package tensorflow.tfprof

Get desktop application:
View/edit binary Protocol Buffers messages

• map<string, AdviceProto.Checker> checkers = 1

  checker name -> a list of reports from the checker.

Used in: AdviceProto

• repeated string reports = 2

• map<string, AdvisorOptionsProto.CheckerOption> checkers = 1

  checker name -> a dict of key-value options.

Used in: AdvisorOptionsProto

• map<string, string> options = 1

It specifies the Python callstack that creates an op.

Used in: CallTraceback, OpLogEntry, ProfileNode

• repeated CodeDef.Trace traces = 1

Used in: CodeDef

• string file = 1

  deprecated by file_id.

• int64 file_id = 6

• int32 lineno = 2

• string function = 3

  deprecated by function_id.

• int64 function_id = 7

• string line = 4

  deprecated line_id.

• int64 line_id = 8

• int32 func_start_line = 5

Used in: ExecProfile

• int64 memory_micros = 1

  This is the timestamp when the memory information was tracked.

• int64 host_temp_bytes = 2

  NOTE: Please don't depend on the following 4 fields yet. Due to TensorFlow internal tracing issues, the numbers can be quite wrong. TODO(xpan): Fix the TensorFlow internal tracing.

• int64 host_persistent_bytes = 3

• int64 accelerator_temp_bytes = 4

• int64 accelerator_persistent_bytes = 5

• int64 requested_bytes = 6

  Total bytes requested by the op.

• int64 peak_bytes = 7

  Total bytes requested by the op and released before op end.

• int64 residual_bytes = 8

  Total bytes requested by the op and not released after op end.

• int64 output_bytes = 9

  Total bytes output by the op (not necessarily requested by the op).

• int64 allocator_bytes_in_use = 10

  The total number of bytes currently allocated by the allocator if >0.

• map<int32, Memory> output_memory = 11

  The memory of each output of the operation.

Used in: ProfileNode

• int64 run_count = 1

  Can be larger than 1 if run multiple times in loop.

• int64 all_start_micros = 2

  The earliest/latest time including scheduling and execution.

• int64 latest_end_micros = 3

• map<string, ExecTime> accelerator_execs = 4

  device -> vector of {op_start_micros, op_exec_micros} pairs. accelerator_execs: gpu:id/stream:all -> {op_start_micros, op_exec_micros} For accelerator, vector size can be larger than 1, multiple kernel fires or in tf.while_loop.

• map<string, ExecTime> cpu_execs = 5

  cpu_execs: cpu/gpu:id -> {op_start_micros, op_exec_micros} For cpu, vector size can be larger than 1 if in tf.while_loop.

• repeated ExecMemory memory_execs = 7

  Each entry to memory information of a scheduling of the node. Normally, there will be multiple entries in while_loop.

• repeated AllocationRecord allocations = 11

  The allocation and deallocation times and sizes throughout execution.

• repeated string devices = 6

  The devices related to this execution.

Used in: ExecProfile

• repeated Tuple times = 1

A node in TensorFlow graph. Used by scope/graph view.

Used in: MultiGraphNodeProto

• string name = 1

  op name.

• optional TFProfTensorProto tensor_value = 15

  tensor value restored from checkpoint.

• int64 run_count = 21

  op execution time. A node can be defined once but run multiple times in tf.while_loop. the times sum up all different runs.

• int64 exec_micros = 2

• int64 accelerator_exec_micros = 17

• int64 cpu_exec_micros = 18

• int64 requested_bytes = 3

  Total bytes requested by the op.

• int64 peak_bytes = 24

  Max bytes allocated and being used by the op at a point.

• int64 residual_bytes = 25

  Total bytes requested by the op and not released before end.

• int64 output_bytes = 26

  Total bytes output by the op (not necessarily allocated by the op).

• int64 parameters = 4

  Number of parameters if available.

• int64 float_ops = 13

  Number of float operations.

• repeated string devices = 10

  Device the op is assigned to. Since an op can fire multiple kernel calls, there can be multiple devices.

• int64 total_definition_count = 23

  The following are the aggregated stats from all *accounted* children and the node itself. The actual children depend on the data structure used. In graph view, children are inputs recursively. In scope view, children are nodes under the name scope.

• int64 total_run_count = 22

• int64 total_exec_micros = 6

• int64 total_accelerator_exec_micros = 19

• int64 total_cpu_exec_micros = 20

• int64 total_requested_bytes = 7

• int64 total_peak_bytes = 27

• int64 total_residual_bytes = 28

• int64 total_output_bytes = 29

• int64 total_parameters = 8

• int64 total_float_ops = 14

• repeated TensorShapeProto shapes = 11

  shape information, if available. TODO(xpan): Why is this repeated?

• map<int32, TensorShapeProto> input_shapes = 16

• repeated GraphNodeProto children = 12

  Descendants of the graph. The actual descendants depend on the data structure used (scope, graph).

