def version_11(cls, node, **kwargs):
x = kwargs["tensor_dict"][node.inputs[0]]
x_rank = len(x.get_shape())
storage_format, compute_format = get_data_format(x_rank)
attrs = copy.deepcopy(node.attrs)
attrs["data_format"] = storage_format
mode = attrs.get("mode", "DCR")
if mode == "CRD":
# need native computation
bsize = attrs.get("blocksize")
batch, channel, height, width = x.shape
csize = channel // (bsize**2)
reshape_node = tf.reshape(
x, [batch, csize, bsize, bsize, height, width])
transpose_node = tf.transpose(reshape_node,
perm=[0, 1, 4, 2, 5, 3])
return [
tf.reshape(transpose_node,
[batch, csize, height * bsize, width * bsize])
]
else:
return [
cls.make_tensor_from_onnx_node(node,
attrs=attrs,
c_first_cuda_only=True,
**kwargs)
]
def _common(cls, node, **kwargs):
x = kwargs["tensor_dict"][node.inputs[0]]
x_rank = len(x.get_shape())
storage_format, compute_format = get_data_format(x_rank)
attrs = copy.deepcopy(node.attrs)
attrs["data_format"] = storage_format
if sys_config.device == 'CUDA' and x.dtype not in {
tf.uint8, tf.float16, tf.float32
}:
# Tensorflow GPU version doesn't support these datatype but CPU version support
with tf.device("/cpu:0"): # run it on cpu
compute_format = compute_format.replace("C", "") + "C"
pre_perm = get_perm_from_formats(storage_format,
compute_format)
post_perm = get_perm_from_formats(compute_format,
storage_format)
x_t = tf.transpose(x, perm=pre_perm)
y = tf.nn.space_to_depth(x_t, attrs["blocksize"],
compute_format)
y = tf.transpose(y, perm=post_perm)
else:
y = cls.make_tensor_from_onnx_node(node,
attrs=attrs,
c_first_cuda_only=True,
**kwargs)
return [y]
def pool_v11(cls, node, input_dict, pooling_type, strict=True):
x = input_dict[node.inputs[0]]
kernel_shape = node.attrs["kernel_shape"]
spatial_size = len(kernel_shape)
x_rank = spatial_size + 2
kernel_shape = node.attrs["kernel_shape"]
strides = node.attrs.get("strides", [1] * spatial_size)
dilations = node.attrs.get("dilations", [1] * spatial_size)
ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pads = node.attrs.get("auto_pad", "NOTSET")
if pads == "NOTSET":
pads = node.attrs.get("pads", [0] * spatial_size * 2)
if spatial_size > 3:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPool with {}D input".format(x_rank), "Tensorflow")
if pooling_type == "MAX_WITH_ARGMAX" and x_rank != 4:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPool with {}D input".format(x_rank), "Tensorflow")
if node.attrs.get("storage_order", 0) != 0:
exception.OP_UNSUPPORTED_EXCEPT("MaxPool with column major",
"Tensorflow")
storage_format, _ = get_data_format(x_rank)
need_trans = storage_format.startswith("NC")
if need_trans:
compute_format = "N" + storage_format[2:] + "C"
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, compute_format))
dp = DilatedPooling(input=x,
kernel_shape=kernel_shape,
strides=strides,
dilations=dilations,
padding=pads,
ceil_mode=ceil_mode)
# select correct op depending on the pooling type
pooling_op = lambda : (dp.dilated_maxpool(), None) if \
pooling_type == "MAX" else dp.dilated_maxpool_with_argmax()
# select the correct transpose ops depending on the input storage format
perm = get_perm_from_formats(compute_format, storage_format)
postprocess_op = lambda pooled, argmax: (
tf.transpose(pooled, perm=perm)
if need_trans else pooled, tf.transpose(argmax, perm=perm)
if need_trans and argmax is not None else argmax)
pooled, argmax = pooling_op()
pooled, argmax = postprocess_op(pooled, argmax)
result = [pooled] if argmax is None else [pooled, argmax]
return result
def version_1(cls, node, **kwargs):
x = kwargs["tensor_dict"][node.inputs[0]]
x_rank = len(x.get_shape())
storage_format, compute_format = get_data_format(x_rank)
attrs = copy.deepcopy(node.attrs)
attrs["data_format"] = storage_format
return [
cls.make_tensor_from_onnx_node(node,
attrs=attrs,
c_first_cuda_only=True,
**kwargs)
]
def c_last_only(cls, tf_func, inputs, attrs):
""" Handle operator that channel last only is supported.
Add two transposes anyway.
:param tf_func: Callable Tf function.
:param inputs: Inputs tensor.
:param attrs: Attributes.
:return: Tensor.
