• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
16 #include <string>
17 #include <vector>
18 
19 #include "absl/strings/numbers.h"
20 #include "absl/strings/str_join.h"
21 #include "absl/strings/str_split.h"
22 #include "absl/strings/strip.h"
23 #include "absl/types/optional.h"
24 #include "tensorflow/lite/toco/toco_cmdline_flags.h"
25 #include "tensorflow/lite/toco/toco_port.h"
26 #include "tensorflow/core/platform/logging.h"
27 #include "tensorflow/core/util/command_line_flags.h"
28 
29 namespace toco {
30 
ParseTocoFlagsFromCommandLineFlags(int * argc,char * argv[],std::string * msg,ParsedTocoFlags * parsed_toco_flags_ptr)31 bool ParseTocoFlagsFromCommandLineFlags(
32     int* argc, char* argv[], std::string* msg,
33     ParsedTocoFlags* parsed_toco_flags_ptr) {
34   using tensorflow::Flag;
35   ParsedTocoFlags& parsed_flags = *parsed_toco_flags_ptr;
36   std::vector<tensorflow::Flag> flags = {
37       Flag("input_file", parsed_flags.input_file.bind(),
38            parsed_flags.input_file.default_value(),
39            "Input file (model of any supported format). For Protobuf "
40            "formats, both text and binary are supported regardless of file "
41            "extension."),
42       Flag("savedmodel_directory", parsed_flags.savedmodel_directory.bind(),
43            parsed_flags.savedmodel_directory.default_value(),
44            "Deprecated. Full path to the directory containing the SavedModel."),
45       Flag("output_file", parsed_flags.output_file.bind(),
46            parsed_flags.output_file.default_value(),
47            "Output file. "
48            "For Protobuf formats, the binary format will be used."),
49       Flag("input_format", parsed_flags.input_format.bind(),
50            parsed_flags.input_format.default_value(),
51            "Input file format. One of: TENSORFLOW_GRAPHDEF, TFLITE."),
52       Flag("output_format", parsed_flags.output_format.bind(),
53            parsed_flags.output_format.default_value(),
54            "Output file format. "
55            "One of TENSORFLOW_GRAPHDEF, TFLITE, GRAPHVIZ_DOT."),
56       Flag("savedmodel_tagset", parsed_flags.savedmodel_tagset.bind(),
57            parsed_flags.savedmodel_tagset.default_value(),
58            "Deprecated. Comma-separated set of tags identifying the "
59            "MetaGraphDef within the SavedModel to analyze. All tags in the tag "
60            "set must be specified."),
61       Flag("default_ranges_min", parsed_flags.default_ranges_min.bind(),
62            parsed_flags.default_ranges_min.default_value(),
63            "If defined, will be used as the default value for the min bound "
64            "of min/max ranges used for quantization of uint8 arrays."),
65       Flag("default_ranges_max", parsed_flags.default_ranges_max.bind(),
66            parsed_flags.default_ranges_max.default_value(),
67            "If defined, will be used as the default value for the max bound "
68            "of min/max ranges used for quantization of uint8 arrays."),
69       Flag("default_int16_ranges_min",
70            parsed_flags.default_int16_ranges_min.bind(),
71            parsed_flags.default_int16_ranges_min.default_value(),
72            "If defined, will be used as the default value for the min bound "
73            "of min/max ranges used for quantization of int16 arrays."),
74       Flag("default_int16_ranges_max",
75            parsed_flags.default_int16_ranges_max.bind(),
76            parsed_flags.default_int16_ranges_max.default_value(),
77            "If defined, will be used as the default value for the max bound "
78            "of min/max ranges used for quantization of int16 arrays."),
79       Flag("inference_type", parsed_flags.inference_type.bind(),
80            parsed_flags.inference_type.default_value(),
81            "Target data type of arrays in the output file (for input_arrays, "
82            "this may be overridden by inference_input_type). "
83            "One of FLOAT, QUANTIZED_UINT8."),
84       Flag("inference_input_type", parsed_flags.inference_input_type.bind(),
85            parsed_flags.inference_input_type.default_value(),