Used in: ExecMemory

• int64 bytes = 1

• uint64 ptr = 2

A node that groups multiple GraphNodeProto. Depending on the 'view', the semantics of the TFmultiGraphNodeProto is different: code view: A node groups all TensorFlow graph nodes created by the Python code. op view: A node groups all TensorFlow graph nodes that are of type of the op (e.g. MatMul, Conv2D).

• string name = 1

  Name of the node.

• int64 exec_micros = 2

  code execution time.

• int64 accelerator_exec_micros = 12

• int64 cpu_exec_micros = 13

• int64 requested_bytes = 3

  Total requested bytes by the code.

• int64 peak_bytes = 16

  Max bytes allocated and being used by the op at a point.

• int64 residual_bytes = 17

  Total bytes requested by the op and not released before end.

• int64 output_bytes = 18

  Total bytes output by the op (not necessarily allocated by the op).

• int64 parameters = 4

  Number of parameters if available.

• int64 float_ops = 5

  Number of float operations.

• int64 total_exec_micros = 6

  The following are the aggregated stats from descendants. The actual descendants depend on the data structure used.

• int64 total_accelerator_exec_micros = 14

• int64 total_cpu_exec_micros = 15

• int64 total_requested_bytes = 7

• int64 total_peak_bytes = 19

• int64 total_residual_bytes = 20

• int64 total_output_bytes = 21

• int64 total_parameters = 8

• int64 total_float_ops = 9

• repeated GraphNodeProto graph_nodes = 10

  TensorFlow graph nodes contained by the MultiGraphNodeProto.

• repeated MultiGraphNodeProto children = 11

  Descendants of the node. The actual descendants depend on the data structure used.

Used in: OpLogProto

• string name = 1

  op name.

• int64 float_ops = 2

  float_ops is filled by tfprof Python API when called. It requires the op has RegisterStatistics defined. Currently, Conv2D, MatMul, etc, are implemented.

• repeated string types = 3

  User can define extra op type information for an op. This allows the user to select a group of ops precisely using op_type as a key.

• optional CodeDef code_def = 4

  Used to support tfprof "code" view.

Used in: CallTraceback

• repeated OpLogEntry log_entries = 1

• map<int64, string> id_to_string = 2

  Maps from id of CodeDef file,function,line to its string In the future can also map other id of other fields to string.

Refers to tfprof_options.h/cc for documentation. Only used to pass tfprof options from Python to C++.

• int64 max_depth = 1

• int64 min_bytes = 2

• int64 min_peak_bytes = 19

• int64 min_residual_bytes = 20

• int64 min_output_bytes = 21

• int64 min_micros = 3

• int64 min_accelerator_micros = 22

• int64 min_cpu_micros = 23

• int64 min_params = 4

• int64 min_float_ops = 5

• int64 min_occurrence = 17

• int64 step = 18

• string order_by = 7

• repeated string account_type_regexes = 8

• repeated string start_name_regexes = 9

• repeated string trim_name_regexes = 10

• repeated string show_name_regexes = 11

• repeated string hide_name_regexes = 12

• bool account_displayed_op_only = 13

• repeated string select = 14

• string output = 15

• string dump_to_file = 16

Used in: ProfileProto

• string name = 1

  graph node name.

• string op = 9

  graph operation type.

• int64 id = 13

  A unique id for the node.

• map<int32, int64> inputs = 2

• map<int32, Tuple> input_shapes = 16

• map<int32, int64> outputs = 3

• map<int32, Tuple> output_shapes = 15

• map<int64, int32> src_output_index = 14

  A map from source node id to its output index to current node.

• repeated int64 shape = 4

• repeated string op_types = 5

• string canonical_device = 6

• string host_device = 7

• int64 float_ops = 8

• optional CodeDef trace = 10

• map<string, AttrValue> attrs = 11

• map<int64, ExecProfile> execs = 12

A proto representation of the profiler's profile. It allows serialization, shipping around and deserialization of the profiles. Please don't depend on the internals of the profile proto.

• map<int64, ProfileNode> nodes = 1

• bool has_trace = 2

  Whether or not has code traces.

• bool miss_accelerator_stream = 5

  Whether or not the TF device tracer fails to return accelerator information (which could lead to 0 accelerator execution time).

• repeated int64 steps = 3

  Traced steps.

• map<int64, string> id_to_string = 4

  Maps from id of CodeDef file,function,line to its string In the future can also map other id of other fields to string.

Used in: GraphNodeProto

• DataType dtype = 1

• repeated double value_double = 2

  Flatten tensor in row-major. Only one of the following array is set.

• repeated int64 value_int64 = 3

• repeated string value_str = 4

Used in: ExecTime, ProfileNode

• repeated int64 int64_values = 1