def two_conv_graph():
G = NNGraph(name='two_conv_graph')
ti = G.add_input(Dim.unnamed([10, 10, 2]))
c1filt = Conv2DFilterDim(3, 3, 2, in_c=2)
c1filt.impose_order(['out_c', 'h', 'w', 'in_c'])
n1 = Conv2DParameters("node1",
filt=c1filt,
stride=StrideDim(1, 1),
padding=PadDim(0),
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]))
G.add_node(n1)
w1 = [[0.25, 0.25], [0.25, 0.25], [0.25, 0.25]]
w1 = [w1, w1, w1]
w2 = [[0.75, 0.75], [0.75, 0.75], [0.75, 0.75]]
w2 = [w2, w2, w2]
n1.weights = np.array([w1, w2])
c2filt = Conv2DFilterDim(3, 3, 2, in_c=2)
c2filt.impose_order(['out_c', 'h', 'w', 'in_c'])
n2 = Conv2DParameters("node2",
filt=c2filt,
stride=StrideDim(1, 1),
padding=PadDim(0),
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]))
G.add_node(n2)
w3 = [[0.75, 0.25], [0.75, 0.25], [0.75, 0.25]]
w3 = [w3, w3, w3]
n2.weights = np.array([w3, w3])
to = G.add_output()
G.add_edge(NNEdge(ti, n1))
G.add_edge(NNEdge(n1, n2))
G.add_edge(NNEdge(n2, to))
G.add_dimensions()
yield G
def add_pool(G,
tensors,
name,
subgraph,
op_name,
op,
load_tensors=False,
dequantize=False):
pool_opts = Pool2DOptions.Pool2DOptions()
pool_opts.Init(op.BuiltinOptions().Bytes, op.BuiltinOptions().Pos)
pad = get_tf_padding(pool_opts.Padding())
pool_type = TF_POOL_OPS[op_name]
inp = get_input_size(None, subgraph, op, 0, order=TF_LITE_IN_OUT_ORDER)
check(inp['n'] == 1, "Multi batch not supported")
node = PoolingParameters(name,
filt=PoolFilterDim(pool_opts.FilterHeight(),
pool_opts.FilterWidth()),
stride=StrideDim(pool_opts.StrideH(),
pool_opts.StrideW()),
padding=pad,
pool_type=pool_type,
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]))
return fuse_activation(G, pool_opts, name, node)
def test_conf2d_q2(caplog):
caplog.set_level(logging.INFO)
weights_q = QType(16, 1, True)
weights = weights_q.quantize(np.full([1, 1, 2, 2], 1.0))
filt = Conv2DFilterDim(2, 2, 1, 1)
stride = StrideDim(1)
pad = PadDim.valid()
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
in_q = QType(16, 0, True)
calc_q = QType(weights_q.bits + in_q.bits, weights_q.q + in_q.q, True)
qrec = FilterQuantizationRecord(in_qs=[in_q],
out_qs=[in_q],
weights_q=weights_q,
acc_q=calc_q,
calc_q=calc_q)
input_ = in_q.quantize(np.full([1, 2, 2], 1.0))
in_dims = Dim.named(c=1, h=2, w=2).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = conv2d(params,
in_dims,
out_dims[0],
input_,
weights,
None,
qrec=qrec)
output_ = in_q.dequantize(output_)
assert np.array_equal(output_, [[[4.]]])
def test_conf2d_normal():
weights = np.arange(9).reshape([1, 1, 3, 3])
filt = Conv2DFilterDim(3, 3, 1, 1)
stride = StrideDim(1)
pad = PadDim(0)
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
input_ = np.arange(16).reshape([1, 4, 4])
in_dims = Dim.named(c=1, h=4, w=4).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
details = {}
output_ = conv2d(params,
in_dims,
out_dims[0],
input_,
weights,
None,
details=details)
# assert details['max_acc'] == 438.0 and details['min_acc'] == 258.0
assert np.array_equal(output_, [[[258, 294], [402, 438]]])
def add_mean(G,
tensors,
name,
subgraph,
op_name,
op,
load_tensors=False,
dequantize=False):
check(op.InputsLength() == 2,\
"Very odd " + str(op.InputsAsNumpy()))
mean_dims = get_tensor(G.model, tensors, subgraph, op, 1, dequantize=False)
if len(mean_dims) != 2 or mean_dims[0] != 1 or mean_dims[1] != 2:
LOG.warning(
"MEAN operator seen but can't convert to global average pool")
return add_unconverted(G, name, subgraph, op_name, op, load_tensors,
dequantize)
else:
LOG.info("MEAN operator converted to global average pool")
inp = get_input_size(None, subgraph, op, 0, order=TF_LITE_IN_OUT_ORDER)
check(inp['n'] == 1, "Multi batch not supported")
return add_node(
G,
PoolingParameters(name,
filt=PoolFilterDim(inp['h'], inp['w']),
stride=StrideDim(1, 1),
padding=PadDim.valid(),
pool_type="average",
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']])))
def test_conf2d_depth():
# TF Lite depthwise convolution
weights = np.arange(9).reshape([3, 3])
weights = np.repeat(weights, 2).reshape([1, 3, 3, 2])
filt = Conv2DFilterDim(3, 3, 2,
1).impose_order(["in_c", "h", "w", "out_c"])
stride = StrideDim(1)
pad = PadDim(0)
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
groups=1,
multiplier=2,
tf_depthwise=True,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
input_ = np.arange(16).reshape([1, 4, 4])
in_dims = Dim.named(c=1, h=4, w=4).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output1 = conv2d(params, in_dims, out_dims[0], input_, weights, None)
assert np.array_equal(output1,
[[[258, 294], [402, 438]], [[258, 294], [402, 438]]])
output2 = conv2d(params,
in_dims,
out_dims[0],
input_,
weights,
None,
allow_faster=False)
assert np.array_equal(output1, output2)
def actfusion_graph():
G = NNGraph(name='actfusion_graph')
ti1 = G.add_input(Dim.unnamed([10, 10, 2])).name
ti2 = G.add_input(Dim.unnamed([10, 10, 2])).name
c1filt = Conv2DFilterDim(3, 3, 2, in_c=2)
c1filt.impose_order(['out_c', 'h', 'w', 'in_c'])
n1 = Conv2DParameters("node1",
filt=c1filt,
stride=StrideDim(1, 1),
padding=PadDim(0),
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]))
G.add_node(n1)
w1 = [[0.25, 0.25], [0.25, 0.25], [0.25, 0.25]]
w1 = [w1, w1, w1]
w2 = [[0.75, 0.75], [0.75, 0.75], [0.75, 0.75]]
w2 = [w2, w2, w2]
n1.weights = np.array([w1, w2])
n1a = ReluActivationParameters("node1a")
G.add_node(n1a)