"""
storage_format, compute_format = get_data_format(len(inputs[0].get_shape()))
compute_format = compute_format.replace("C", "") + "C"
return cls._tuck_transpose(tf_func, inputs, attrs,
(storage_format, compute_format))
def max_unpool(cls, node, input_dict):
"""
MaxUnpooling operation
"""
x = input_dict[node.inputs[0]]
ind = input_dict[node.inputs[1]]
if len(node.inputs) > 2:
output_shape = input_dict.get(node.inputs[2], None)
else:
output_shape = None
kernel_shape = node.attrs["kernel_shape"]
spatial_size = len(kernel_shape)
x_rank = spatial_size + 2
storage_format, _ = get_data_format(x_rank)
# if strides are not provided default is 1 along each spatial axis
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", None)
input_shape = tf_shape(x)
default_shape = cls._get_default_shape(input_shape, kernel_shape,
strides)
need_trans = storage_format != "NHWC"
if need_trans:
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, "NHWC"))
ind = tf.transpose(ind,
perm=get_perm_from_formats(
storage_format, "NHWC"))
# default_shape to NHWC storage format
default_shape = [input_shape[0]] + default_shape + \
[input_shape[1]]
unpooled = cls._unpool(x, ind, default_shape)
if need_trans:
unpooled = tf.transpose(unpooled,
perm=get_perm_from_formats(
"NHWC", storage_format))
if output_shape is not None:
pads = cls._get_pads_from_output_shape(unpooled, output_shape)
if pads is not None:
unpooled = cls._pad_output(unpooled, pads, 0)
return [unpooled]
def _tuck_transpose(cls, tf_func, inputs, attrs, data_format=None):
x = inputs[0]
x_rank = len(x.get_shape())
if not data_format:
data_format = get_data_format(x_rank)
pre_perm = get_perm_from_formats(data_format[0], data_format[1])
post_perm = get_perm_from_formats(data_format[1], data_format[0])
attrs["data_format"] = data_format[1]
if pre_perm != list(range(x_rank)):
x_t = tf.transpose(x, perm=pre_perm)
y = cls._run_tf_func(tf_func, [x_t] + inputs[1:], attrs)
y_t = tf.transpose(y, perm=post_perm)
return y_t
return cls._run_tf_func(tf_func, inputs, attrs)
def __init__(self,
input,
kernel_shape,
strides,
dilations,
padding="VALID",
ceil_mode=False,
count_include_pad=False,
pooling_type="MAX",
p=2):
self.input = tf.convert_to_tensor(input)
self.kernel_shape = kernel_shape
self.strides = strides
self.dilations = dilations
self.padding = padding
self.is_explicit_padding = type(padding) is list
self.ceil_mode = ceil_mode
self.count_include_pad = count_include_pad
self.pooling_type = pooling_type.upper()
self.p = p
self.is_known_shape = self.input.shape.is_fully_defined()
self.spatial_size = len(kernel_shape)
self.input_rank = self.spatial_size + 2
# if the rank is not defined, set it to the calculated input_rank
# rank should be known for ops like tf.gather_nd
if not input.shape.rank:
input.set_shape([None] * self.input_rank)
self.orig_input_shape = tf_shape(input)
self.input_shape = self.orig_input_shape
if pooling_type.startswith("MAX"):
self.padding_constant = input.dtype.min
else:
self.padding_constant = 0
self.storage_format, self.compute_format = get_data_format(
self.spatial_size + 2)
self.need_trans = self.storage_format != self.compute_format
def _common(cls, node, **kwargs):
tensor_dict = kwargs['tensor_dict']
feat = tensor_dict[node.inputs[0]]
boxes = tensor_dict[node.inputs[1]]
indx = tensor_dict[node.inputs[2]]
output_height = node.attrs['output_height']
output_width = node.attrs['output_width']
sampling_ratio = node.attrs['sampling_ratio']
spatial_scale = node.attrs['spatial_scale']
adaptive_ratio = False
if sampling_ratio <= 0:
sampling_ratio = int((output_height + output_width) / 2)
adaptive_ratio = True
logger.warning("Do not fully support sampling_ratio <= 0.")
boxes = boxes * spatial_scale
feat_rank = len(feat.shape)
storage_format, _ = get_data_format(feat_rank)
need_trans = storage_format.startswith("NC")
if need_trans:
compute_format = "N" + storage_format[2:] + "C"
feat = tf.transpose(feat,
perm=get_perm_from_formats(
storage_format, compute_format))
ret = crop_and_resize(feat,
boxes,
tf.cast(indx, tf.int32),
(output_height, output_width),
sampling_ratio,
adaptive_ratio=adaptive_ratio)
ret = tf.nn.avg_pool(ret, [1, sampling_ratio, sampling_ratio, 1],
[1, sampling_ratio, sampling_ratio, 1],
padding='SAME',
data_format='NHWC')
ret = tf.transpose(ret, perm=(0, 3, 1, 2))
return [ret]
def conv(cls, node, input_dict, transpose=False):
""" Convolution method for both conv and transposed conv
For transposed conv,
Attr pads is not used for input, but declares how much output is padded.
Here, output means output from transposed conv which already pad output_padding if set.
So the pseudo explanation for output should be:
output = conv_transpose_output + output_padding - pads
And conv_transpose_output shape should be:
conv_transpose_output_shape[i] = strides[i] * (input_shape[i] - 1) + kernel_shape[i]
"""
x = input_dict[node.inputs[0]]
x_rank = len(x.get_shape())
x_shape = tf_shape(x, tf.int32)
spatial_size = x_rank - 2
storage_format, compute_format = get_data_format(x_rank)
compute_c_idx = compute_format.find("C")
spatial_format = "".join([d for d in compute_format if d not in ["N", "C"]])
in_weights = input_dict[node.inputs[1]]
weights_rank = len(in_weights.get_shape())
if transpose:
# Translate weights from (C x M x KH x KW) to (KH x KW X M X C)
perm = list(range(2, weights_rank)) + [1, 0]
else:
# Translate weights from (M x C x KH x KW) to (KH x KW X C X M)
perm = list(range(2, weights_rank)) + [1, 0]
if "kernel_shape" in node.attrs.keys():
kernel_shape = node.attrs["kernel_shape"]
if in_weights.get_shape().is_fully_defined():
assert in_weights.get_shape().as_list()[2:] == kernel_shape, (
"kernel_shape "
"attr of convolution does not match the actual weight "
"passed to this operation, attr {}, actual {}").format(
kernel_shape,
in_weights.get_shape().as_list())
else:
kernel_shape = tf_shape(in_weights, tf.int32)[2:]
weights = tf.transpose(in_weights, perm)
dilations = node.attrs.get("dilations", [1] * spatial_size)
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", [0, 0] * spatial_size)
# Check auto_pad nonexistent or NOTSET first
if "auto_pad" not in node.attrs or node.attrs["auto_pad"] == "NOTSET":
if not transpose:
if pads != [0, 0] * spatial_size:
x = PadMixin.get_padding_as_op(x, pads)
pad_mode = "VALID"
else:
pad_mode = "NOTSET"
# Then we use auto_pad to setup pad_mode
elif node.attrs["auto_pad"] == "SAME_UPPER":
pad_mode = "SAME"
elif node.attrs["auto_pad"] == "VALID":
pad_mode = "VALID"
elif node.attrs["auto_pad"] == "SAME_LOWER":
pad_mode = PAD_TF_INCOMPATIBLE
else:
raise ValueError("Invalid auto_pad attribute: {}".format(
node.attrs["auto_pad"]))
# Currently auto_pad = SAME_LOWER is not supported
if pad_mode is PAD_TF_INCOMPATIBLE:
if transpose:
exception.OP_UNSUPPORTED_EXCEPT(
"ConvTranspose with auto_pad `SAME_LOWER`", "Tensorflow")
else:
exception.OP_UNSUPPORTED_EXCEPT("Conv with auto_pad `SAME_LOWER`",
"Tensorflow")
group = node.attrs.get("group", 1)
weight_shape = weights.get_shape().as_list()
