1# Copyright 2017 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""Library for arbitrary expression evaluation based on a debugger data dump.""" 16import re 17 18import numpy as np # pylint: disable=unused-import 19 20from tensorflow.python.debug.lib import debug_data 21 22_DUMP_TENSOR_PATTERN = re.compile(r"`.*?`") 23_DEVICE_NAME_PREFIX_PATTERN = re.compile( 24 r"/job:(\w)+/replica:(\d)+/task:(\d)+/(\w)+:(\d)+:") 25_EXEC_INDEX_SUFFIX_PATTERN = re.compile(r"\[(\d)*\]$") 26 27_DEFAULT_DEBUG_OP = "DebugIdentity" 28 29 30def _parse_debug_tensor_name(debug_tensor_name): 31 # pylint: disable=line-too-long 32 """Parse a debug tensor name in a to-be-evaluated expression. 33 34 Args: 35 debug_tensor_name: name of the debug tensor, with or without 36 device name as a prefix, with or without debug op, with or 37 without '[<exec_index>]' as a suffix. 38 E.g., without device name prefix, without debug op suffix: 39 "hidden_0/MatMul:0" 40 E.g., with device name prefix: 41 "/job:worker/replica:0/task:1/gpu:0:hidden_0/MatMul:0" 42 E.g., with debug op suffix: 43 "hidden_0/MatMul:0:DebugNumericSummary" 44 E.g., with device name prefix and debug op suffix: 45 "/job:worker/replica:0/task:1/gpu:0:hidden_0/MatMul:0:DebugNumericSummary" 46 E.g., with device name prefix, debug op and an exec index: 47 "/job:worker/replica:0/task:1/gpu:0:hidden_0/MatMul:0:DebugNumericSummary[1]" 48 49 Returns: 50 device_name: If device name prefix exists, the device name; otherwise, 51 `None`. 52 node_name: Name of the node. 53 output_slot: Output slot index as an `int`. 54 debug_op: If the debug op suffix exists, the debug op name; otherwise, 55 `None`. 56 exec_index: Execution index (applicable to cases in which a debug tensor 57 is computed multiple times in a `tf.Session.run` call, e.g., due to 58 `tf.while_loop`). If the exec_index suffix does not exist, this value 59 defaults to `0`. 60 61 Raises: 62 ValueError: If the input `debug_tensor_name` is malformed. 63 """ 64 # pylint: enable=line-too-long 65 device_prefix_match = re.match(_DEVICE_NAME_PREFIX_PATTERN, debug_tensor_name) 66 if device_prefix_match: 67 device_name = debug_tensor_name[ 68 device_prefix_match.start() : device_prefix_match.end() - 1] 69 debug_tensor_name = debug_tensor_name[device_prefix_match.end():] 70 else: 71 device_name = None 72 73 split_items = debug_tensor_name.split(":") 74 if len(split_items) not in (2, 3): 75 raise ValueError( 76 "The debug tensor name in the to-be-evaluated expression is malformed: " 77 "'%s'" % debug_tensor_name) 78 # TODO(cais): Provide examples of good debug tensor names in the error 79 # message. 80 81 exec_index_match = re.search(_EXEC_INDEX_SUFFIX_PATTERN, split_items[-1]) 82 if exec_index_match: 83 exec_index = int(split_items[-1][ 84 exec_index_match.start() + 1 : exec_index_match.end() - 1]) 85 split_items[-1] = split_items[-1][:exec_index_match.start()] 86 else: 87 exec_index = 0 88 89 if len(split_items) == 2: 90 node_name = split_items[0] 91 output_slot = int(split_items[1]) 92 debug_op = _DEFAULT_DEBUG_OP 93 else: 94 split_items = debug_tensor_name.split(":") 95 node_name = split_items[0] 96 output_slot = int(split_items[1]) 97 debug_op = split_items[2] 98 99 return device_name, node_name, output_slot, debug_op, exec_index 100 101 102class ExpressionEvaluator(object): 103 """Evaluates Python expressions using debug tensor values from a dump.""" 104 105 def __init__(self, dump): 106 """Constructor of ExpressionEvaluator. 107 108 Args: 109 dump: an instance of `DebugDumpDir`. 110 """ 111 self._dump = dump 112 self._cached_tensor_values = {} 113 114 def evaluate(self, expression): 115 """Parse an expression. 116 117 Args: 118 expression: the expression to be parsed. 119 120 Returns: 121 The result of the evaluation. 122 123 Raises: 124 ValueError: If the value of one or more of the debug tensors in the 125 expression are not available. 126 """ 127 dump_tensors_iter = re.finditer(_DUMP_TENSOR_PATTERN, expression) 128 rewritten_expression = expression 129 for match in reversed(list(dump_tensors_iter)): 130 tensor_name = match.group(0)[1:-1].strip() 131 device_name, node_name, output_slot, debug_op, exec_index = ( 132 _parse_debug_tensor_name(tensor_name)) 133 if tensor_name not in self._cached_tensor_values: 134 try: 135 value = self._dump.get_tensors( 136 node_name, output_slot, debug_op, 137 device_name=device_name)[exec_index] 138 except debug_data.WatchKeyDoesNotExistInDebugDumpDirError: 139 raise ValueError( 140 "Eval failed due to the value of %s:%d:DebugIdentity being " 141 "unavailable" % (node_name, output_slot)) 142 self._cached_tensor_values[tensor_name] = value 143 rewritten_expression = ( 144 rewritten_expression[:match.start(0)] + 145 "self._cached_tensor_values['" + tensor_name + "']" + 146 rewritten_expression[match.end(0):]) 147 148 return eval(rewritten_expression) # pylint: disable=eval-used 149