c2filt = Conv2DFilterDim(3, 3, 2, in_c=2)
c2filt.impose_order(['out_c', 'h', 'w', 'in_c'])
n2 = Conv2DParameters("node2",
filt=c2filt,
stride=StrideDim(1, 1),
padding=PadDim(0),
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]))
G.add_node(n2)
w3 = [[0.75, 0.25], [0.75, 0.25], [0.75, 0.25]]
w3 = [w3, w3, w3]
n2.weights = np.array([w3, w3])
n3 = MatrixAddParameters("node3")
G.add_node(n3)
n4 = ReluActivationParameters("node4")
G.add_node(n4)
to = G.add_output()
G.add_edge(NNEdge(ti1, n1))
G.add_edge(NNEdge(n1, n1a))
G.add_edge(NNEdge(ti2, n2))
G.add_edge(NNEdge(n1a, n3, to_idx=0))
G.add_edge(NNEdge(n2, n3, to_idx=1))
G.add_edge(NNEdge(n3, n4))
G.add_edge(NNEdge(n4, to))
G.add_dimensions()
yield G
def version_1(cls, node: TFLiteNode, **kwargs):
node_opts = node.get_options(TransposeConvOptions)
G = kwargs['G']
opts = kwargs['opts']
all_nodes = kwargs['all_nodes']
inputs = [all_nodes[t] for t in node.input]
x = inputs[2]
x_shape = x[2].shape
in_b, in_h, in_w, in_c = tuple(x_shape)
pout_shape = [
dim if x_shape[idx] is not None else None
for idx, dim in enumerate(cls.get_constant(inputs[0]))
]
out_b, out_h, out_w, out_c = tuple(pout_shape)
filt = inputs[1]
weights_node = filt[0]
filt_shape = filt[2].shape
# # ['in_c', 'h', 'w', 'out_c']
filt_out_c, filt_h, filt_w, filt_in_c = tuple(filt_shape)
filt_dim = Conv2DFilterDim(filt_h, filt_w, filt_out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.TF_LITE_FILTER_ORDER)
stride_w = node_opts.StrideW()
stride_h = node_opts.StrideH()
# compute padding
pad = node_opts.Padding()
if pad == Padding.SAME:
pad_h = ((in_h - 1) * stride_h + filt_h - out_h)
pad_w = ((in_w - 1) * stride_w + filt_w - out_w)
pad_top = pad_h // 2
pad_left = pad_w // 2
pad = PadDim(pad_top,
pad_h - pad_top,
pad_left,
pad_w - pad_left,
same_type='balanced_right')
else:
pad = PadDim(0)
params = TransposeConv2DParameters(
node.name,
filt=filt_dim,
stride=StrideDim(stride_h, stride_w),
padding=pad,
in_dims_hint=[['h', 'w', 'c'],
cls.TF_LITE_FILTER_ORDER.copy()],
out_dims_hint=[['h', 'w', 'c']])
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
G.add_edge(NNEdge(from_node=weights_node, to_node=params, to_idx=1))
pout_dims = ProvisionalDim(pout_shape)
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def add_convolution(out_graph, routes, idx, l):
activation = get_str(l, 'activation', default="logistic")
node_name = "{}_{}".format(l['type'], idx)
routes['in'][idx] = node_name
padding = l.get("padding")
pad = l.get("pad")
size = get_int(l, 'size', 1)
groups = get_int(l, 'groups', 1)
filters_c = get_int(l, 'filters', 1)
stride = get_int(l, 'stride', 1)
batch_normalize = get_int(l, 'batch_normalize', 0)
flipped = get_int(l, 'flipped', 0)
custom = {'batch_normalize': batch_normalize == 1, 'flipped': flipped == 1}
assert 'binary' not in l, "Binary convolutions are not implemented"
assert 'xnor' not in l, "XNOR convolutions are not implemented"
assert 'dot' not in l, "dot is not implemented"
# padding calculation as per Darknet code
if pad is not None:
padding = int(size / 2)
if padding is None:
padding = 0
if activation is None:
routes['in'][idx], routes['out'][idx] =\
out_graph.add_operator(node_name,\
Conv2DParameters(Conv2DFilterDim(size, size, filters_c),\
StrideDim(stride), PadDim(padding), groups=groups,\
custom=custom, has_bias=True))
else:
activation = DARKNET_ACTIVATION_TYPES[activation]
routes['in'][idx], routes['out'][idx] =\
out_graph.add_operators(
node_name,
[
Conv2DParameters(Conv2DFilterDim(size, size, filters_c),\
StrideDim(stride), PadDim(padding), groups=groups,\
custom=custom, has_bias=True),
ActivationParameters(activation)
]
)
return True
def conv(cls, node, **kwargs):
all_nodes = kwargs['all_nodes']
G = kwargs['G']
valid_name = kwargs['valid_name']
inputs = [all_nodes[inp] for inp in node.input]
# input N x C x H x W
x = inputs[0]
x_rank = len(x[2].shape)
x_shape = x[2].shape
spatial_size = x_rank - 2
assert spatial_size <= 2, "only 1D and 2D convolutions supported"
# M x C/group x kH x kW
weights = cls.get_constant(inputs[1])
out_c = weights.shape[0]
group = node.attrs.get("group", 1)
in_c = x_shape[1]
filt_in_c = in_c // group
filt_h = weights.shape[2]
filt_w = weights.shape[2]
h = 1 if spatial_size <= 1 else x_shape[2]
w = 1 if spatial_size == 0 else (x_shape[2] if spatial_size == 1 else x_shape[3])
filt_dim = Conv2DFilterDim(filt_h, filt_w,
out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.ONNX_FILTER_ORDER)
if len(inputs) > 2:
biases = cls.get_constant(inputs[2])
else:
biases = np.zeros([out_c])
dilations = cls.pad_start_with(node.attrs.get("dilations", [1] * spatial_size), [1], 2)
strides = cls.pad_start_with(node.attrs.get("strides", [1] * spatial_size), [1], 2)
pad_dim = cls.calc_pad_dim(node, spatial_size)
params = Conv2DParameters(valid_name,
filt=filt_dim,
stride=StrideDim(strides[0],
strides[1]),
dilation=DilationDim(dilations[0],
dilations[1]),
groups=group,
padding=pad_dim,
has_bias=True,
in_dims_hint=SparseList([['c', 'h', 'w']]),
out_dims_hint=SparseList([['c', 'h', 'w']]),
constant_store=G.constant_store)
params.weights = weights
params.biases = biases
in_dim = Dim.named_ordered(c=in_c, h=h, w=w)
out_dims = params.get_output_size([in_dim])
pout_dims = ProvisionalDim([x_shape[0]] + out_dims[0].shape)
G.add_edge(NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def pool(cls, node, pool_type=None, **kwargs):