# Is this convolution depthwise we can support?
depthwise = (x_rank == 4 and len(weight_shape) == 4 and group != 1 and
not transpose and not (None in weight_shape))
if depthwise and isinstance(x_shape, np.ndarray):
depthwise = bool(group == x_shape[1])
if depthwise is True:
# Depthwise convolution.
# The convolution kernel layout in tf.depthwise_conv is:
# [filter_height, filter_width, in_channels, channel_multiplier]
# Weight is now (KH x KW X C/g X M), or more precisely, (KH x KW X C/g X (g * M/g)),
# we reshape it to (KH x KW x C x M/g)
# NOTE: Assuming weight has fixed shape.
depthwise_filter_shape = weight_shape[0:2] + [
-1, weight_shape[3] // group
]
weights = tf.reshape(weights, depthwise_filter_shape)
if not sys_config.device == 'CUDA':
# transpose input to NHWC layout
x = tf.transpose(x,
perm=get_perm_from_formats(storage_format,
compute_format))
weight_groups = [weights]
xs = [x]
else:
weight_groups = tf.split(weights, num_or_size_splits=group, axis=-1)
if sys_config.device == 'CUDA':
if group == 1:
xs = [x]
else:
xs = tf.split(x, num_or_size_splits=group, axis=1)
else:
x = tf.transpose(x,
perm=get_perm_from_formats(storage_format,
compute_format))
if group == 1:
xs = [x]
else:
xs = tf.split(x, num_or_size_splits=group, axis=-1)
if transpose:
if dilations != [1] * spatial_size:
raise RuntimeError("Cannot set non-1 dilation for conv transpose.")
convolved = []
# this is a workaround for tensorflow AutoGraph not detecting
# corretly x. This is fixed in tf>=2.2.0
x = None
for (x, weight) in zip(xs, weight_groups):
x_spatial_shape = [
x_shape[storage_format.find(d)] for d in spatial_format
]
weights_shape = tf_shape(weights, tf.int32)
output_shape = node.attrs.get("output_shape", None)
conv_output_shape = [x_shape[storage_format.find("N")]]
# calculate output shape
if pad_mode == "NOTSET":
if output_shape is None:
conv_output_shape += [
strides[i] * x_spatial_shape[i] - strides[i] +
(kernel_shape[i] - 1) * dilations[i] + 1
for i in list(range(spatial_size))
]
else:
conv_output_shape += [
s + pads[i] + pads[spatial_size + i]
for i, s in enumerate(output_shape[-2:])
]
conv_output_shape.insert(compute_c_idx, weights_shape[-2])
# make strides to match input rank
strides_full = [1] + strides
strides_full.insert(compute_c_idx, 1)
# get corresponding function in tf
if spatial_size == 1:
conv_func = tf.nn.conv1d_transpose
strides_full = strides[0]
elif spatial_size == 2:
conv_func = tf.nn.conv2d_transpose
elif spatial_size == 3:
conv_func = tf.nn.conv3d_transpose
else:
raise NotImplementedError(
"Transposed convolution for {}d is not implemented in Tensorflow"
.format(spatial_size))
# use raw input x to do transposed conv
conv_rs = conv_func(x,
weight,
conv_output_shape,
strides_full,
padding="VALID",
data_format=compute_format)
# pad output first by output_padding attr
if "output_padding" in node.attrs and output_shape is None:
output_padding = [[0, 0]
] + [[0, p] for p in node.attrs["output_padding"]]
output_padding.insert(compute_c_idx, [0, 0])
conv_rs = tf.pad(conv_rs, output_padding)
# remove pads set in pads attr
conv_rs_shape = tf_shape(conv_rs, tf.int32)
conv_rs_shape_list = [
conv_rs_shape[i] for i in range(conv_rs.shape.rank)
]
begin = [0] + pads[:spatial_size]
begin.insert(compute_c_idx, 0)
size = [
s if d in ["N", "C"] else s - pads[spatial_format.find(d)] -
pads[spatial_format.find(d) + spatial_size]
for d, s in zip(compute_format, conv_rs_shape_list)
]
conv_rs = tf.slice(conv_rs, begin=begin, size=size)
convolved.append(conv_rs)
else:
