def _conv2d(input,
filter,
bias=False,
strides=[1, 1],
pads=[1, 1, 1, 1],
dilations=[1, 1],
group=1,
debugContext=''):
"""Encapsulation of function get_builder().aiOnnx.conv!
args:
x: input tensor
ksize: int,kernel size
stride: int,stride of conv
pads: int, conv padding
c_out: int, output channel
group: int, conv group nums,default:1
"""
args = [input.getIpuIndex(), filter.getIpuIndex()]
if bias:
args.append(bias.getIpuIndex())
output = get_builder().aiOnnx.conv(args,
strides=strides,
pads=pads,
dilations=dilations,
group=group,
debugContext=debugContext)
if get_memory_proportion() is not None:
get_builder().setAvailableMemoryProportion(output,
get_memory_proportion())
return TTensor(output)
def clip(x, minmun=-np.inf, maxmun=np.inf, debugContext=""):
if get_ai_onnx_version() >= 11:
minmun = constant(np.asarray(minmun).astype(np.float32))
maxmun = constant(np.asarray(maxmun).astype(np.float32))
return TTensor(get_builder().aiOnnx.clip(
[x.getIpuIndex(),
minmun.getIpuIndex(),
maxmun.getIpuIndex()],
debugContext=debugContext))
else:
return TTensor(get_builder().aiOnnx.clip([x.getIpuIndex()],
maxmun.getIpuIndex(),
minmun.getIpuIndex(),
debugContext=debugContext))
def _batchNorm(
x,
scale,
biases,
mean,
var,
num_outputs=1,
momentum=0.9,
epsilon=1e-5,
debugContext="",
):
results = get_builder().aiOnnx.batchnormalization(
[
x.getIpuIndex(),
scale.getIpuIndex(),
biases.getIpuIndex(),
mean.getIpuIndex(),
var.getIpuIndex()
],
num_outputs=num_outputs,
epsilon=epsilon,
momentum=momentum,
debugContext=debugContext)
results = results[0] if num_outputs == 1 else results
if isinstance(results, list):
results = [TTensor(r) for r in results]
else:
results = [TTensor(results)]
return results
def softmax_2d(x, axis=1, debugContext=""):
assert axis in [-1, 1]
assert x.shape.ndims == 2
x = get_builder().aiOnnx.softmax(
[x.getIpuIndex()], axis=axis,
debugContext=debugContext)
return TTensor(x)
def reshape(source, target_shape, debugContext=""):
"""
args:
source : tensor name
target_shape: list of int e.g.: [3,4,5,6]
"""
if isinstance(target_shape, TTensor):
target_shape = target_shape.data
if isinstance(target_shape, np.ndarray):
target_shape = target_shape.tolist()
if isinstance(target_shape, list):
target_shape = [scalarTensor2int(ele) for ele in target_shape]
target_shape = constant(np.array(target_shape).astype(np.int64),
debugContext=debugContext)
if check_all_constant([source, target_shape]):
# degrade to np op
result = source.data.reshape(target_shape.data)
result = constant(result)
return result
else:
return TTensor(get_builder().aiOnnx.reshape(
[source.getIpuIndex(),
target_shape.getIpuIndex()],
debugContext=debugContext))
def one_hot(indices, depth, values=None, debugContext=''):
'''
values: [off_value, on_value]
if indice is -1, the corrosponding arr is [0]*depth
'''
if isinstance(depth, int):
depth = to_tensor(depth, dtype='INT64')
if values is None:
values = constant(np.asarray([0, 1]).astype(np.int32),
debugContext=debugContext)
assert indices.dtype in [
'INT32', 'INT64', 'INT16', 'UINT32', 'UINT64', 'UINT16'
]
assert depth.dtype in [
'INT32', 'INT64', 'INT16', 'UINT32', 'UINT64', 'UINT16'
]
assert values.dtype in [
'INT32', 'INT64', 'INT16', 'UINT32', 'UINT64', 'UINT16'
]
result = get_builder().aiOnnx.onehot(
[indices.getIpuIndex(),
depth.getIpuIndex(),
values.getIpuIndex()],
debugContext=debugContext)
result = TTensor(result)
result_shape = list(result.pureShape)
if result_shape[1] == 0:
result_shape[1] = depth
result = result.reshape(result_shape)
return result
def tile(input, repeats, debugContext=""):
if check_all_constant([input, repeats]):
result = np.tile(input.data, repeats.data)
return constant(result)
result = get_builder().aiOnnx.tile(
[input.getIpuIndex(), repeats.getIpuIndex()], debugContext)
return TTensor(result)
def real_init(self, ):
assert not self.initialized
self.data = np.ascontiguousarray(self.data.copy())
name = get_builder().aiOnnx.constant(self.data)
super().__init__(name)
self.__name = name # private attribute can not be inherited
self.initialized = True
