv_1_kernel = op.Constant(Type.f32, Shape(conv_1_kernel_shape),
weights[0:conv_1_kernel_length].tolist())
weights_offset += conv_1_kernel_length
conv_1_node = opset8.convolution(param_node, conv_1_kernel, [1, 1],
padding_begin, padding_end, [1, 1])
# add 1
add_1_kernel_shape, add_1_kernel_length = shape_and_length([1, 20, 1, 1])
add_1_kernel = op.Constant(
Type.f32, Shape(add_1_kernel_shape),
weights[weights_offset:weights_offset + add_1_kernel_length])
weights_offset += add_1_kernel_length
add_1_node = opset8.add(conv_1_node, add_1_kernel)
# maxpool 1
maxpool_1_node = opset1.max_pool(add_1_node, [2, 2], padding_begin,
padding_end, [2, 2], 'ceil')
# convolution 2
conv_2_kernel_shape, conv_2_kernel_length = shape_and_length(
[50, 20, 5, 5])
conv_2_kernel = op.Constant(
Type.f32,
Shape(conv_2_kernel_shape),
weights[weights_offset:weights_offset + conv_2_kernel_length],
)
weights_offset += conv_2_kernel_length
conv_2_node = opset8.convolution(maxpool_1_node, conv_2_kernel, [1, 1],
padding_begin, padding_end, [1, 1])
# add 2
add_2_kernel_shape, add_2_kernel_length = shape_and_length([1, 50, 1, 1])
add_2_kernel = op.Constant(
Type.f32,
Shape(add_2_kernel_shape),
weights[weights_offset:weights_offset + add_2_kernel_length],
)
weights_offset += add_2_kernel_length
add_2_node = opset8.add(conv_2_node, add_2_kernel)
# maxpool 2
maxpool_2_node = opset1.max_pool(add_2_node, [2, 2], padding_begin,
padding_end, [2, 2], 'ceil')
# reshape 1
reshape_1_dims, reshape_1_length = shape_and_length([2])
# workaround to get int64 weights from float32 ndarray w/o unnecessary copying
dtype_weights = np.frombuffer(
weights[weights_offset:weights_offset + 2 * reshape_1_length],
dtype=np.int64,
)
reshape_1_kernel = op.Constant(Type.i64, Shape(list(dtype_weights.shape)),
dtype_weights)
weights_offset += 2 * reshape_1_length
reshape_1_node = opset8.reshape(maxpool_2_node, reshape_1_kernel, True)
# matmul 1
matmul_1_kernel_shape, matmul_1_kernel_length = shape_and_length(
[500, 800])
matmul_1_kernel = op.Constant(
Type.f32,
Shape(matmul_1_kernel_shape),
weights[weights_offset:weights_offset + matmul_1_kernel_length],
)
weights_offset += matmul_1_kernel_length
matmul_1_node = opset8.matmul(reshape_1_node, matmul_1_kernel, False, True)
# add 3
add_3_kernel_shape, add_3_kernel_length = shape_and_length([1, 500])
add_3_kernel = op.Constant(
Type.f32,
Shape(add_3_kernel_shape),
weights[weights_offset:weights_offset + add_3_kernel_length],
)
weights_offset += add_3_kernel_length
add_3_node = opset8.add(matmul_1_node, add_3_kernel)
# ReLU
relu_node = opset8.relu(add_3_node)
# reshape 2
reshape_2_kernel = op.Constant(Type.i64, Shape(list(dtype_weights.shape)),
dtype_weights)
reshape_2_node = opset8.reshape(relu_node, reshape_2_kernel, True)
# matmul 2
matmul_2_kernel_shape, matmul_2_kernel_length = shape_and_length([10, 500])
matmul_2_kernel = op.Constant(
Type.f32,
Shape(matmul_2_kernel_shape),
weights[weights_offset:weights_offset + matmul_2_kernel_length],
)
weights_offset += matmul_2_kernel_length
matmul_2_node = opset8.matmul(reshape_2_node, matmul_2_kernel, False, True)
# add 4
add_4_kernel_shape, add_4_kernel_length = shape_and_length([1, 10])
add_4_kernel = op.Constant(
Type.f32,
Shape(add_4_kernel_shape),
weights[weights_offset:weights_offset + add_4_kernel_length],
)
weights_offset += add_4_kernel_length
add_4_node = opset8.add(matmul_2_node, add_4_kernel)
# softmax
softmax_axis = 1
softmax_node = opset8.softmax(add_4_node, softmax_axis)
return Model(softmax_node, [param_node], 'lenet')