all_nodes = kwargs['all_nodes']
G = kwargs['G']
valid_name = kwargs['valid_name']
inputs = [all_nodes[inp] for inp in node.input]
x = inputs[0]
x_shape = x[2].shape
x_feature_shape = x_shape[2::]
in_c = x_shape[1]
kernel_shape = node.attrs["kernel_shape"]
spatial_size = len(kernel_shape)
x_rank = spatial_size + 2
if spatial_size != 2:
raise ValueError(valid_name + " with {}D input".format(x_rank))
h = x_shape[2]
w = x_shape[3]
strides = node.attrs.get("strides", [1] * spatial_size)
stride_is_one = all(stride == 1 for stride in strides)
dilations = node.attrs.get("dilations", [1] * spatial_size)
if any(dilation > 1 for dilation in dilations):
raise ValueError(valid_name + " with dilation not supported")
# ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pad_dim = cls.calc_pad_dim(node, spatial_size)
# Note: This needs to check dilation if it is added
filter_matches_input = (all(
k_dim >= (x_dim + pad) for k_dim, x_dim, pad in zip(
kernel_shape, x_feature_shape, [pad_dim.h, pad_dim.w])))
if filter_matches_input and stride_is_one:
params = GlobalPoolParameters(valid_name,
pool_type=pool_type,
axis=[1, 2],
keep_dims=True,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
else:
params = PoolingParameters(
valid_name,
filt=PoolFilterDim(kernel_shape[0], kernel_shape[1]),
stride=StrideDim(strides[0], strides[1]),
padding=pad_dim,
pool_type=pool_type,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
in_dim = Dim.named_ordered(c=in_c, h=h, w=w)
out_dims = params.get_output_size([in_dim])
pout_dims = ProvisionalDim([x_shape[0]] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def add_pool(out_graph, routes, idx, l):
pool_type = DARKNET_POOL_TYPES[l['type']]
node_name = "{}_{}".format(l['type'], idx)
if pool_type == "average":
# Darknet average pool averages entire channel
# Pool_w and _h set to None indicates that they must be computed
# when the network sizes are computed
routes['in'][idx], routes['out'][idx] =\
out_graph.add_operator(node_name,\
PoolingParameters(None, PoolFilterDim(1), StrideDim(0), pool_type=pool_type))
else:
stride = get_int(l, 'stride', default=1)
size = get_int(l, 'size', default=stride)
padding = get_int(l, 'padding', default=size - 1)
pad = split_pad(padding)
routes['in'][idx], routes['out'][idx] =\
out_graph.add_operator(node_name,\
PoolingParameters(PoolFilterDim(size), StrideDim(stride), pad, pool_type=pool_type))
return True
def test_paddim():
dim1 = PadDim(1)
assert not dim1.is_same
assert dim1.h == 2 and dim1.w == 2
assert dim1.l == 1 and dim1.r == 1 and dim1.t == 1 and dim1.b == 1
assert dim1.numpy_pad_shape(Dim.named_ordered(w=10, h=10)) == [(1, 1), (1, 1)]
stride_dim = StrideDim(1)
filt_dim = Conv2DFilterDim(5, 5, 1, 1)
in_dim = Dim.named_ordered(c=1, h=20, w=20)
dim1 = PadDim.same()
dim1.calculate_same(in_dim, filt_dim, stride_dim)
assert dim1.shape == [2, 2, 2, 2]
def pool2d(cls, node, pool_type=None, **kwargs):
all_nodes = kwargs['all_nodes']
G = kwargs['G']
opts = kwargs['opts']
node_opts = node.get_options(Pool2DOptions)
inputs = [all_nodes[inp] for inp in node.input]
x = inputs[0]
x = cls.remove_known_batch_dimension(G, x, node)
x_shape = x[2].shape
in_c = x_shape[1]
in_b, h, w, in_c = tuple(x_shape)
filt_h = node_opts.FilterHeight()
filt_w = node_opts.FilterWidth()
stride_h = node_opts.StrideH()
stride_w = node_opts.StrideW()
pad = cls.get_tf_padding(node_opts.Padding())
filter_matches_input = h == filt_h and w == filt_w
stride_is_one = stride_h == 1 and stride_w == 1
if filter_matches_input and stride_is_one:
params = GlobalPoolParameters(node.name,
pool_type=pool_type,
axis=[0, 1],
keep_dims=True,
in_dims_hint=[['h', 'w', 'c']],
out_dims_hint=[['h', 'w', 'c']])
else:
params = PoolingParameters(node.name,
filt=PoolFilterDim(filt_h, filt_w),
stride=StrideDim(stride_h, stride_w),
padding=pad,
pool_type=pool_type,
in_dims_hint=[['h', 'w', 'c']],
out_dims_hint=[['h', 'w', 'c']])
if opts.get('load_quantization'):
G.quantization[NodeId(params)] = cls.load_tf_quantization(
node.input, node.output)
in_dim = Dim.named_ordered(h=h, w=w, c=in_c)
out_dims = params.get_output_size([in_dim])
pout_dims = ProvisionalDim([in_b] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
params = cls.fuse_activation(node_opts, node.name, params, **kwargs)
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def test_max_pool_normal():
filt = PoolFilterDim(2, 2)
stride = StrideDim(1)
pad = PadDim(0)
params = PoolingParameters("test",
filt=filt,
stride=stride,
padding=pad,
pool_type="max")
input_ = np.arange(9).reshape([1, 3, 3])
in_dims = Dim.named(c=1, h=3, w=3).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = max_pool(params, in_dims, out_dims[0], input_)
assert np.array_equal(output_, [[[4, 5], [7, 8]]])
def add_convolution(G,
tensors,
name,
subgraph,
_,
op,
load_tensors=False,
dequantize=False):
conv_opts = Conv2DOptions.Conv2DOptions()
conv_opts.Init(op.BuiltinOptions().Bytes, op.BuiltinOptions().Pos)
# get filter dimensions
filt = get_input_size(tensors, subgraph, op, 1, order=TF_LITE_FILTER_ORDER)
filt = Conv2DFilterDim(filt['h'], filt['w'],\
filt['out_c'], in_c=filt['in_c'])
filt = filt.impose_order(TF_LITE_FILTER_ORDER)
# compute padding
pad = get_tf_padding(conv_opts.Padding())
# does it have biases
has_bias = op.InputsLength() > 2
node = Conv2DParameters(name,
filt=filt,
stride=StrideDim(conv_opts.StrideH(),
conv_opts.StrideW()),