def flatten(x):
'''implements the np.flatten function
'''
if check_all_constant([x]):
x = x.data.flatten()
return constant(x)
x = get_builder().aiOnnx.flatten([x.getIpuIndex()], 0)
return TTensor(x).squeeze(0)
def min(tensor_list, debugContext=""):
if check_all_constant(tensor_list):
# degrade to np op
arr_list = [t.data for t in tensor_list]
result = np.min(arr_list)
return constant(result)
return TTensor(get_builder().aiOnnx.min(
[t.getIpuIndex() for t in tensor_list], debugContext=debugContext))
def reduceMean(x, axes, keepdims=False, debugContext=''):
if isinstance(axes, int):
axes = [axes]
assert isinstance(axes, list) or axes is None
return TTensor(get_builder().aiOnnx.reducemean([x.getIpuIndex()],
axes=axes,
keepdims=keepdims,
debugContext=debugContext))
def sub(tensors, debugContext=""):
assert len(tensors) == 2
if check_all_constant(tensors):
# degrade to np op
result = tensors[0].data - tensors[1].data
return constant(result, debugContext=debugContext)
return TTensor(get_builder().aiOnnx.sub([t.getIpuIndex() for t in tensors],
debugContext=debugContext))
def add(tensors, debugContext=""):
if check_all_constant(tensors):
# degrade to np op
result = 0
for t in tensors:
result = result + t.data
return constant(result, debugContext=debugContext)
return TTensor(get_builder().aiOnnx.add([t.getIpuIndex() for t in tensors],
debugContext=debugContext))
def transpose(x, dim_order, debugContext=""):
"""dim_order: list of int. eg:[0,2,3,1]"""
if check_all_constant([x]):
# degrade to np op
result = np.transpose(x.data, dim_order)
return constant(result)
return TTensor(get_builder().aiOnnx.transpose([x.getIpuIndex()],
dim_order,
debugContext=debugContext))
def mul(tensors, debugContext=""):
if check_all_constant(tensors):
# degrade to np op
result = 1
for t in tensors:
result = t.data * result
return constant(result, debugContext=debugContext)
return TTensor(get_builder().aiOnnx.mul([t.getIpuIndex() for t in tensors],
debugContext=debugContext))
def matmul(x, y, debugContext=""):
if check_all_constant([x, y]):
# degrade to np op
result = np.matmul(x.data, y.data)
return constant(result)
else:
assert x.dtype in ['FLOAT', "FLOAT16"]
assert y.dtype in ['FLOAT', "FLOAT16"]
return TTensor(get_builder().aiOnnx.matmul(
[x.getIpuIndex(), y.getIpuIndex()], debugContext=debugContext))
def _concat(tensor_list, dim, debugContext=""):
if check_all_constant(tensor_list):
# degrade to np op
np_arr_list = [t.data for t in tensor_list]
result = np.concatenate(np_arr_list, axis=dim)
return constant(result)
return TTensor(get_builder().aiOnnx.concat(
[tensor.getIpuIndex() for tensor in tensor_list],
dim,
debugContext=debugContext))
def pad(data, pads, mode='constant', constant_value=0, debugContext=''):
constant_value = constant(constant_value).cast(data.dtype.upper())
pads = to_tensor(pads).cast('INT64').flatten()
result = get_builder().aiOnnx.pad(
[data.getIpuIndex(),
pads.getIpuIndex(),
constant_value.getIpuIndex()],
mode=mode,
debugContext=debugContext)
return TTensor(result)
def reduceprod(x, dim, keepdims=False, debugContext=""):
"""
args:
dim: int .which dim to do prod
"""
x = get_builder().aiOnnx.reduceprod([x.getIpuIndex()],
axes=[dim],
keepdims=keepdims,
debugContext=debugContext)
return TTensor(x)
def relu(x, debugContext=""):
"""
args:
x: input tensor
"""
if isinstance(x, list):
x = [ele.getIpuIndex() for ele in x]
else:
x = [x.getIpuIndex()]
x = get_builder().aiOnnx.relu(x, debugContext=debugContext)
return TTensor(x)
def split(x, lenOfSplit, dim, debugContext=""):
"""
args:
lenOfSplit: (4,1) split into two piece，one's length is 4 ,
the other is 1
"""
return TTensor(get_builder().aiOnnx.split([x.getIpuIndex()],
len(lenOfSplit),
dim,
lenOfSplit,
debugContext=debugContext))
def addInitializedInputTensor(array, debugContext=""):
"""
args:
array: an numpy array that will be copy to IPU
return:
str: tensor name
"""
name = get_builder().addInitializedInputTensor(array,
debugContext=debugContext)
return TTensor(name)
def nllloss(prob,
label,
reductionType=popart.ReductionType.Mean,
debugPrefix=''):