# No need to check pads if auto_pad is specifically provided.
# The assumption is that once auto_pad is provided as either VALID
# or SAME_UPPER (SAME_LOWER is currently not supported in TF) the
# output_shape will always be inferred. That is, the output_shape
# and output_padding will not be used in this case.
if pad_mode == "VALID":
conv_output_shape += [
strides[i] * (x_spatial_shape[i] - 1) + weights_shape[i]
for i in list(range(spatial_size))
]
else:
conv_output_shape += [
strides[i] * x_spatial_shape[i]
for i in list(range(spatial_size))
]
conv_output_shape.insert(compute_c_idx, weights_shape[-2])
# make strides to match input rank
strides_full = [1] + strides
strides_full.insert(compute_c_idx, 1)
# get corresponding function in tf
if spatial_size == 1:
conv_func = tf.nn.conv1d_transpose
strides_full = strides[0]
elif spatial_size == 2:
conv_func = tf.nn.conv2d_transpose
elif spatial_size == 3:
conv_func = tf.nn.conv3d_transpose
else:
raise NotImplementedError(
"Transposed convolution for {}d is not implemented in Tensorflow"
.format(spatial_size))
# use raw input x to do transposed conv
conv_rs = conv_func(x,
weight,
conv_output_shape,
strides_full,
padding=pad_mode,
data_format=compute_format)
convolved.append(conv_rs)
else: # not transpose:
if depthwise is True:
if compute_format == "NHWC":
strides = [1] + strides + [1]
elif compute_format == 'NCHW':
strides = [1, 1] + strides
else:
raise ValueError("Invalid compute_format: {}".format(compute_format))
convolved = [
tf.nn.depthwise_conv2d(x,
weight,
padding=pad_mode,
strides=strides,
dilations=dilations,
data_format=compute_format)
for (x, weight) in zip(xs, weight_groups)
]
else:
convolved = [
tf.nn.convolution(x,
weight,
padding=pad_mode,
strides=strides,
dilations=dilations,
data_format=compute_format)
for (x, weight) in zip(xs, weight_groups)
]
if len(node.inputs) == 2:
if sys_config.device == 'CUDA':
output = tf.concat(convolved, axis=1)
else:
output = tf.concat(convolved, axis=-1)
output = tf.transpose(output,
perm=get_perm_from_formats(
compute_format, storage_format))
else:
bias = input_dict[node.inputs[2]]
bias = cls.explicit_broadcast([x, bias], compute_c_idx)
if sys_config.device == 'CUDA':
output = tf.concat(convolved, axis=1)
output = tf.add(output, bias)
else:
output = tf.concat(convolved, axis=-1)
output = tf.add(output, bias)
output = tf.transpose(output,
perm=get_perm_from_formats(
compute_format, storage_format))
return [output]
def conv(cls, node, input_dict, transpose=False):
""" Convolution method for both conv and transposed conv
For transposed conv,
Attr pads is not used for input, but declares how much output is padded.
Here, output means output from transposed conv which already pad output_padding if set.
So the pseudo explanation for output should be:
output = conv_transpose_output + output_padding - pads
And conv_transpose_output shape should be:
conv_transpose_output_shape[i] = strides[i] * (input_shape[i] - 1) + kernel_shape[i]
"""
x = input_dict[node.inputs[0]]
x_rank = len(x.get_shape())
x_shape = x.get_shape().as_list()
spatial_size = x_rank - 2
support_cuda = supports_device("CUDA")
storage_format, compute_format = get_data_format(x_rank)
compute_c_idx = compute_format.find("C")
spatial_format = "".join([d for d in compute_format if d not in ["N", "C"]])
in_weights = input_dict[node.inputs[1]]
weights_rank = len(in_weights.get_shape())
if transpose:
# Translate weights from (C x M x KH x KW) to (KH x KW X M X C)
perm = list(range(2, weights_rank)) + [1, 0]
else:
# Translate weights from (M x C x KH x KW) to (KH x KW X C X M)
perm = list(range(2, weights_rank)) + [1, 0]
if "kernel_shape" in node.attrs.keys():
kernel_shape = node.attrs["kernel_shape"]
assert in_weights.get_shape().as_list()[2:] == kernel_shape, (
"kernel_shape "
"attr of convolution does not match the actual weight "
"passed to this operation, attr {}, actual {}").format(
kernel_shape,
in_weights.get_shape().as_list())
weights = tf.transpose(in_weights, perm)
dilations = node.attrs.get("dilations", [1] * spatial_size)
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", [0, 0] * spatial_size)
if not transpose:
x = PadMixin.get_padding_as_op(x, pads)
group = node.attrs.get("group", 1)
weight_groups = tf.split(weights, num_or_size_splits=group, axis=-1)
if support_cuda:
xs = tf.split(x, num_or_size_splits=group, axis=1)
else:
x = tf.transpose(
x, perm=get_perm_from_formats(storage_format, compute_format))
xs = tf.split(x, num_or_size_splits=group, axis=-1)
if transpose:
if dilations != [1] * spatial_size:
raise RuntimeError("Cannot set non-1 dilation for conv transpose.")
convolved = []
for (x, weight) in zip(xs, weight_groups):
x_spatial_shape = [
x_shape[storage_format.find(d)] for d in spatial_format
]
weights_shape = weights.get_shape().as_list()
# calculate output shape
output_shape = node.attrs.get("output_shape", None)
conv_output_shape = [x_shape[storage_format.find("N")]]
if output_shape is None:
conv_output_shape += [
strides[i] * (x_spatial_shape[i] - 1) + weights_shape[i]
for i in list(range(spatial_size))
]
else:
conv_output_shape += [
s + pads[i] + pads[spatial_size + i]
for i, s in enumerate(output_shape[-2:])
]
conv_output_shape.insert(compute_c_idx, weights_shape[-2])
# make strides to match input rank
strides_full = [1] + strides
strides_full.insert(compute_c_idx, 1)
# get corresponding function in tf
if spatial_size == 1:
conv_func = tf.contrib.nn.conv1d_transpose
strides_full = strides[0]
elif spatial_size == 2:
conv_func = tf.nn.conv2d_transpose
elif spatial_size == 3:
conv_func = tf.nn.conv3d_transpose
else:
raise NotImplementedError(
"Transposed convolution for {}d is not implemented in Tensorflow".