padding=pad,
has_bias=has_bias,
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]),
constant_store=G.constant_store)
if load_tensors:
node.weights = get_tensor(G.model,
tensors,
subgraph,
op,
1,
dequantize=dequantize)
if has_bias:
node.biases = get_tensor(G.model,
tensors,
subgraph,
op,
2,
dequantize=dequantize)
return fuse_activation(G, conv_opts, name, node)
def test_conf2d_depth_q():
calc_q = QType(32, 9, True)
biases_q = acc_q = out_q = QType(16, 4, True)
weights_q = QType(16, 4, True)
in_q = QType(16, 5, True)
# TF Lite depthwise convolution
biases = np.full([2], 0.5)
qbiases = biases_q.quantize(biases)
weights = np.full([3, 3], 0.5)
weights = np.repeat(weights, 2).reshape([1, 3, 3, 2])
qweights = weights_q.quantize(weights)
filt = Conv2DFilterDim(3, 3, 2,
1).impose_order(["in_c", "h", "w", "out_c"])
stride = StrideDim(1)
pad = PadDim(0)
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
groups=1,
multiplier=2,
tf_depthwise=True,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
qrec = FilterQuantizationRecord(in_qs=[in_q],
out_qs=[out_q],
weights_q=weights_q,
biases_q=biases_q,
acc_q=acc_q,
calc_q=calc_q)
input_ = np.full([1, 4, 4], 2)
qinput_ = in_q.quantize(input_)
in_dims = Dim.named(c=1, h=4, w=4).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = conv2d(params, in_dims, out_dims[0], input_, weights, biases)
qoutput_ = conv2d(params,
in_dims,
out_dims[0],
qinput_,
qweights,
qbiases,
qrec=qrec)
dqoutput_ = out_q.dequantize(qoutput_)
assert np.array_equal(output_, dqoutput_)
def test_max_pool_q():
filt = PoolFilterDim(2, 2)
stride = StrideDim(1)
pad = PadDim(0)
params = PoolingParameters("test",
filt=filt,
stride=stride,
padding=pad,
pool_type="max")
in_q = QType(16, 0, True)
qrec = QuantizationRecord([in_q], [in_q])
input_ = in_q.quantize(np.arange(9)).reshape([1, 3, 3])
in_dims = Dim.named(c=1, h=3, w=3).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = max_pool(params, in_dims, out_dims[0], input_)
output_ = in_q.dequantize(output_)
assert np.array_equal(output_, [[[4, 5], [7, 8]]])
def __init__(self,
*args,
stride=None,
padding=None,
pad_type="zero",
**kwargs):
if stride is None:
stride = StrideDim(1, 1)
if padding is None:
padding = PadDim(0)
super(FilterLikeParameters, self).__init__(*args, **kwargs)
self.stride = stride
self.padding = padding
self.pad_type = pad_type
self._ker_in_order = [['c', 'h', 'w']]
self._ker_out_order = [['c', 'h', 'w']]
def test_conf2d_pad_dilate():
weights = np.arange(9).reshape([1, 1, 3, 3])
filt = Conv2DFilterDim(3, 3, 1, 1)
stride = StrideDim(1)
pad = PadDim.same()
dilation = DilationDim(2)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
input_ = np.arange(16).reshape([1, 4, 4])
in_dims = Dim.named(c=1, h=4, w=4).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = conv2d(params, in_dims, out_dims[0], input_, weights, None, None)
assert np.array_equal(output_, [[[266., 206.], [98., 66.]]])
def test_conf2d_pad():
weights = np.arange(9).reshape([1, 1, 3, 3])
filt = Conv2DFilterDim(3, 3, 1, 1)
stride = StrideDim(1)
pad = PadDim.same()
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
input_ = np.arange(16).reshape([1, 4, 4])
in_dims = Dim.named(c=1, h=4, w=4).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = conv2d(params, in_dims, out_dims[0], input_, weights, None)
assert np.array_equal(output_, [[[73, 121, 154, 103], [171, 258, 294, 186],\
[279, 402, 438, 270], [139, 187, 202, 113]]])
def calc_shapes(node, spatial_size, input_size, kernel_shape):
padding = expand_dim(node.attrs.get('pads', None), 4 - spatial_size * 2, 0)
auto_pad = node.attrs.get('auto_pad', 'NOTSET')
output_shape = expand_dim(node.attrs.get('output_shape', None), 2 - spatial_size, 1)
output_padding = Dim2D(*expand_dim(node.attrs.get('output_padding', None), 2 - spatial_size, 0))
dilations = DilationDim(*expand_dim(node.attrs.get('dilations', None), 2 - spatial_size, 1))
strides = StrideDim(*expand_dim(node.attrs.get('strides', None), 2 - spatial_size, 1))
if output_shape:
total_padding = strides * (input_size - 1) + output_padding + ((kernel_shape - 1) * dilations + 1) - output_shape
if auto_pad == 'SAME_UPPER':
pad_start = total_padding // 2
pad_end = total_padding - pad_start
else:
pad_end = total_padding // 2
pad_start = total_padding - pad_end
padding = PadDim(pad_start.h, pad_end.h, pad_start.w, pad_end.w)
elif auto_pad == 'NOTSET':
assert padding, 'pads not set and auto_pad is NOTSET'
padding = PadDim(*padding)
elif auto_pad == 'VALID':
padding = PadDim.valid()
return padding, dilations, strides, output_padding
def test_conf2d_2_in_2_out_c():
weights = np.arange(4).reshape([1, 2, 2])
weights = np.append(weights, weights, axis=0)
weights = np.append(weights, weights, axis=0)
weights = weights.reshape([2, 2, 2, 2])
filt = Conv2DFilterDim(2, 2, 2, 2)
stride = StrideDim(1)
pad = PadDim.valid()
dilation = DilationDim(1)
params = Conv2DParameters("test",
filt=filt,
stride=stride,
padding=pad,
dilation=dilation,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
input_ = np.arange(9).reshape([1, 3, 3])
input_ = np.append(input_, input_, axis=0)
in_dims = Dim.named(c=2, h=3, w=3).impose_order(['c', 'h', 'w'])
out_dims = params.get_output_size([in_dims])
output_ = conv2d(params, in_dims, out_dims[0], input_, weights, None, None)
assert np.array_equal(output_, [[[38., 50.], [74., 86.]],\
[[38., 50.], [74., 86.]]])