# prob: scaled probabilities, [batch, classes], float
# label: labels, [batch,], int32
with name_scope(debugPrefix):
loss = get_builder().aiGraphcore.nllloss(
[prob.getIpuIndex(), label.getIpuIndex()],
reductionType,
debugContext=debugPrefix)
return TTensor(loss)
def cast(x, target_type='FLOAT', debugContext=''):
"""
target_type:
FLOAT|FLOAT16|INT8|INT16|INT32|UINT8|UINT16|UINT32|BOOL
"""
target_type = 'FLOAT' if target_type == 'FLOAT32' else target_type
if check_all_constant([x]):
# degrade to np op
data = x.data.astype(mappin_gc2npy[target_type])
return constant(data)
else:
return TTensor(get_builder().aiOnnx.cast([x.getIpuIndex()],
target_type.upper(),
debugContext))
def unsqueeze(x, dims, debugContext=""):
"""
args:
dim: list of int of which dim will delete
eg:[3] or [1,3]
"""
if check_all_constant([x]):
# degrade to np op
result = np.expand_dims(x.data, axis=dims)
return constant(result)
x = get_builder().aiOnnx.unsqueeze([x.getIpuIndex()],
axes=dims,
debugContext=debugContext)
return TTensor(x)
def avgPooling(x,
strides=2,
kernel_size=2,
padding=0,
count_include_pad=0,
debugContext=""):
x = get_builder().aiOnnx.averagepool(
[x.getIpuIndex()],
kernel_shape=[kernel_size, kernel_size],
count_include_pad=count_include_pad,
pads=[padding] * 4,
strides=[strides, strides],
debugContext=debugContext)
return TTensor(x)
def topk(x, k, sorted=True, dim=-1, debugContext=""):
"""
args:
k: the count of return
dim: in which dim to sort and clip
"""
if k.shape.ndims == 0:
k = k.unsqueeze(0)
else:
assert k.shape.ndims == 1
values, order = get_builder().aiOnnx.topk(
[x.getIpuIndex(), k.getIpuIndex()],
axis=dim,
sorted=sorted,
debugContext=debugContext)
return TTensor(values), TTensor(order)
def resize(x,
roi=None,
scales=None,
sizes=None,
coordinate_transformation_mode='half_pixel',
cubic_coeff_a=-0.75,
exclude_outside=0,
extrapolation_value=0.0,
mode='nearest',
nearest_mode='round_prefer_floor',
debugContext=''):
# TODO Check whether each parameter is correct
# x:N-D tensor
# roi: 1-D tensor given as [start1, ..., startN, end1, ..., endN], where N is the rank of X
# scales: tensor(float),The scale array along each dimension.
# sizes: tensor(int64),The size of the output tensor.
# Only one of 'scales' and 'sizes' can be specified.
assert None in [scales, sizes] and set([scales, sizes]) == 2
if roi is None:
assert coordinate_transformation_mode == 'tf_crop_and_resize'
else:
raise not NotImplementedError
roi = constant(
np.array([0, -1] * x.shape.ndims).astype(
mappin_gc2npy[x.dtype])) if roi is None else roi
scales = constant(np.array(
[1.0] * x.shape.ndims).astype('FLOAT32')) if scales is None else scales
sizes = constant(np.array(
[1] * x.shape.ndims).astype('INT64')) if sizes is None else sizes
inputs_list = [
x.getIpuIndex(),
roi.getIpuIndex(),
scales.getIpuIndex(),
sizes.getIpuIndex()
]
inputs_dic = {
'coordinate_transformation_mode': coordinate_transformation_mode,
'cubic_coeff_a': cubic_coeff_a,
'exclude_outside': exclude_outside,
'extrapolation_value': extrapolation_value,
'mode': mode,
'nearest_mode': nearest_mode,
'debugContext': debugContext
}
result = TTensor(get_builder().aiOnnx.resize(inputs_list, **inputs_dic))
return result
def gc_slice(x, axes, starts, ends, debugContext=""):
if check_all_constant([x, axes, starts, ends]):
# degrade to np op
x = x.data
x_slices = []
for start, end in zip(starts.data.tolist(), ends.data.tolist()):
x_slices.append(slice(start, end))
return constant(x[x_slices])
else:
x = get_builder().aiOnnx.slice([
x.getIpuIndex(),
starts.getIpuIndex(),
ends.getIpuIndex(),
axes.getIpuIndex()
],
debugContext=debugContext)
return TTensor(x)
def squeeze(x, dims, debugContext=""):
if check_all_constant([x]):
# degrade to np op
x = x.data
current_dim = float('inf')
for dim in reversed(dims):
assert current_dim > dim
current_dim = dim
x = x.squeeze(dim)
return constant(x)
if isinstance(dims, int):
dims = [dims]
for dim in dims:
assert x.pureShape[dim] == 1
x = get_builder().aiOnnx.squeeze([x.getIpuIndex()],
axes=dims,
debugContext=debugContext)
return TTensor(x)