format(spatial_size))
# use raw input x to do transposed conv
conv_rs = conv_func(
x,
weight,
conv_output_shape,
strides_full,
padding="VALID",
data_format=compute_format)
# pad output first by output_padding attr
if "output_padding" in node.attrs and output_shape is None:
output_padding = [[0, 0]
] + [[0, p] for p in node.attrs["output_padding"]]
output_padding.insert(compute_c_idx, [0, 0])
conv_rs = tf.pad(conv_rs, output_padding)
# remove pads set in pads attr
conv_rs_shape = conv_rs.get_shape().as_list()
begin = [0] + pads[:spatial_size]
begin.insert(compute_c_idx, 0)
size = [
s if d in ["N", "C"] else s - pads[spatial_format.find(d)] -
pads[spatial_format.find(d) + spatial_size]
for d, s in zip(compute_format, conv_rs_shape)
]
conv_rs = tf.slice(conv_rs, begin=begin, size=size)
convolved.append(conv_rs)
else:
if group != weights.shape[-1]:
convolved = [
tf.nn.convolution(
x,
weight,
"VALID",
strides=strides,
dilation_rate=dilations,
data_format=compute_format)
for (x, weight) in zip(xs, weight_groups)
]
else:
# convert to depthwise convolutions if num group==channels
convolved = [
tf.nn.depthwise_conv2d(
x,
tf.transpose(weights, [0, 1, 3, 2]), # [filter_height, filter_width, in_channels, multiplier (=1)]
strides=_get_sequence(strides, 2, channel_index=3, name="strides"), # requires a 4-d list
padding="VALID",
rate=dilations, # NOTE I'm not sure if it's a correct. In the newer tensorflow versions there is dilations parameter.
data_format=compute_format,
)
]
if len(node.inputs) == 2:
if support_cuda:
output = tf.concat(convolved, axis=1)
else:
output = tf.concat(convolved, axis=-1)
output = tf.transpose(
output, perm=get_perm_from_formats(compute_format, storage_format))
else:
bias = input_dict[node.inputs[2]]
bias = cls.explicit_broadcast([x, bias], compute_c_idx)
if support_cuda:
output = tf.concat(convolved, axis=1)
output = tf.add(output, bias)
else:
output = tf.concat(convolved, axis=-1)
output = tf.add(output, bias)
output = tf.transpose(
output, perm=get_perm_from_formats(compute_format, storage_format))
return [output]
def conv(cls, node, input_dict, transpose=False):
""" Convolution method for both conv and transposed conv
For transposed conv,
Attr pads is not used for input, but declares how much output is padded.
Here, output means output from transposed conv which already pad output_padding if set.
So the pseudo explanation for output should be:
output = conv_transpose_output + output_padding - pads
And conv_transpose_output shape should be:
conv_transpose_output_shape[i] = strides[i] * (input_shape[i] - 1) + kernel_shape[i]
"""
x = input_dict[node.inputs[0]]
x_rank = len(x.get_shape())
x_shape = x.get_shape().as_list()
spatial_size = x_rank - 2
support_cuda = supports_device("CUDA")
storage_format, compute_format = get_data_format(x_rank)
compute_c_idx = compute_format.find("C")
spatial_format = "".join(
[d for d in compute_format if d not in ["N", "C"]])
in_weights = input_dict[node.inputs[1]]
weights_rank = len(in_weights.get_shape())
if transpose:
# Translate weights from (C x M x KH x KW) to (KH x KW X M X C)
perm = list(range(2, weights_rank)) + [1, 0]
else:
# Translate weights from (M x C x KH x KW) to (KH x KW X C X M)
perm = list(range(2, weights_rank)) + [1, 0]
if "kernel_shape" in node.attrs.keys():
kernel_shape = node.attrs["kernel_shape"]
assert in_weights.get_shape().as_list()[2:] == kernel_shape, (
"kernel_shape "
"attr of convolution does not match the actual weight "
"passed to this operation, attr {}, actual {}").format(
kernel_shape,
in_weights.get_shape().as_list())
weights = tf.transpose(in_weights, perm)
dilations = node.attrs.get("dilations", [1] * spatial_size)
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", [0, 0] * spatial_size)
# Check auto_pad nonexistent or NOTSET first
if "auto_pad" not in node.attrs or node.attrs["auto_pad"] == "NOTSET":
if not transpose:
if pads != [0, 0] * spatial_size:
x = PadMixin.get_padding_as_op(x, pads)
pad_mode = "VALID"
else:
pad_mode = "NOTSET"
# Then we use auto_pad to setup pad_mode
elif node.attrs["auto_pad"] == "SAME_UPPER":
pad_mode = "SAME"
elif node.attrs["auto_pad"] == "VALID":
pad_mode = "VALID"
elif node.attrs["auto_pad"] == "SAME_LOWER":
pad_mode = PAD_TF_INCOMPATIBLE
else:
raise ValueError("Invalid auto_pad attribute: {}".format(
node.attrs["auto_pad"]))
# Currently auto_pad = SAME_LOWER is not supported
if pad_mode is PAD_TF_INCOMPATIBLE:
if transpose:
exception.OP_UNSUPPORTED_EXCEPT(
"ConvTranspose with auto_pad `SAME_LOWER`", "Tensorflow")
else:
exception.OP_UNSUPPORTED_EXCEPT(
"Conv with auto_pad `SAME_LOWER`", "Tensorflow")
group = node.attrs.get("group", 1)
weight_groups = tf.split(weights, num_or_size_splits=group, axis=-1)
if support_cuda:
xs = tf.split(x, num_or_size_splits=group, axis=1)
else:
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, compute_format))
xs = tf.split(x, num_or_size_splits=group, axis=-1)
if transpose:
if dilations != [1] * spatial_size:
raise RuntimeError(
"Cannot set non-1 dilation for conv transpose.")