def conv(cls, node, quantized=False, **kwargs):
all_nodes = kwargs['all_nodes']
G = kwargs['G']
valid_name = kwargs['valid_name']
inputs = [all_nodes[inp] for inp in node.input]
# input N x C x H x W
x = inputs[0]
x_rank = len(x[2].shape)
x_shape = x[2].shape
if x_shape[0] is not None:
real_in_shape = tuple(x_shape.copy())
if x_shape[0] > 1:
# support for multi batch is very limited
batch = x_shape[0]
logger.warning(
f"{valid_name} has a non 1 batch dimension of {batch} -"
" this is not supported by nntool or autotiler kernels")
else:
# if the batch is specified but is 1 then the input will be reshaped
# and the output will have the batch dim set as unknown.
batch = None
else:
real_in_shape = tuple(x_shape[1:])
batch = None
spatial_size = x_rank - 2
assert spatial_size == 2 or spatial_size == 1, "only 1D and 2D convolutions supported"
# Input error checking
undefined = []
if x_shape[1] is None:
# cope with swapped batch and channel due to bad initial reshape
if x_shape[0] == 1:
batch = None
x_shape = [x_shape[1], x_shape[0]] + list(x_shape[2:])
real_in_shape = x_shape[1:]
else:
undefined.append(f"input channel size of filter {valid_name} must be defined.")
if not all(dim is not None for dim in x_shape[-spatial_size:]):
undefined.append(f"input spatial size {x_shape} of filter {valid_name} must be defined.")
if undefined:
raise ValueError(f"{' '.join(undefined)}. You may need to override input dimensions.")
# M x C/group x kH x kW
weights_idx = 3 if quantized else 1
weights_node = inputs[weights_idx][0]
weights_node.name = f'{valid_name}_weights'
weights = cls.get_constant(inputs[weights_idx])
out_c = weights.shape[0]
group = node.attrs.get("group", 1)
in_c = x_shape[1]
filt_in_c = in_c // group
if in_c != weights.shape[1] * group:
raise ValueError(f'node {valid_name} has incorrect input channel '
f'dimension {in_c} expecting {weights.shape[1] * group}')
if spatial_size == 1:
filt_w = weights.shape[-1]
filt_h = h = 1
w = x_shape[-1]
# create a new constant node since we are changing the shape
weights = np.reshape(weights, (out_c, filt_in_c, filt_h, filt_w))
weights_node = ConstantInputParameters(f'{valid_name}_weights', value=weights,
dims=Dim.unnamed(
weights.shape))
cls.record_constant_qrec(inputs[1], weights_node, **kwargs)
else:
filt_h = weights.shape[-2]
filt_w = weights.shape[-1]
h = x_shape[-2]
w = x_shape[-1]
conv_in_shape = (in_c, h, w)
# h = 1 if spatial_size == 1 else (
# x_shape[-2] if x_shape[-2] is not None else 1)
# w = x_shape[-1] if x_shape[-1] is not None else 1
filt_dim = Conv2DFilterDim(filt_h, filt_w,
out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.ONNX_FILTER_ORDER)
biases_idx = 8 if quantized else 2
if len(inputs) > biases_idx:
biases_node = inputs[biases_idx][0]
biases = cls.get_constant(inputs[biases_idx])
else:
biases = np.zeros([out_c], dtype=np.float32)
biases_node = ConstantInputParameters(f'{valid_name}_biases', value=biases,
dims=Dim.unnamed(
biases.shape))
dilations = cls.pad_start_with(node.attrs.get("dilations", []), [1], 2)
strides = cls.pad_start_with(node.attrs.get("strides", []), [1], 2)
pad_dim = cls.calc_pad_dim(node, 4)
if batch is not None:
in_hint = ['n', 'c', 'h', 'w']
out_hint = ['n', 'c', 'h', 'w']
in_dim = Dim.named_ordered(n=batch, c=in_c, h=h, w=w)
ker_in_order = [
['n', 'c', 'h', 'w'],
['out_c', 'in_c', 'h', 'w'],
['out_c']]
ker_out_order = [['n', 'c', 'h', 'w']]
else:
in_hint = ['c', 'h', 'w']
out_hint = ['c', 'h', 'w']
in_dim = Dim.named_ordered(c=in_c, h=h, w=w)
ker_in_order = [
['c', 'h', 'w'],
['out_c', 'in_c', 'h', 'w'],
['out_c']]
ker_out_order = [['c', 'h', 'w']]
params = Conv2DParameters(valid_name,
filt=filt_dim,
stride=StrideDim(strides[0],
strides[1]),
dilation=DilationDim(dilations[0],
dilations[1]),
batch=batch,
groups=group,
padding=pad_dim,
ker_in_order=ker_in_order,
ker_out_order=ker_out_order,
has_bias=True,
in_dims_hint=[in_hint,
cls.ONNX_FILTER_ORDER, ['c']],
out_dims_hint=[out_hint])
if quantized:
qrecs = kwargs['qrecs']
x_zp = cls.get_constant(inputs[2])
x_scale = cls.get_constant(inputs[1])
x_qtype = QType(dtype=x_zp.dtype, scale=x_scale, zero_point=x_zp)
w_zp = cls.get_constant(inputs[5])
w_scale = cls.get_constant(inputs[4])
weights_node.qtype = w_qtype = QType(
dtype=w_zp.dtype, scale=w_scale,
zero_point=w_zp, quantized_dimension=0 if len(w_scale) > 1 else None)
o_zp = cls.get_constant(inputs[7])
o_scale = cls.get_constant(inputs[6])
o_qtype = QType(dtype=o_zp.dtype, scale=o_scale, zero_point=o_zp)
biases_node.qtype = b_qtype = QType(
dtype=biases.dtype, scale=w_scale*x_scale)
qrecs[NodeId(params)] = QRec.scaled(
in_qs=[x_qtype, w_qtype, b_qtype],
out_qs=[o_qtype],
)
else:
o_qtype = None
w_dim = Dim.named_ordered(
out_c=out_c, in_c=filt_in_c, h=filt_h, w=filt_w)
b_dim = Dim.named_ordered(c=out_c)
out_dims = params.get_output_size([in_dim, w_dim, b_dim])
G.add_edge(NNEdge(from_node=weights_node,
to_node=params, from_idx=0, to_idx=1))
G.add_edge(NNEdge(from_node=biases_node,
to_node=params, from_idx=0, to_idx=2))
# check if input needs a reshape
if conv_in_shape != real_in_shape:
r1_params = ReshapeParameters(f'{valid_name}_reshape_in',
