def testTranslateNumericDTypes(self):
x = np.zeros([2, 2], dtype=np.float32)
self.assertEqual('float32', tensor_helper.translate_dtype(x.dtype))
x = np.zeros([2], dtype=np.int16)
self.assertEqual('int16', tensor_helper.translate_dtype(x.dtype))
x = np.zeros([], dtype=np.uint8)
self.assertEqual('uint8', tensor_helper.translate_dtype(x.dtype))
def query(self, time_indices):
"""Query the values at given time indices.
Args:
time_indices: 0-based time indices to query, as a `list` of `int`.
Returns:
Values as a list of `numpy.ndarray` (for time indices in memory) or
`None` (for time indices discarded).
"""
if self._disposed:
raise ValueError(
'Cannot query: this _WatchStore instance is already disposed')
if not isinstance(time_indices, (tuple, list)):
time_indices = [time_indices]
output = []
for time_index in time_indices:
if isinstance(self._data[time_index], _TensorValueDiscarded):
output.append(None)
else:
data_item = self._data[time_index]
if (hasattr(data_item, 'dtype') and
tensor_helper.translate_dtype(data_item.dtype) == 'string'):
_, _, data_item = tensor_helper.array_view(data_item)
data_item = np.array(
tensor_helper.process_buffers_for_display(data_item),
dtype=np.object)
output.append(data_item)
return output
def _comm_tensor_data(device_name, node_name, maybe_base_expanded_node_name,
output_slot, debug_op, tensor_value, wall_time):
"""Create a dict() as the outgoing data in the tensor data comm route.
Note: The tensor data in the comm route does not include the value of the
tensor in its entirety in general. Only if a tensor satisfies the following
conditions will its entire value be included in the return value of this
method:
1. Has a numeric data type (e.g., float32, int32) and has fewer than 5
elements.
2. Is a string tensor and has fewer than 5 elements. Each string element is
up to 40 bytes.
Args:
device_name: Name of the device that the tensor is on.
node_name: (Original) name of the node that produces the tensor.
maybe_base_expanded_node_name: Possbily base-expanded node name.
output_slot: Output slot number.
debug_op: Name of the debug op.
tensor_value: Value of the tensor, as a numpy.ndarray.
wall_time: Wall timestamp for the tensor.
Returns:
A dict representing the tensor data.
"""
output_slot = int(output_slot)
tf.logging.info('Recording tensor value: %s, %d, %s', node_name,
output_slot, debug_op)
tensor_values = None
if isinstance(tensor_value, debug_data.InconvertibleTensorProto):
if not tensor_value.initialized:
tensor_dtype = UNINITIALIZED_TAG
tensor_shape = UNINITIALIZED_TAG
else:
tensor_dtype = UNSUPPORTED_TAG
tensor_shape = UNSUPPORTED_TAG
tensor_values = NA_TAG
else:
tensor_dtype = tensor_helper.translate_dtype(tensor_value.dtype)
tensor_shape = tensor_value.shape
# The /comm endpoint should respond with tensor values only if the tensor is
# small enough. Otherwise, the detailed values sould be queried through a
# dedicated tensor_data that supports slicing.
if tensor_helper.numel(tensor_shape) < 5:
_, _, tensor_values = tensor_helper.array_view(tensor_value)
if tensor_dtype == 'string' and tensor_value is not None:
tensor_values = tensor_helper.process_buffers_for_display(
tensor_values, limit=STRING_ELEMENT_MAX_LEN)
return {
'type': 'tensor',
'timestamp': wall_time,
'data': {
'device_name': device_name,
'node_name': node_name,
'maybe_base_expanded_node_name': maybe_base_expanded_node_name,
'output_slot': output_slot,
'debug_op': debug_op,
'dtype': tensor_dtype,
'shape': tensor_shape,
'values': tensor_values,
},
}
def testTranslateStringDType(self):
x = np.array(['abc'], dtype=np.object)
self.assertEqual('string', tensor_helper.translate_dtype(x.dtype))
def testTranslateBooleanDType(self):
x = np.zeros([2, 2], dtype=np.bool)
self.assertEqual('bool', tensor_helper.translate_dtype(x.dtype))
def _comm_tensor_data(device_name,
node_name,
maybe_base_expanded_node_name,
output_slot,
debug_op,
tensor_value,
wall_time):
"""Create a dict() as the outgoing data in the tensor data comm route.
Note: The tensor data in the comm route does not include the value of the
tensor in its entirety in general. Only if a tensor satisfies the following
conditions will its entire value be included in the return value of this
method:
1. Has a numeric data type (e.g., float32, int32) and has fewer than 5
elements.
2. Is a string tensor and has fewer than 5 elements. Each string element is
up to 40 bytes.
Args:
device_name: Name of the device that the tensor is on.
node_name: (Original) name of the node that produces the tensor.
maybe_base_expanded_node_name: Possbily base-expanded node name.
output_slot: Output slot number.
debug_op: Name of the debug op.
tensor_value: Value of the tensor, as a numpy.ndarray.
wall_time: Wall timestamp for the tensor.
Returns:
A dict representing the tensor data.
"""
output_slot = int(output_slot)
tf.logging.info(
'Recording tensor value: %s, %d, %s', node_name, output_slot, debug_op)
tensor_values = None
if isinstance(tensor_value, debug_data.InconvertibleTensorProto):
if not tensor_value.initialized:
tensor_dtype = UNINITIALIZED_TAG
tensor_shape = UNINITIALIZED_TAG
else:
tensor_dtype = UNSUPPORTED_TAG
tensor_shape = UNSUPPORTED_TAG
tensor_values = NA_TAG
else:
tensor_dtype = tensor_helper.translate_dtype(tensor_value.dtype)
tensor_shape = tensor_value.shape
# The /comm endpoint should respond with tensor values only if the tensor is
# small enough. Otherwise, the detailed values sould be queried through a
# dedicated tensor_data that supports slicing.
if tensor_helper.numel(tensor_shape) < 5:
_, _, tensor_values = tensor_helper.array_view(tensor_value)
if tensor_dtype == 'string' and tensor_value is not None:
tensor_values = tensor_helper.process_buffers_for_display(
tensor_values, limit=STRING_ELEMENT_MAX_LEN)
return {
'type': 'tensor',
'timestamp': wall_time,
'data': {
'device_name': device_name,
'node_name': node_name,
'maybe_base_expanded_node_name': maybe_base_expanded_node_name,
'output_slot': output_slot,
'debug_op': debug_op,
'dtype': tensor_dtype,
'shape': tensor_shape,
'values': tensor_values,
},
}