86            "Target data type of input arrays. "
87            "If not specified, inference_type is used. "
88            "One of FLOAT, QUANTIZED_UINT8."),
89       Flag("input_type", parsed_flags.input_type.bind(),
90            parsed_flags.input_type.default_value(),
91            "Deprecated ambiguous flag that set both --input_data_types and "
92            "--inference_input_type."),
93       Flag("input_types", parsed_flags.input_types.bind(),
94            parsed_flags.input_types.default_value(),
95            "Deprecated ambiguous flag that set both --input_data_types and "
96            "--inference_input_type. Was meant to be a "
97            "comma-separated list, but this was deprecated before "
98            "multiple-input-types was ever properly supported."),
99 
100       Flag("drop_fake_quant", parsed_flags.drop_fake_quant.bind(),
101            parsed_flags.drop_fake_quant.default_value(),
102            "Ignore and discard FakeQuant nodes. For instance, to "
103            "generate plain float code without fake-quantization from a "
104            "quantized graph."),
105       Flag(
106           "reorder_across_fake_quant",
107           parsed_flags.reorder_across_fake_quant.bind(),
108           parsed_flags.reorder_across_fake_quant.default_value(),
109           "Normally, FakeQuant nodes must be strict boundaries for graph "
110           "transformations, in order to ensure that quantized inference has "
111           "the exact same arithmetic behavior as quantized training --- which "
112           "is the whole point of quantized training and of FakeQuant nodes in "
113           "the first place. "
114           "However, that entails subtle requirements on where exactly "
115           "FakeQuant nodes must be placed in the graph. Some quantized graphs "
116           "have FakeQuant nodes at unexpected locations, that prevent graph "
117           "transformations that are necessary in order to generate inference "
118           "code for these graphs. Such graphs should be fixed, but as a "
119           "temporary work-around, setting this reorder_across_fake_quant flag "
120           "allows TOCO to perform necessary graph transformaitons on them, "
121           "at the cost of no longer faithfully matching inference and training "
122           "arithmetic."),
123       Flag("allow_custom_ops", parsed_flags.allow_custom_ops.bind(),
124            parsed_flags.allow_custom_ops.default_value(),
125            "If true, allow TOCO to create TF Lite Custom operators for all the "
126            "unsupported TensorFlow ops."),
127       Flag("custom_opdefs", parsed_flags.custom_opdefs.bind(),
128            parsed_flags.custom_opdefs.default_value(),
129            "List of strings representing custom ops OpDefs that are included "
130            "in the GraphDef."),
131       Flag("allow_dynamic_tensors", parsed_flags.allow_dynamic_tensors.bind(),
132            parsed_flags.allow_dynamic_tensors.default_value(),
133            "Boolean flag indicating whether the converter should allow models "
134            "with dynamic Tensor shape. When set to False, the converter will "
135            "generate runtime memory offsets for activation Tensors (with 128 "
136            "bits alignment) and error out on models with undetermined Tensor "
137            "shape. (Default: True)"),
138       Flag(
139           "drop_control_dependency",
140           parsed_flags.drop_control_dependency.bind(),
141           parsed_flags.drop_control_dependency.default_value(),
142           "If true, ignore control dependency requirements in input TensorFlow "
143           "GraphDef. Otherwise an error will be raised upon control dependency "
144           "inputs."),
145       Flag("debug_disable_recurrent_cell_fusion",
146            parsed_flags.debug_disable_recurrent_cell_fusion.bind(),
147            parsed_flags.debug_disable_recurrent_cell_fusion.default_value(),
148            "If true, disable fusion of known identifiable cell subgraphs into "
149            "cells. This includes, for example, specific forms of LSTM cell."),
150       Flag("propagate_fake_quant_num_bits",