convolved = []
for (x, weight) in zip(xs, weight_groups):
x_spatial_shape = [
x_shape[storage_format.find(d)] for d in spatial_format
]
weights_shape = weights.get_shape().as_list()
output_shape = node.attrs.get("output_shape", None)
conv_output_shape = [x_shape[storage_format.find("N")]]
# calculate output shape
if pad_mode == "NOTSET":
if output_shape is None:
conv_output_shape += [
strides[i] * x_spatial_shape[i] +
max(weights_shape[i] - strides[i], 0)
for i in list(range(spatial_size))
]
else:
conv_output_shape += [
s + pads[i] + pads[spatial_size + i]
for i, s in enumerate(output_shape[-2:])
]
conv_output_shape.insert(compute_c_idx, weights_shape[-2])
# make strides to match input rank
strides_full = [1] + strides
strides_full.insert(compute_c_idx, 1)
# get corresponding function in tf
if spatial_size == 1:
conv_func = tf.nn.conv1d_transpose
strides_full = strides[0]
elif spatial_size == 2:
conv_func = tf.nn.conv2d_transpose
elif spatial_size == 3:
conv_func = tf.nn.conv3d_transpose
else:
raise NotImplementedError(
"Transposed convolution for {}d is not implemented in Tensorflow"
.format(spatial_size))
# use raw input x to do transposed conv
conv_rs = conv_func(x,
weight,
conv_output_shape,
strides_full,
padding="VALID",
data_format=compute_format)
# pad output first by output_padding attr
if "output_padding" in node.attrs and output_shape is None:
output_padding = [[
0, 0
]] + [[0, p] for p in node.attrs["output_padding"]]
output_padding.insert(compute_c_idx, [0, 0])
conv_rs = tf.pad(conv_rs, output_padding)
# remove pads set in pads attr
conv_rs_shape = conv_rs.get_shape().as_list()
begin = [0] + pads[:spatial_size]
begin.insert(compute_c_idx, 0)
size = [
s if d in ["N", "C"] else s -
pads[spatial_format.find(d)] -
pads[spatial_format.find(d) + spatial_size]
for d, s in zip(compute_format, conv_rs_shape)
]
conv_rs = tf.slice(conv_rs, begin=begin, size=size)
convolved.append(conv_rs)
else:
# No need to check pads if auto_pad is specifically provided.
# The assumption is that once auto_pad is provided as either VALID
# or SAME_UPPER (SAME_LOWER is currently not supported in TF) the
# output_shape will always be inferred. That is, the output_shape
# and output_padding will not be used in this case.
if pad_mode == "VALID":
conv_output_shape += [
strides[i] * (x_spatial_shape[i] - 1) +
weights_shape[i] for i in list(range(spatial_size))
]
else:
conv_output_shape += [
strides[i] * x_spatial_shape[i]
for i in list(range(spatial_size))
]
conv_output_shape.insert(compute_c_idx, weights_shape[-2])
# make strides to match input rank
strides_full = [1] + strides
strides_full.insert(compute_c_idx, 1)
# get corresponding function in tf
if spatial_size == 1:
conv_func = tf.contrib.nn.conv1d_transpose
strides_full = strides[0]
elif spatial_size == 2:
conv_func = tf.nn.conv2d_transpose
elif spatial_size == 3:
conv_func = tf.nn.conv3d_transpose
else:
raise NotImplementedError(
"Transposed convolution for {}d is not implemented in Tensorflow"
.format(spatial_size))
# use raw input x to do transposed conv
conv_rs = conv_func(x,
weight,
conv_output_shape,
strides_full,
padding=pad_mode,
data_format=compute_format)
convolved.append(conv_rs)
else:
convolved = [
tf.nn.convolution(x,
weight,
padding=pad_mode,
strides=strides,
dilations=dilations,
data_format=compute_format)
for (x, weight) in zip(xs, weight_groups)
]
if len(node.inputs) == 2:
if support_cuda:
output = tf.concat(convolved, axis=1)
else:
output = tf.concat(convolved, axis=-1)
output = tf.transpose(output,
perm=get_perm_from_formats(
compute_format, storage_format))
else:
bias = input_dict[node.inputs[2]]
bias = cls.explicit_broadcast([x, bias], compute_c_idx)
if support_cuda:
output = tf.concat(convolved, axis=1)
output = tf.add(output, bias)
else:
output = tf.concat(convolved, axis=-1)
output = tf.add(output, bias)
output = tf.transpose(output,
perm=get_perm_from_formats(
compute_format, storage_format))
return [output]
def pool(cls, node, input_dict, pool_func, pooling_type, strict=True):
x = input_dict[node.inputs[0]]
x_rank = len(x.get_shape())
x_shape = x.get_shape().as_list()
spatial_size = x_rank - 2
support_cuda = supports_device("CUDA")
storage_format, compute_format = get_data_format(x_rank)
kernel_shape = node.attrs["kernel_shape"]
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", None)
pad = PAD_TF_INCOMPATIBLE
# from version 7
count_include_pad = node.attrs.get("count_include_pad", 0)
# If padding is specified, try to recover it from explicit padding
# specification to tensorflow padding mode:
if pads is not None:
pad = cls._get_tf_pad(x_shape[2:], kernel_shape, strides, pads)
else:
# Neither pad nor auto_pad is specified, assume no padding.
if "auto_pad" not in node.attrs:
pad = "VALID"
