experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Improve error logs when converting model with control flow. Also update the g3doc to clearly convey that 'control flow' is not currently supported.

Browse Source 
PiperOrigin-RevId: 248773536
This commit is contained in:
Haoliang Zhang 2019-05-17 13:05:36 -07:00 committed by TensorFlower Gardener
parent 39e91628f5
commit 5260e5f979
3 changed files with 202 additions and 56 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
tensorflow/lite/g3doc/guide/faq.md
View File

@ -30,8 +30,16 @@ Since the number of TensorFlow Lite operations is smaller than TensorFlow's,
some inference models may not be able to convert. For unimplemented operations,
take a look at the question on
[missing operators](faq.md#why-are-some-operations-not-implemented-in-tensorflow-lite).
Unsupported operators include embeddings and LSTM/RNNs. For conversion issues
not related to missing operations, search our
Unsupported operators include embeddings and LSTM/RNNs. For models with
LSTM/RNNs, you can also try the experimental API
[OpHint](https://www.tensorflow.org/api_docs/python/tf/lite/OpHint) to convert.
Models with control flow ops (Switch, Merge, etc) are not convertible at the
moment, but we are working on adding support for control flow in Tensorflow
Lite, please see
[GitHub issues](https://github.com/tensorflow/tensorflow/issues/28485).
For conversion issues not related to missing operations or control flow ops,
search our
[GitHub issues](https://github.com/tensorflow/tensorflow/issues?q=label%3Acomp%3Alite+)
or file a [new one](https://github.com/tensorflow/tensorflow/issues).