151            parsed_flags.propagate_fake_quant_num_bits.bind(),
152            parsed_flags.propagate_fake_quant_num_bits.default_value(),
153            "If true, use FakeQuant* operator num_bits attributes to adjust "
154            "array data_types."),
155       Flag("allow_nudging_weights_to_use_fast_gemm_kernel",
156            parsed_flags.allow_nudging_weights_to_use_fast_gemm_kernel.bind(),
157            parsed_flags.allow_nudging_weights_to_use_fast_gemm_kernel
158                .default_value(),
159            "Some fast uint8 GEMM kernels require uint8 weights to avoid the "
160            "value 0. This flag allows nudging them to 1 to allow proceeding, "
161            "with moderate inaccuracy."),
162       Flag("dedupe_array_min_size_bytes",
163            parsed_flags.dedupe_array_min_size_bytes.bind(),
164            parsed_flags.dedupe_array_min_size_bytes.default_value(),
165            "Minimum size of constant arrays to deduplicate; arrays smaller "
166            "will not be deduplicated."),
167       Flag("split_tflite_lstm_inputs",
168            parsed_flags.split_tflite_lstm_inputs.bind(),
169            parsed_flags.split_tflite_lstm_inputs.default_value(),
170            "Split the LSTM inputs from 5 tensors to 18 tensors for TFLite. "
171            "Ignored if the output format is not TFLite."),
172       Flag("quantize_to_float16", parsed_flags.quantize_to_float16.bind(),
173            parsed_flags.quantize_to_float16.default_value(),
174            "Used in conjunction with post_training_quantize. Specifies that "
175            "the weights should be quantized to fp16 instead of the default "
176            "(int8)"),
177       Flag("quantize_weights", parsed_flags.quantize_weights.bind(),
178            parsed_flags.quantize_weights.default_value(),
179            "Deprecated. Please use --post_training_quantize instead."),
180       Flag("post_training_quantize", parsed_flags.post_training_quantize.bind(),
181            parsed_flags.post_training_quantize.default_value(),
182            "Boolean indicating whether to quantize the weights of the "
183            "converted float model. Model size will be reduced and there will "
184            "be latency improvements (at the cost of accuracy)."),
185       // TODO(b/118822804): Unify the argument definition with `tflite_convert`.
186       // WARNING: Experimental interface, subject to change
187       Flag("enable_select_tf_ops", parsed_flags.enable_select_tf_ops.bind(),
188            parsed_flags.enable_select_tf_ops.default_value(), ""),
189       // WARNING: Experimental interface, subject to change
190       Flag("force_select_tf_ops", parsed_flags.force_select_tf_ops.bind(),
191            parsed_flags.force_select_tf_ops.default_value(), ""),
192       // WARNING: Experimental interface, subject to change
193       Flag("unfold_batchmatmul", parsed_flags.unfold_batchmatmul.bind(),
194            parsed_flags.unfold_batchmatmul.default_value(), ""),
195       // WARNING: Experimental interface, subject to change
196       Flag("accumulation_type", parsed_flags.accumulation_type.bind(),
197            parsed_flags.accumulation_type.default_value(),
198            "Accumulation type to use with quantize_to_float16"),
199       // WARNING: Experimental interface, subject to change
200       Flag("allow_bfloat16", parsed_flags.allow_bfloat16.bind(),
201            parsed_flags.allow_bfloat16.default_value(), "")};
202 
203   bool asked_for_help =
204       *argc == 2 && (!strcmp(argv[1], "--help") || !strcmp(argv[1], "-help"));
205   if (asked_for_help) {
206     *msg += tensorflow::Flags::Usage(argv[0], flags);
207     return false;
208   } else {
209     return tensorflow::Flags::Parse(argc, argv, flags);
210   }
211 }
212 
213 namespace {
214 
215 // Defines the requirements for a given flag. kUseDefault means the default
216 // should be used in cases where the value isn't specified by the user.
217 enum class FlagRequirement {
218   kNone,
219   kMustBeSpecified,
220   kMustNotBeSpecified,
221   kUseDefault,
222 };
223 
224 // Enforces the FlagRequirements are met for a given flag.