old_shape=Dim.unnamed(real_in_shape),
shape=Dim.unnamed(conv_in_shape))
G.add_edge(
NNEdge(from_node=x[0], to_node=r1_params, from_idx=x[1], to_idx=0))
G.add_edge(NNEdge(from_node=r1_params,
to_node=params, from_idx=0, to_idx=0))
else:
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
# check if output needs a reshape
if spatial_size == 1:
if batch is not None:
oned_out_shape = [batch, out_dims[0].c, out_dims[0].w]
pout_dims = ProvisionalDim(oned_out_shape)
else:
oned_out_shape = [out_dims[0].c, out_dims[0].w]
pout_dims = ProvisionalDim([None] + oned_out_shape)
r2_params = ReshapeParameters(f'{valid_name}_reshape_out',
old_shape=out_dims[0],
shape=Dim.unnamed(oned_out_shape))
G.add_edge(NNEdge(from_node=params,
to_node=r2_params, from_idx=0, to_idx=0))
params = r2_params
else:
pout_dims = ProvisionalDim([batch] + out_dims[0].shape)
all_nodes[node.output[0]] = (params, 0, pout_dims, o_qtype)
return params
def version_1(cls, node: TFLiteNode, **kwargs):
node_opts = node.get_options(Conv2DOptions)
G = kwargs['G']
opts = kwargs['opts']
all_nodes = kwargs['all_nodes']
inputs = [all_nodes[t] for t in node.input]
x = inputs[0]
x_shape = x[2].shape
in_b, h, w, in_c = tuple(x_shape)
filt = inputs[1]
weights_node = filt[0]
filt_shape = filt[2].shape
# ['in_c', 'h', 'w', 'out_c']
filt_out_c, filt_h, filt_w, filt_in_c = tuple(filt_shape)
# get filter dimensions
if filt_h > h or filt_w > w:
LOG.warning(
"Filter %s of shape [%dx%d] is bigger than input of shape [%dx%d]",
node.name, filt_h, filt_w, h, w)
filt_dim = Conv2DFilterDim(filt_h, filt_w, filt_out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.TF_LITE_FILTER_ORDER)
# compute padding
pad = cls.get_tf_padding(node_opts.Padding())
# does it have biases
if len(inputs) > 2:
bias = inputs[2]
bias_node = bias[0]
else:
bias_node = ConstantInputParameters(
f'{node.name}_bias',
dims=Dim.unnamed([filt_out_c]),
value=np.zeros([filt_out_c], dtype=np.float32)) # TODO - check
params = Conv2DParameters(
node.name,
filt=filt_dim,
stride=StrideDim(node_opts.StrideH(), node_opts.StrideW()),
dilation=DilationDim(node_opts.DilationHFactor(),
node_opts.DilationWFactor()),
padding=pad,
has_bias=True,
in_dims_hint=SparseList([['h', 'w', 'c'],
cls.TF_LITE_FILTER_ORDER.copy(),
['out_c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]),
constant_store=G.constant_store)
G.add_edge(NNEdge(from_node=weights_node, to_node=params, to_idx=1))
G.add_edge(NNEdge(from_node=bias_node, to_node=params, to_idx=2))
cls.new_load_filter_parameters(G, params, node.input[0], weights_node,
bias_node, node.output[0], opts)
# if opts.get('load_dequantized'):
# weights_node.value, bias_node.value = cls.load_dequantized_filter_parameters(
# node.input, bias_node.value)
# else:
# qrec, weights_node.value, bias_node.value = cls.load_filter_parameters(G, params, node.input, bias_node.value,
# node.output, opts)
# if qrec:
# G.quantization[NodeId(weights_node)].out_qs[0] = qrec.in_qs[1]
# G.quantization[NodeId(bias_node)].out_qs[0] = qrec.in_qs[2]
in_dim = Dim.named_ordered(h=h, w=w, c=in_c)
out_dims = params.get_output_size(
[in_dim,
Dim.unnamed(filt_dim.shape),
Dim.unnamed([filt_out_c])])
pout_dims = ProvisionalDim([in_b] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
params = cls.fuse_activation(node_opts, node.name, params, **kwargs)
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def pool(cls, node, pool_type=None, copy_qtype=False, **kwargs):
all_nodes = kwargs['all_nodes']
G = kwargs['G']
valid_name = kwargs['valid_name']
inputs = [all_nodes[inp] for inp in node.input]
x = inputs[0]
x_shape = x[2].shape
x_feature_shape = x_shape[2::]
input_rank = len(x_feature_shape)
in_c = x_shape[1]
kernel_shape = node.attrs["kernel_shape"]
kernel_rank = len(kernel_shape)
if input_rank != kernel_rank:
raise ValueError(
f'error in ONNX graph. {pool_type} pool {valid_name} '
f'has a different input spatial rank {input_rank} to kernel rank {kernel_rank}'
)
spatial_size = kernel_rank
if kernel_rank > 2:
raise NotImplementedError(
f'{pool_type} pool {valid_name} is a {kernel_rank}D pool '
'which is not supported by NNTOOL')
strides = node.attrs.get("strides", [1] * spatial_size)
stride_is_one = all(stride == 1 for stride in strides)
dilations = node.attrs.get("dilations", [1] * spatial_size)
if any(dilation > 1 for dilation in dilations):
raise ValueError(valid_name + " with dilation not supported")
# ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pad_dim = cls.calc_pad_dim(node, 2)
if spatial_size == 1:
strides = [1] + strides
dilations = [1] + dilations
kernel_shape = [1] + kernel_shape
h = 1
w = x_shape[2]
x_feature_shape = [1] + x_feature_shape
else:
h = x_shape[2]
w = x_shape[3]
strides = node.attrs.get("strides", [1] * spatial_size)
stride_is_one = all(stride == 1 for stride in strides)
dilations = node.attrs.get("dilations", [1] * spatial_size)
if any(dilation > 1 for dilation in dilations):
raise ValueError(valid_name + " with dilation not supported")
# ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pad_dim = cls.calc_pad_dim(node, spatial_size)
if pad_dim.is_same:
pad_dim.calculate_same(
Dim.named_ordered(h=h, w=w),
PoolFilterDim(kernel_shape[0], kernel_shape[1]),
StrideDim(strides[0], strides[1]))
# Note: This needs to check dilation if it is added
filter_matches_input = (all(
k_dim >= (x_dim + pad) for k_dim, x_dim, pad in zip(
kernel_shape, x_feature_shape, [pad_dim.h, pad_dim.w])))
if filter_matches_input and stride_is_one:
params = GlobalPoolingParameters(valid_name,
pool_type=pool_type,
axis=[1, 2],
keep_dims=True,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
else:
params = PoolingParameters(
valid_name,
filt=PoolFilterDim(kernel_shape[0], kernel_shape[1]),
stride=StrideDim(strides[0], strides[1]),
padding=pad_dim,
pool_type=pool_type,
in_dims_hint=[['c', 'h', 'w']],
out_dims_hint=[['c', 'h', 'w']])
in_dim = Dim.named_ordered(c=in_c, h=h, w=w)
out_dims = params.get_output_size([in_dim])
pout_dims = ProvisionalDim([x_shape[0]] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
all_nodes[node.output[0]] = (params, 0, pout_dims,
x[3] if copy_qtype else None)
return params
def version_1(cls, node: TFLiteNode, **kwargs):
node_opts = node.get_options(Conv2DOptions)
G = kwargs['G']
opts = kwargs['opts']
all_nodes = kwargs['all_nodes']
inputs = [all_nodes[t] for t in node.input]
x = inputs[0]
x_shape = x[2].shape
in_b, h, w, in_c = tuple(x_shape)
filt = inputs[1]
filt_tensor = node.input[1]
filt_shape = filt[2].shape
# ['in_c', 'h', 'w', 'out_c']
filt_out_c, filt_h, filt_w, filt_in_c = tuple(filt_shape)
# get filter dimensions
filt_tensor.used = True
if filt_h > h or filt_w > w:
LOG.warning(
"Filter %s of shape [%dx%d] is bigger than input of shape [%dx%d]",
node.name, filt_h, filt_w, h, w)
filt_dim = Conv2DFilterDim(filt_h, filt_w, filt_out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.TF_LITE_FILTER_ORDER)
# compute padding
pad = cls.get_tf_padding(node_opts.Padding())
# does it have biases
has_bias = len(inputs) > 2
if has_bias:
node.input[2].used = True
params = Conv2DParameters(node.name,
filt=filt_dim,
stride=StrideDim(node_opts.StrideH(),
node_opts.StrideW()),
dilation=DilationDim(
node_opts.DilationHFactor(),
node_opts.DilationWFactor()),
padding=pad,
has_bias=has_bias,
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]),
constant_store=G.constant_store)
if opts.get('load_dequantized'):
cls.load_dequantized_filter_parameters(params, node.input)
else:
cls.load_filter_parameters(G, params, node.input, node.output,
opts)
in_dim = Dim.named_ordered(h=h, w=w, c=in_c)
out_dims = params.get_output_size([in_dim])
pout_dims = ProvisionalDim([in_b] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
params = cls.fuse_activation(node_opts, node.name, params, **kwargs)
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def version_1(cls, node: TFLiteNode, **kwargs):
node_opts = node.get_options(Conv2DOptions)
G = kwargs['G']
opts = kwargs['opts']
all_nodes = kwargs['all_nodes']
inputs = [all_nodes[t] for t in node.input]
x = inputs[0]
x = cls.remove_known_batch_dimension(G, x, node)
x_shape = x[2].shape
in_b, h, w, in_c = tuple(x_shape)
filt = inputs[1]
weights_node = filt[0]
filt_shape = filt[2].shape
# ['in_c', 'h', 'w', 'out_c']
filt_out_c, filt_h, filt_w, filt_in_c = tuple(filt_shape)
# get filter dimensions
if filt_h > h or filt_w > w:
LOG.warning(
"Filter %s of shape [%dx%d] is bigger than input of shape [%dx%d]",
node.name, filt_h, filt_w, h, w)
filt_dim = Conv2DFilterDim(filt_h, filt_w, filt_out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.TF_LITE_FILTER_ORDER)
# compute padding
pad = cls.get_tf_padding(node_opts.Padding())
# does it have biases
if len(inputs) > 2:
bias = inputs[2]
bias_node = bias[0]
else:
bias_node = ConstantInputParameters(
f'{node.name}_bias',
dims=Dim.unnamed([filt_out_c]),
value=np.zeros([filt_out_c], dtype=np.float32)) # TODO - check
groups = in_c // filt_in_c
params = Conv2DParameters(
node.name,
filt=filt_dim,
stride=StrideDim(node_opts.StrideH(), node_opts.StrideW()),
dilation=DilationDim(node_opts.DilationHFactor(),
node_opts.DilationWFactor()),
groups=groups,
padding=pad,
has_bias=True,
in_dims_hint=[['h', 'w', 'c'],
cls.TF_LITE_FILTER_ORDER.copy(), ['out_c']],
out_dims_hint=[['h', 'w', 'c']])
G.add_edge(NNEdge(from_node=weights_node, to_node=params, to_idx=1))
G.add_edge(NNEdge(from_node=bias_node, to_node=params, to_idx=2))
cls.new_load_filter_parameters(G, params, params.filter.actual_shape,
params.filter.get_order_idx('out_c'),
node.input[0], weights_node, bias_node,
node.output[0], opts)
in_dim = Dim.named_ordered(h=h, w=w, c=in_c)
out_dims = params.get_output_size(
[in_dim,
Dim.unnamed(filt_dim.shape),
Dim.unnamed([filt_out_c])])
pout_dims = ProvisionalDim([None] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
oparams = cls.fuse_activation(node_opts, node.name, params, **kwargs)
all_nodes[node.output[0]] = (oparams, 0, pout_dims)
return oparams
def _common(cls, node: TFLiteNode, **kwargs):
node_opts = node.get_options(DepthwiseConv2DOptions)
G = kwargs['G']
opts = kwargs['opts']
all_nodes = kwargs['all_nodes']
inputs = [all_nodes[t] for t in node.input]