127
tensorflow/lite/toco/tflite/export.cc
View File

@ -16,6 +16,7 @@ limitations under the License.
#include "flatbuffers/flexbuffers.h"
#include "absl/strings/str_join.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/lite/context.h"
#include "tensorflow/lite/schema/schema_generated.h"
#include "tensorflow/lite/toco/tflite/operator.h"
@ -431,6 +432,31 @@ tensorflow::Status Export(const Model& model, string* output_file_contents,
return Export(model, output_file_contents, params, ops_by_type);
}
void ParseControlFlowErrors(std::set<string>* custom_ops,
std::vector<string>* error_msgs) {
std::set<string> unsupported_control_flow_ops;
// Check if unsupported ops contains control flow ops. It's impossible
// to implement these ops as custom ops at the moment.
for (const auto& op : *custom_ops) {
if (IsControlFlowOp(op)) {
unsupported_control_flow_ops.insert(op);
}
}
if (!unsupported_control_flow_ops.empty()) {
error_msgs->push_back(absl::StrCat(
"TensorFlow Lite currently doesn't support control flow ops: ",
absl::StrJoin(unsupported_control_flow_ops, ", "), ".",
" We are working on supporting control flow ops, please see github "
"issue at "
"https://github.com/tensorflow/tensorflow/issues/28485."));
}
// Remove control flow ops from `custom_ops` set so that they won't be
// reported again in later messages.
for (const auto& op : unsupported_control_flow_ops) {
custom_ops->erase(op);
}
}
tensorflow::Status Export(
const Model& model, string* output_file_contents,
const ExportParams& params,
@ -483,6 +509,18 @@ tensorflow::Status Export(
if (!custom_ops.empty()) {
if (!params.allow_custom_ops) {
auto please_report_bug_message = []() {
return "We are continually in the process of adding support to "
"TensorFlow Lite for more ops. It would be helpful if you could "
"inform us of how this conversion went by opening a github "
"issue at "
"https://github.com/tensorflow/tensorflow/issues/new?template="
"40-tflite-op-request.md\n and pasting the following:\n\n";
};
std::vector<string> error_msgs;
ParseControlFlowErrors(&custom_ops, &error_msgs);
// Remove ExpandDims and ReorderAxes from unimplemented list unless they
// compose the list. Both ops are removed during graph transformations.
// However, if an op is unimplemented earlier in the model, the graph
@ -499,62 +537,45 @@ tensorflow::Status Export(
custom_ops_final = custom_ops;
}
auto please_report_bug_message = []() {
return "We are continually in the process of adding support to "
"TensorFlow Lite for more ops. It would be helpful if you could "
"inform us of how this conversion went by opening a github "
"issue at "
"https://github.com/tensorflow/tensorflow/issues/new?template="
"40-tflite-op-request.md\n and pasting the following:\n\n";
};
if (params.enable_select_tf_ops) {
return tensorflow::errors::InvalidArgument(absl::StrCat(
please_report_bug_message(),
"Some of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime and are not recognized by "
"TensorFlow. If you have a custom "
"implementation for them you can disable this error with "
"--allow_custom_ops, or by setting allow_custom_ops=True "
"when calling tf.lite.TFLiteConverter(). Here is a list "
"of builtin operators you are using: ",
absl::StrJoin(builtin_ops, ", "),
". Here is a list "
"of operators for which you will need custom implementations: ",
absl::StrJoin(custom_ops_final, ", "), "."));
} else {
return tensorflow::errors::InvalidArgument(absl::StrCat(
please_report_bug_message(),
"Some of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime. If those are native "
"TensorFlow operators, you might be able to use the extended "
"runtime by passing --enable_select_tf_ops, or by setting "
"target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling "
"tf.lite.TFLiteConverter(). Otherwise, if you have a "
"custom implementation for them you can disable this error with "
"--allow_custom_ops, or by setting allow_custom_ops=True "
"when calling tf.lite.TFLiteConverter(). Here is a list "
"of builtin operators you are using: ",
absl::StrJoin(builtin_ops, ", "),
". Here is a list "
"of operators for which you will need custom implementations: ",
absl::StrJoin(custom_ops_final, ", "), "."));
if (!custom_ops_final.empty()) {
if (params.enable_select_tf_ops) {
error_msgs.push_back(absl::StrCat(
"Some of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime and are not recognized "
"by "
"TensorFlow. If you have a custom "
"implementation for them you can disable this error with "
"--allow_custom_ops, or by setting allow_custom_ops=True "
"when calling tf.lite.TFLiteConverter(). Here is a list "
"of builtin operators you are using: ",
absl::StrJoin(builtin_ops, ", "),
". Here is a list "
"of operators for which you will need custom implementations: ",
absl::StrJoin(custom_ops_final, ", "), "."));
} else {
error_msgs.push_back(absl::StrCat(
"Some of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime. If those are native "
"TensorFlow operators, you might be able to use the extended "
"runtime by passing --enable_select_tf_ops, or by setting "
"target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling "
"tf.lite.TFLiteConverter(). Otherwise, if you have a "
"custom implementation for them you can disable this error "
"with "
"--allow_custom_ops, or by setting allow_custom_ops=True "
"when calling tf.lite.TFLiteConverter(). Here is a list "
"of builtin operators you are using: ",
absl::StrJoin(builtin_ops, ", "),
". Here is a list "
"of operators for which you will need custom implementations: ",
absl::StrJoin(custom_ops_final, ", "), "."));
}
}
}
std::set<string> unsupported_control_flow_ops;
// Check if unsupported ops contains control flow ops. It's impossible
// to implement these ops as custom ops at the moment.
for (const auto& op : custom_ops) {
if (IsControlFlowOp(op)) {
unsupported_control_flow_ops.insert(op);
if (!error_msgs.empty()) {
return tensorflow::errors::InvalidArgument(absl::StrCat(
please_report_bug_message(), absl::StrJoin(error_msgs, " ")));
}
}
if (!unsupported_control_flow_ops.empty()) {
return tensorflow::errors::InvalidArgument(absl::StrCat(
"TensorFlow Lite currently doesn't support control flow ops: ",
absl::StrJoin(unsupported_control_flow_ops, ", "), "."));
}
}
if (!unsupported_flex_ops.empty()) {