225 template <typename T>
EnforceFlagRequirement(const T & flag,const std::string & flag_name,FlagRequirement requirement)226 void EnforceFlagRequirement(const T& flag, const std::string& flag_name,
227                             FlagRequirement requirement) {
228   if (requirement == FlagRequirement::kMustBeSpecified) {
229     QCHECK(flag.specified()) << "Missing required flag " << flag_name;
230   }
231   if (requirement == FlagRequirement::kMustNotBeSpecified) {
232     QCHECK(!flag.specified())
233         << "Given other flags, this flag should not have been specified: "
234         << flag_name;
235   }
236 }
237 
238 // Gets the value from the flag if specified. Returns default if the
239 // FlagRequirement is kUseDefault.
240 template <typename T>
GetFlagValue(const Arg<T> & flag,FlagRequirement requirement)241 absl::optional<T> GetFlagValue(const Arg<T>& flag,
242                                FlagRequirement requirement) {
243   if (flag.specified()) return flag.value();
244   if (requirement == FlagRequirement::kUseDefault) return flag.default_value();
245   return absl::optional<T>();
246 }
247 
248 }  // namespace
249 
ReadTocoFlagsFromCommandLineFlags(const ParsedTocoFlags & parsed_toco_flags,TocoFlags * toco_flags)250 void ReadTocoFlagsFromCommandLineFlags(const ParsedTocoFlags& parsed_toco_flags,
251                                        TocoFlags* toco_flags) {
252   namespace port = toco::port;
253   port::CheckInitGoogleIsDone("InitGoogle is not done yet");
254 
255 #define READ_TOCO_FLAG(name, requirement)                                \
256   do {                                                                   \
257     EnforceFlagRequirement(parsed_toco_flags.name, #name, requirement);  \
258     auto flag_value = GetFlagValue(parsed_toco_flags.name, requirement); \
259     if (flag_value.has_value()) {                                        \
260       toco_flags->set_##name(flag_value.value());                        \
261     }                                                                    \
262   } while (false)
263 
264 #define PARSE_TOCO_FLAG(Type, name, requirement)                         \
265   do {                                                                   \
266     EnforceFlagRequirement(parsed_toco_flags.name, #name, requirement);  \
267     auto flag_value = GetFlagValue(parsed_toco_flags.name, requirement); \
268     if (flag_value.has_value()) {                                        \
269       Type x;                                                            \
270       QCHECK(Type##_Parse(flag_value.value(), &x))                       \
271           << "Unrecognized " << #Type << " value "                       \
272           << parsed_toco_flags.name.value();                             \
273       toco_flags->set_##name(x);                                         \
274     }                                                                    \
275   } while (false)
276 
277   PARSE_TOCO_FLAG(FileFormat, input_format, FlagRequirement::kUseDefault);
278   PARSE_TOCO_FLAG(FileFormat, output_format, FlagRequirement::kUseDefault);
279   PARSE_TOCO_FLAG(IODataType, inference_type, FlagRequirement::kNone);
280   PARSE_TOCO_FLAG(IODataType, inference_input_type, FlagRequirement::kNone);
281   READ_TOCO_FLAG(default_ranges_min, FlagRequirement::kNone);
282   READ_TOCO_FLAG(default_ranges_max, FlagRequirement::kNone);
283   READ_TOCO_FLAG(default_int16_ranges_min, FlagRequirement::kNone);
284   READ_TOCO_FLAG(default_int16_ranges_max, FlagRequirement::kNone);
285   READ_TOCO_FLAG(drop_fake_quant, FlagRequirement::kNone);
286   READ_TOCO_FLAG(reorder_across_fake_quant, FlagRequirement::kNone);
287   READ_TOCO_FLAG(allow_custom_ops, FlagRequirement::kNone);
288   READ_TOCO_FLAG(drop_control_dependency, FlagRequirement::kNone);
289   READ_TOCO_FLAG(debug_disable_recurrent_cell_fusion, FlagRequirement::kNone);
