def retrieve_tensor_value(self,
name,
detail,
shape,
graph_name=None,
prev=False):
"""Retrieve the tensor value."""
log.info("Retrieve tensor value: name: %s, detail: %s, shape: %s",
name, detail, shape)
self.validate_tensor_param(name, detail)
# Limit to query max two dimensions for tensor in table view.
parsed_shape = TensorUtils.parse_shape(shape,
limit=MAX_DIMENSIONS_FOR_TENSOR)
node_type, tensor_name = self._get_tensor_name_and_type_by_ui_name(
name, graph_name)
reply = self.cache_store.get_stream_handler(Streams.TENSOR).get({
'name':
tensor_name,
'node_type':
node_type,
'shape':
parsed_shape,
'prev':
prev
})
reply['tensor_value']['name'] = name
return reply
def tensor_comparisons(self,
name,
shape,
detail='data',
tolerance='0',
rank_id=0,
graph_name=None):
"""
Get tensor comparisons data for given name, detail, shape and tolerance.
Args:
name (str): The name of tensor for ui.
shape (str): Specify concrete dimensions of shape.
detail (str): Specify which data to query. Current available value is 'data' which means
concrete tensor data. Histogram or unique count can be supported in the future.
rank_id (int): The id of rank. Default: 0.
tolerance (str): Specify tolerance of difference between current step tensor and previous
step tensor. Default value is 0.
graph_name (str): The graph name. Default: None.
Returns:
dict, the retrieved data.
"""
if self.cache_store.get_stream_handler(
Streams.METADATA).state != ServerStatus.WAITING.value:
log.error(
"Failed to compare tensors as the MindSpore is not in waiting state."
)
raise DebuggerCompareTensorError(
"Failed to compare tensors as the MindSpore is not in waiting state."
)
self.validate_tensor_param(name, detail)
# Limit to query max two dimensions for tensor in table view.
parsed_shape = TensorUtils.parse_shape(shape,
limit=MAX_DIMENSIONS_FOR_TENSOR)
node_type, tensor_name, graph_name = self._get_tensor_name_and_type_by_ui_name(
name, graph_name, rank_id)
tolerance = to_float(tolerance, 'tolerance')
tensor_stream = self.cache_store.get_stream_handler(
Streams.TENSOR).get_tensor_handler_by_rank_id(rank_id)
cur_step = self.cache_store.get_stream_handler(Streams.METADATA).step
if node_type == NodeTypeEnum.PARAMETER.value:
reply = tensor_stream.get_tensors_diff(tensor_name, parsed_shape,
tolerance, cur_step)
else:
raise DebuggerParamValueError(
"The node type must be parameter, but got {}.".format(
node_type))
if reply.pop('view_cmd', False):
self._send_view_cmd(name, graph_name, rank_id, tensor_name,
node_type)
return reply
def _get_train_tensors(self, train_id, tags, step, dims, detail):
"""
Get tensor data for given train_id, tags, step, dims and detail.
Args:
train_id (str): Specify list of train job ID.
tags (list): Specify list of tag.
step (int): Specify step of tensor, it's necessary when detail is set to 'data'.
dims (str): Specify dims of tensor, it's necessary when detail is set to 'data'.
detail (str): Specify which data to query, available values: 'stats', 'histogram' and 'data'.
Returns:
list[dict], a list of dictionaries containing the `train_id`, `tag`, `values`.
Raises:
TensorNotExistError, If tensor with specific train_id and tag is not exist in cache.
ParamValueError, If the value of detail is not within available values:
'stats', 'histogram' and 'data'.
"""
tensors_response = []
for tag in tags:
try:
tensors = self._data_manager.list_tensors(train_id, tag)
except ParamValueError as err:
raise TensorNotExistError(err.message)
if tensors and not isinstance(tensors[0].value, TensorContainer):
raise TensorNotExistError(
"there is no tensor data in this tag: {}".format(tag))
if detail is None or detail == 'stats':
values = self._get_tensors_summary(detail, tensors)
elif detail == 'data':
Validation.check_param_empty(step=step, dims=dims)
# Limit to query max two dimensions for tensor in table view.
dims = TensorUtils.parse_shape(dims,
limit=MAX_DIMENSIONS_FOR_TENSOR)
step = to_int(step, "step")
values = self._get_tensors_data(step, dims, tensors)
elif detail == 'histogram':
values = self._get_tensors_histogram(tensors)
else:
raise ParamValueError(
'Can not support this value: {} of detail.'.format(detail))
tensor = {"train_id": train_id, "tag": tag, "values": values}
tensors_response.append(tensor)
return tensors_response