x = inputs[0]
x = cls.remove_known_batch_dimension(G, x, node)
x_shape = x[2].shape
in_b, h, w, in_c = tuple(x_shape)
filt = inputs[1]
weights_node = filt[0]
filt_shape = filt[2].shape
# ['in_c', 'h', 'w', 'out_c']
filt_in_c, filt_h, filt_w, filt_out_c = tuple(filt_shape)
# get filter dimensions
if filt_h > h or filt_w > w:
LOG.warning(
"Filter %s of shape [%dx%d] is bigger than input of shape [%dx%d]",
node.name, filt_h, filt_w, h, w)
filt_dim = Conv2DFilterDim(filt_h, filt_w, filt_out_c, in_c=filt_in_c)
filt_dim = filt_dim.impose_order(cls.TF_LITE_DW_FILTER_ORDER)
# multiplier should match filter
check(filt_dim.out_c == node_opts.DepthMultiplier() * in_c,
"invalid multiplier")
groups = filt_dim.out_c // node_opts.DepthMultiplier()
# compute padding
pad = cls.get_tf_padding(node_opts.Padding())
# does it have biases
if len(inputs) > 2:
bias = inputs[2]
bias_node = bias[0]
else:
bias_node = ConstantInputParameters(
f'{node.name}_bias',
dims=Dim.unnamed([filt_out_c]),
value=np.zeros([filt_out_c], dtype=np.float32)) # TODO - check
# TFLITE produces single channel input DW convolutions with the
# multiplier equal to the number of out channels. This is just
# a normal convolution and since we don't handle the channel
# multiplier at present (but can) just convert them to normal
# convolutions
convert_to_conv = in_c == 1 and groups == 1
if convert_to_conv:
filt_dim.impose_order(cls.TF_LITE_FILTER_ORDER)
params = Conv2DParameters(
node.name,
filt=filt_dim,
stride=StrideDim(node_opts.StrideH(), node_opts.StrideW()),
dilation=DilationDim(node_opts.DilationHFactor(),
node_opts.DilationWFactor()),
padding=pad,
has_bias=True,
in_dims_hint=[['h', 'w', 'c'],
cls.TF_LITE_FILTER_ORDER.copy(), ['out_c']],
out_dims_hint=[['h', 'w', 'c']])
else:
filt_dim.impose_order(cls.TF_LITE_DW_FILTER_ORDER)
params = Conv2DParameters(
node.name,
filt=filt_dim,
stride=StrideDim(node_opts.StrideH(), node_opts.StrideW()),
dilation=DilationDim(node_opts.DilationHFactor(),
node_opts.DilationWFactor()),
padding=pad,
groups=groups,
multiplier=node_opts.DepthMultiplier(),
has_bias=True,
tf_depthwise=True,
in_dims_hint=[['h', 'w', 'c'],
cls.TF_LITE_DW_FILTER_ORDER.copy(), ['out_c']],
out_dims_hint=[['h', 'w', 'c']])
G.add_edge(NNEdge(from_node=weights_node, to_node=params, to_idx=1))
G.add_edge(NNEdge(from_node=bias_node, to_node=params, to_idx=2))
cls.new_load_filter_parameters(G,
params,
params.filter.actual_shape,
params.filter.get_order_idx('out_c'),
node.input[0],
weights_node,
bias_node,
node.output[0],
opts,
dw_to_pw=convert_to_conv)
in_dim = Dim.named_ordered(h=h, w=w, c=in_c)
out_dims = params.get_output_size(
[in_dim,
Dim.unnamed(filt_dim.shape),
Dim.unnamed([filt_out_c])])
pout_dims = ProvisionalDim([in_b] + out_dims[0].shape)
G.add_edge(
NNEdge(from_node=x[0], to_node=params, from_idx=x[1], to_idx=0))
params = cls.fuse_activation(node_opts, node.name, params, **kwargs)
all_nodes[node.output[0]] = (params, 0, pout_dims)
return params
def add_depthwise_convolution(G,
tensors,
name,
subgraph,
_,
op,
load_tensors=False,
dequantize=False):
conv_opts = DepthwiseConv2DOptions.DepthwiseConv2DOptions()
conv_opts.Init(op.BuiltinOptions().Bytes, op.BuiltinOptions().Pos)
# get filter dimensions
inp = get_input_size(tensors, subgraph, op, 0, order=TF_LITE_IN_OUT_ORDER)
filt = get_input_size(tensors,
subgraph,
op,
1,
order=TF_LITE_DW_FILTER_ORDER)
filt = Conv2DFilterDim(filt['h'], filt['w'],\
filt['out_c'], in_c=1)
# multiplier should match filter
check(filt.out_c == conv_opts.DepthMultiplier() * inp['c'],
"invalid multiplier")
groups = filt.out_c // conv_opts.DepthMultiplier()
# compute padding
pad = get_tf_padding(conv_opts.Padding())
# does it have biases
has_bias = op.InputsLength() > 2
# TFLITE produces single channel input DW convolutions with the
# multiplier equal to the number of out channels. This is just
# a normal convolution and since we don't handle the channel
# multiplier at present (but can) just convert them to normal
# convolutions
convert_to_conv = inp['c'] == 1 and groups == 1
if convert_to_conv:
filt.impose_order(TF_LITE_FILTER_ORDER)
node = Conv2DParameters(name,
filt=filt,
stride=StrideDim(conv_opts.StrideH(),
conv_opts.StrideW()),
padding=pad,
has_bias=has_bias,
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]),
constant_store=G.constant_store)
else:
filt.impose_order(TF_LITE_DW_FILTER_ORDER)
node = Conv2DParameters(name,
filt=filt,
stride=StrideDim(conv_opts.StrideH(),
conv_opts.StrideW()),
padding=pad,
groups=groups,
multiplier=conv_opts.DepthMultiplier(),
has_bias=has_bias,
tf_depthwise=True,
in_dims_hint=SparseList([['h', 'w', 'c']]),
out_dims_hint=SparseList([['h', 'w', 'c']]),
constant_store=G.constant_store)
if load_tensors:
node.weights = get_tensor(G.model,
tensors,
subgraph,
op,
1,
dequantize=dequantize)
# If we've converted to a normal conv then change the weight order
if convert_to_conv:
node.weights = node.weights.transpose(TF_LITE_DW_FILTER_TRANSPOSE)
if has_bias:
node.biases = get_tensor(G.model,
tensors,
subgraph,
op,
2,
dequantize=dequantize)
return fuse_activation(G, conv_opts, name, node)