290   READ_TOCO_FLAG(propagate_fake_quant_num_bits, FlagRequirement::kNone);
291   READ_TOCO_FLAG(allow_nudging_weights_to_use_fast_gemm_kernel,
292                  FlagRequirement::kNone);
293   READ_TOCO_FLAG(dedupe_array_min_size_bytes, FlagRequirement::kNone);
294   READ_TOCO_FLAG(split_tflite_lstm_inputs, FlagRequirement::kNone);
295   READ_TOCO_FLAG(quantize_weights, FlagRequirement::kNone);
296   READ_TOCO_FLAG(quantize_to_float16, FlagRequirement::kNone);
297   READ_TOCO_FLAG(post_training_quantize, FlagRequirement::kNone);
298   READ_TOCO_FLAG(enable_select_tf_ops, FlagRequirement::kNone);
299   READ_TOCO_FLAG(force_select_tf_ops, FlagRequirement::kNone);
300   READ_TOCO_FLAG(unfold_batchmatmul, FlagRequirement::kNone);
301   PARSE_TOCO_FLAG(IODataType, accumulation_type, FlagRequirement::kNone);
302   READ_TOCO_FLAG(allow_bfloat16, FlagRequirement::kNone);
303 
304   if (parsed_toco_flags.force_select_tf_ops.value() &&
305       !parsed_toco_flags.enable_select_tf_ops.value()) {
306     // TODO(ycling): Consider to enforce `enable_select_tf_ops` when
307     // `force_select_tf_ops` is true.
308     LOG(WARNING) << "--force_select_tf_ops should always be used with "
309                     "--enable_select_tf_ops.";
310   }
311 
312   // Deprecated flag handling.
313   if (parsed_toco_flags.input_type.specified()) {
314     LOG(WARNING)
315         << "--input_type is deprecated. It was an ambiguous flag that set both "
316            "--input_data_types and --inference_input_type. If you are trying "
317            "to complement the input file with information about the type of "
318            "input arrays, use --input_data_type. If you are trying to control "
319            "the quantization/dequantization of real-numbers input arrays in "
320            "the output file, use --inference_input_type.";
321     toco::IODataType input_type;
322     QCHECK(toco::IODataType_Parse(parsed_toco_flags.input_type.value(),
323                                   &input_type));
324     toco_flags->set_inference_input_type(input_type);
325   }
326   if (parsed_toco_flags.input_types.specified()) {
327     LOG(WARNING)
328         << "--input_types is deprecated. It was an ambiguous flag that set "
329            "both --input_data_types and --inference_input_type. If you are "
330            "trying to complement the input file with information about the "
331            "type of input arrays, use --input_data_type. If you are trying to "
332            "control the quantization/dequantization of real-numbers input "
333            "arrays in the output file, use --inference_input_type.";
334     std::vector<std::string> input_types =
335         absl::StrSplit(parsed_toco_flags.input_types.value(), ',');
336     QCHECK(!input_types.empty());
337     for (size_t i = 1; i < input_types.size(); i++) {
338       QCHECK_EQ(input_types[i], input_types[0]);
339     }
340     toco::IODataType input_type;
341     QCHECK(toco::IODataType_Parse(input_types[0], &input_type));
342     toco_flags->set_inference_input_type(input_type);
343   }
344   if (parsed_toco_flags.quantize_weights.value()) {
345     LOG(WARNING)
346         << "--quantize_weights is deprecated. Falling back to "
347            "--post_training_quantize. Please switch --post_training_quantize.";
348     toco_flags->set_post_training_quantize(
349         parsed_toco_flags.quantize_weights.value());
350   }
351   if (parsed_toco_flags.quantize_weights.value()) {
352     if (toco_flags->inference_type() == IODataType::QUANTIZED_UINT8) {
353       LOG(WARNING)
354           << "--post_training_quantize quantizes a graph of inference_type "
355              "FLOAT. Overriding inference type QUANTIZED_UINT8 to FLOAT.";
356       toco_flags->set_inference_type(IODataType::FLOAT);
357     }
358   }
359 
360 #undef READ_TOCO_FLAG
361 #undef PARSE_TOCO_FLAG
362 }
363 }  // namespace toco
364