# We consult auto_pad if pad is not specified and auto_pad
# is available.
else:
if node.attrs["auto_pad"] == "SAME_UPPER":
pad = "SAME"
elif node.attrs["auto_pad"] == "VALID":
pad = "VALID"
elif node.attrs["auto_pad"] == "SAME_LOWER":
pad = PAD_TF_INCOMPATIBLE
if count_include_pad == 1:
_, pads = cls._pool_get_shapes(node.attrs["auto_pad"],
x_shape[2:], kernel_shape,
strides,
[0] * spatial_size * 2)
if strict and count_include_pad == 0:
if pad is PAD_TF_INCOMPATIBLE:
return cls._compatibility_pool(node, input_dict, pooling_type)
else:
if pads != [0] * spatial_size * 2:
x = PadMixin.get_padding_as_op(x, pads)
pad = "VALID"
if support_cuda:
pooled = pool_func(x,
kernel_shape,
padding=pad,
strides=strides,
data_format=compute_format)
else:
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, compute_format))
pooled = pool_func(x,
kernel_shape,
padding=pad,
strides=strides,
data_format=compute_format)
pooled = tf.transpose(pooled,
perm=get_perm_from_formats(
compute_format, storage_format))
return [pooled]
def pool(cls, node, input_dict, pooling_type, strict=True):
x = input_dict[node.inputs[0]]
orig_x = x
kernel_shape = node.attrs["kernel_shape"]
spatial_size = len(kernel_shape)
x_rank = spatial_size + 2
kernel_shape = node.attrs["kernel_shape"]
strides = node.attrs.get("strides", [1] * spatial_size)
dilations = node.attrs.get("dilations", [1] * spatial_size)
ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pads = node.attrs.get("auto_pad", "NOTSET")
p = node.attrs.get("p", 2)
if pads == "NOTSET":
pads = node.attrs.get("pads", [0] * spatial_size * 2)
# In case shape is fully defined, check if pads match
# SAME padding in Tensorflow
if x.shape.is_fully_defined() and pads != [0] * spatial_size * 2:
in_shape = x.get_shape().as_list()
same_paddings = calc_pads_same(in_shape[1:x_rank - 1],
kernel_shape, strides,
dilations, "SAME_UPPER")
if pads == same_paddings:
pads = "SAME_UPPER"
count_include_pad = bool(node.attrs.get("count_include_pad", 0))
if pooling_type == "AVG":
pooling_name = "AveragePool"
elif pooling_type == "MAX":
pooling_name = "MaxPool"
elif pooling_type == "MAX_WITH_ARGMAX":
pooling_name = "MaxPoolWithArgmax"
elif pooling_type == "LP":
pooling_name = "LpPool"
if spatial_size > 3:
exception.OP_UNSUPPORTED_EXCEPT(
pooling_name + " with {}D input".format(x_rank), "Tensorflow")
if pooling_type == "MAX_WITH_ARGMAX" and x_rank != 4:
exception.OP_UNSUPPORTED_EXCEPT(
pooling_name + " with {}D input".format(x_rank), "Tensorflow")
if node.attrs.get("storage_order", 0) != 0:
exception.OP_UNSUPPORTED_EXCEPT(
pooling_name + " with column major", "Tensorflow")
storage_format, _ = get_data_format(x_rank)
need_trans = storage_format.startswith("NC")
if need_trans:
compute_format = "N" + storage_format[2:] + "C"
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, compute_format))
dp = DilatedPooling(input=x,
kernel_shape=kernel_shape,
strides=strides,
dilations=dilations,
padding=pads,
ceil_mode=ceil_mode,
pooling_type=pooling_type,
count_include_pad=count_include_pad,
p=p)
if not dp.is_supported():
if strict:
logger.warning(
"Using the pooling op in compatibility mode. "
"This means your graph cannot be serialized.", UserWarning)
result = tf.py_func(py_pool, [
orig_x, kernel_shape, strides, dilations, pads, ceil_mode,
pooling_type, False
], orig_x.dtype)
if orig_x.shape.is_fully_defined():
shape = orig_x.get_shape().as_list()
output_shape = shape[0:2] + calc_output_shape(
shape[2:x_rank], kernel_shape, strides, dilations,
pads, ceil_mode)
else:
output_shape = [None] * x_rank
result.set_shape(output_shape)
return [result]
else:
exception.OP_UNSUPPORTED_EXCEPT(
"strict == 0 and " + pooling_name +
" arguments not compatible", "Tensorflow")
def dilated_pool():
return (dp.dilated_pool(), None)
# select correct op depending on the pooling type
pooling_op = dilated_pool if pooling_type in ["MAX", "AVG", "LP"] else \
dp.dilated_maxpool_with_argmax
# select the correct transpose ops depending on the input storage format
perm = get_perm_from_formats(compute_format, storage_format)
def postprocess(pooled, argmax):
return (tf.transpose(pooled, perm=perm) if need_trans else pooled,
tf.transpose(argmax, perm=perm)
if need_trans and argmax is not None else argmax)
pooled, argmax = pooling_op()
pooled, argmax = postprocess(pooled, argmax)
result = [pooled] if argmax is None else [pooled, argmax]
return result
def pool(cls, node, input_dict, pool_func, pooling_type, strict=True):
x = input_dict[node.inputs[0]]
x_rank = len(x.get_shape())
x_shape = x.get_shape().as_list()
spatial_size = x_rank - 2
if spatial_size > 3:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPool with {}D input".format(x_rank), "Tensorflow")
support_cuda = supports_device("CUDA")
storage_format, compute_format = get_data_format(x_rank)
kernel_shape = node.attrs["kernel_shape"]
strides = node.attrs.get("strides", [1] * spatial_size)
pads = node.attrs.get("pads", None)