119
tensorflow/lite/toco/tflite/export_test.cc
View File

@ -227,6 +227,123 @@ TEST_F(ExportTest, Export) {
EXPECT_THAT(ExportAndGetOperatorIndices(params), ElementsAre(1, 0, 2, 3));
}
TEST_F(ExportTest, UnsupportedControlFlowErrors) {
AddOperatorsByName({"Conv", "Add", "Switch", "Merge"});
ExportParams params;
params.allow_custom_ops = false;
// The model contains control flow ops which are not convertible, so we should
// check the returned error message.
string output;
const auto ops_by_type = BuildOperatorByTypeMap();
auto status = Export(input_model_, &output, params, ops_by_type);
EXPECT_EQ(status.error_message(),
"We are continually in the process of adding support to TensorFlow "
"Lite for more ops. It would be helpful if you could inform us of "
"how this conversion went by opening a github issue at "
"https://github.com/tensorflow/tensorflow/issues/"
"new?template=40-tflite-op-request.md\n and pasting the "
"following:\n\nTensorFlow Lite currently doesn't support control "
"flow ops: Merge, Switch. We are working on supporting control "
"flow ops, please see github issue at "
"https://github.com/tensorflow/tensorflow/issues/28485.");
}
TEST_F(ExportTest, UnsupportedOpsAndNeedEnableFlex) {
AddOperatorsByName({"Conv", "Add", "BatchNormWithGlobalNormalization"});
ExportParams params;
params.allow_custom_ops = false;
params.enable_select_tf_ops = false;
string output;
const auto ops_by_type = BuildOperatorByTypeMap();
auto status = Export(input_model_, &output, params, ops_by_type);
EXPECT_EQ(
status.error_message(),
"We are continually in the process of adding support to TensorFlow Lite "
"for more ops. It would be helpful if you could inform us of how this "
"conversion went by opening a github issue at "
"https://github.com/tensorflow/tensorflow/issues/"
"new?template=40-tflite-op-request.md\n and pasting the "
"following:\n\nSome of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime. If those are native TensorFlow "
"operators, you might be able to use the extended runtime by passing "
"--enable_select_tf_ops, or by setting "
"target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling "
"tf.lite.TFLiteConverter(). Otherwise, if you have a custom "
"implementation for them you can disable this error with "
"--allow_custom_ops, or by setting allow_custom_ops=True when calling "
"tf.lite.TFLiteConverter(). Here is a list of builtin operators you are "
"using: ADD, CONV_2D. Here is a list of operators for which you will "
"need custom implementations: BatchNormWithGlobalNormalization.");
}
TEST_F(ExportTest, UnsupportedOpsNeedCustomImplementation) {
AddOperatorsByName({"Conv", "Add", "MyCustomOp1", "MyCustomOp2"});
ExportParams params;
params.allow_custom_ops = false;
params.enable_select_tf_ops = true;
string output;
const auto ops_by_type = BuildOperatorByTypeMap();
auto status = Export(input_model_, &output, params, ops_by_type);
EXPECT_EQ(
status.error_message(),
"We are continually in the process of adding support to TensorFlow Lite "
"for more ops. It would be helpful if you could inform us of how this "
"conversion went by opening a github issue at "
"https://github.com/tensorflow/tensorflow/issues/"
"new?template=40-tflite-op-request.md\n and pasting the "
"following:\n\nSome of the operators in the model are not supported by "
"the standard TensorFlow Lite runtime and are not recognized by "
"TensorFlow. If you have a custom implementation for them you can "
"disable this error with --allow_custom_ops, or by setting "
"allow_custom_ops=True when calling tf.lite.TFLiteConverter(). Here is a "
"list of builtin operators you are using: ADD, CONV_2D. Here is a list "
"of operators for which you will need custom implementations: "
"MyCustomOp1, MyCustomOp2.");
}
TEST_F(ExportTest, UnsupportedControlFlowAndCustomOpsErrors) {
AddOperatorsByName(
{"Conv", "Add", "Switch", "Merge", "MyCustomOp1", "MyCustomOp2"});
ExportParams params;
params.allow_custom_ops = false;
// The model contains control flow ops which are not convertible, so we should
// check the returned error message.
string output;
const auto ops_by_type = BuildOperatorByTypeMap();
auto status = Export(input_model_, &output, params, ops_by_type);
EXPECT_EQ(
status.error_message(),
"We are continually in the process of adding support to TensorFlow Lite "
"for more ops. It would be helpful if you could inform us of how this "
"conversion went by opening a github issue at "
"https://github.com/tensorflow/tensorflow/issues/"
"new?template=40-tflite-op-request.md\n and pasting the "
"following:\n\nTensorFlow Lite currently doesn't support control flow "
"ops: Merge, Switch. We are working on supporting control flow ops, "
"please see github issue at "
"https://github.com/tensorflow/tensorflow/issues/28485. Some of the "
"operators in the model are not supported by the standard TensorFlow "
"Lite runtime. If those are native TensorFlow operators, you might be "
"able to use the extended runtime by passing --enable_select_tf_ops, or "
"by setting target_ops=TFLITE_BUILTINS,SELECT_TF_OPS when calling "
"tf.lite.TFLiteConverter(). Otherwise, if you have a custom "
"implementation for them you can disable this error with "
"--allow_custom_ops, or by setting allow_custom_ops=True when calling "
"tf.lite.TFLiteConverter(). Here is a list of builtin operators you are "
"using: ADD, CONV_2D. Here is a list of operators for which you will "
"need custom implementations: MyCustomOp1, MyCustomOp2.");
}
TEST_F(ExportTest, QuantizeWeights) {
// Sanity check for quantize_weights parameter.
BuildQuantizableTestModel();
@ -335,7 +452,7 @@ TEST_F(OpSetsTest, BuiltinsAndTfSelect) {
// This test is based on a hypothetical scenario that dilation is supported
// only in Conv version 2. So Toco populates version=1 when dialation
// parameters are all 1, and version=2 otehrwise.
// parameters are all 1, and version=2 otherwise.
class FakeConvolutionOperator
: public BuiltinOperator<ConvOperator, ::tflite::Conv2DOptions,
::tflite::BuiltinOptions_Conv2DOptions> {