pad = PAD_TF_INCOMPATIBLE
# from version 7
count_include_pad = node.attrs.get("count_include_pad", 0)
auto_pad = node.attrs.get("auto_pad", "NOTSET")
# if auto_pad is NOTSET, we check pads
if auto_pad == "NOTSET":
# If padding is specified, try to recover it from explicit padding
# specification to tensorflow padding mode:
if pads is not None:
pad = cls._get_tf_pad(x_shape[2:], kernel_shape, strides, pads)
else:
pad = "VALID"
else:
if auto_pad == "SAME_UPPER":
pad = "SAME"
elif auto_pad == "VALID":
pad = "VALID"
elif auto_pad == "SAME_LOWER":
pad = PAD_TF_INCOMPATIBLE
if count_include_pad == 1:
_, pads = cls._pool_get_shapes(auto_pad, x_shape[2:],
kernel_shape, strides,
[0] * spatial_size * 2)
if pooling_type in ("AVG", "MAX"):
if strict and count_include_pad == 0:
if pad is PAD_TF_INCOMPATIBLE:
return cls._compatibility_pool(node, input_dict,
pooling_type)
else:
if pads != [0] * spatial_size * 2:
x = PadMixin.get_padding_as_op(x, pads)
pad = "VALID"
elif pooling_type == "MAX_WITH_ARGMAX":
if pad is PAD_TF_INCOMPATIBLE:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPoolWithArgmax with pad is None or incompatible mode",
"Tensorflow")
if x_rank != 4:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPoolWithArgmax with {}D input".format(x_rank),
"Tensorflow")
if node.attrs.get("storage_order", 0) != 0:
exception.OP_UNSUPPORTED_EXCEPT(
"MaxPoolWithArgmax with column major", "Tensorflow")
need_trans = storage_format != "NHWC"
if need_trans:
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, "NHWC"))
pooled, argmax = pool_func(x, [1] + kernel_shape + [1],
padding=pad,
strides=[1] + strides + [1])
if need_trans:
pooled = tf.transpose(pooled,
perm=get_perm_from_formats(
"NHWC", storage_format))
argmax = tf.transpose(argmax,
perm=get_perm_from_formats(
"NHWC", storage_format))
return [pooled, argmax]
if support_cuda:
pooled = pool_func(x,
kernel_shape,
padding=pad,
strides=strides,
data_format=compute_format)
else:
x = tf.transpose(x,
perm=get_perm_from_formats(
storage_format, compute_format))
pooled = pool_func(x,
kernel_shape,
padding=pad,
strides=strides,
data_format=compute_format)
pooled = tf.transpose(pooled,
perm=get_perm_from_formats(
compute_format, storage_format))
return [pooled]
def pool_v11(cls, node, input_dict, pooling_type, strict=True):
x = input_dict[node.inputs[0]]
orig_x = x
kernel_shape = node.attrs["kernel_shape"]
spatial_size = len(kernel_shape)
x_rank = spatial_size + 2
kernel_shape = node.attrs["kernel_shape"]
strides = node.attrs.get("strides", [1] * spatial_size)
dilations = node.attrs.get("dilations", [1] * spatial_size)
ceil_mode = bool(node.attrs.get("ceil_mode", 0))
pads = node.attrs.get("auto_pad", "NOTSET")
if pads == "NOTSET":
pads = node.attrs.get("pads", [0] * spatial_size * 2)
count_include_pad = bool(node.attrs.get("count_include_pad", 0))
if pooling_type == "AVG":
pooling_name = "AveragePool"
elif pooling_type == "MAX":
pooling_name = "MaxPool"
elif pooling_type == "MAX_WITH_ARGMAX":
pooling_name = "MaxPoolWithArgmax"
if spatial_size > 3:
exception.OP_UNSUPPORTED_EXCEPT(
pooling_name + " with {}D input".format(x_rank), "Tensorflow")
if pooling_type == "MAX_WITH_ARGMAX" and x_rank != 4:
exception.OP_UNSUPPORTED_EXCEPT(
pooling_name + " with {}D input".format(x_rank), "Tensorflow")
if node.attrs.get("storage_order", 0) != 0:
exception.OP_UNSUPPORTED_EXCEPT(pooling_name + " with column major",
"Tensorflow")
storage_format, _ = get_data_format(x_rank)
need_trans = storage_format.startswith("NC")
if need_trans:
compute_format = "N" + storage_format[2:] + "C"
x = tf.transpose(x, perm=get_perm_from_formats(storage_format,
compute_format))
dp = DilatedPooling(input=x, kernel_shape=kernel_shape, strides=strides,
dilations=dilations, padding=pads, ceil_mode=ceil_mode,
pooling_type=pooling_type,
count_include_pad=count_include_pad)
if not dp.is_supported():
if strict:
warnings.warn(
"Using the pooling op in compatibility mode. "
"This means your graph cannot be serialized.", UserWarning)
return [tf.py_func(py_pool, [orig_x, kernel_shape, strides,
dilations, pads, ceil_mode, "AVG",
False], orig_x.dtype)]
else:
exception.OP_UNSUPPORTED_EXCEPT("strict == 0 and average pool"
" arguments not compatible",
"Tensorflow")
def dilated_pool():
return (dp.dilated_pool(), None)
# select correct op depending on the pooling type
pooling_op = dilated_pool if pooling_type in ["MAX", "AVG"] else \
dp.dilated_maxpool_with_argmax
# select the correct transpose ops depending on the input storage format
perm = get_perm_from_formats(compute_format, storage_format)
def postprocess(pooled, argmax):
return (tf.transpose(pooled, perm=perm) if need_trans else pooled,
tf.transpose(argmax, perm=perm) if need_trans and argmax
is not None else argmax)
pooled, argmax = pooling_op()
pooled, argmax = postprocess(pooled, argmax)
result = [pooled] if argmax is None else [pooled, argmax]
return result