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

Fail the model evaluation if there's no successful TFLite interpreter invocations in all attempts.

Browse Source 
PiperOrigin-RevId: 258908895
This commit is contained in:
Chao Mei 2019-07-18 22:38:27 -07:00 committed by TensorFlower Gardener
parent 69359f86c9
commit 303384cce6
3 changed files with 15 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
tensorflow/lite/tools/accuracy/ilsvrc/imagenet_accuracy_eval.cc
View File

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. limitations under the License.
==============================================================================*/ ==============================================================================*/
#include <cstdlib>
#include <iomanip> #include <iomanip>
#include <memory> #include <memory>
#include <mutex> // NOLINT(build/c++11) #include <mutex> // NOLINT(build/c++11)
@ -178,17 +179,17 @@ int Main(int argc, char* argv[]) {
std::unique_ptr<ImagenetModelEvaluator> evaluator; std::unique_ptr<ImagenetModelEvaluator> evaluator;
if (output_file_path.empty()) { if (output_file_path.empty()) {
LOG(ERROR) << "Invalid output file path."; LOG(ERROR) << "Invalid output file path.";
return 0; return EXIT_FAILURE;
} }
if (num_threads <= 0) { if (num_threads <= 0) {
LOG(ERROR) << "Invalid number of threads."; LOG(ERROR) << "Invalid number of threads.";
return 0; return EXIT_FAILURE;
} }
if (ImagenetModelEvaluator::Create(argc, argv, num_threads, &evaluator) != if (ImagenetModelEvaluator::Create(argc, argv, num_threads, &evaluator) !=
kTfLiteOk) kTfLiteOk)
return 0; return EXIT_FAILURE;
std::ofstream output_stream(output_file_path, std::ios::out); std::ofstream output_stream(output_file_path, std::ios::out);
if (!output_stream) { if (!output_stream) {
@ -210,7 +211,10 @@ int Main(int argc, char* argv[]) {
absl::make_unique<CSVWriter>(columns, &output_stream)); absl::make_unique<CSVWriter>(columns, &output_stream));
evaluator->AddObserver(&results_writer); evaluator->AddObserver(&results_writer);
LOG(ERROR) << "Starting evaluation with: " << num_threads << " threads."; LOG(ERROR) << "Starting evaluation with: " << num_threads << " threads.";
evaluator->EvaluateModel(); if (evaluator->EvaluateModel() != kTfLiteOk) {
LOG(ERROR) << "Failed to evaluate the model!";
return EXIT_FAILURE;
}
if (!proto_output_file_path.empty()) { if (!proto_output_file_path.empty()) {
std::ofstream proto_out_file(proto_output_file_path, std::ofstream proto_out_file(proto_output_file_path,
@ -220,7 +224,7 @@ int Main(int argc, char* argv[]) {
proto_out_file.close(); proto_out_file.close();
} }
return 0; return EXIT_SUCCESS;
} }
} // namespace metrics } // namespace metrics

4
tensorflow/lite/tools/accuracy/ilsvrc/imagenet_model_evaluator.cc
View File

@ -263,7 +263,7 @@ TfLiteStatus ImagenetModelEvaluator::EvaluateModel() const {
&all_okay]() { &all_okay]() {
if (EvaluateModelForShard(shard_id, image_label, model_labels, params_, if (EvaluateModelForShard(shard_id, image_label, model_labels, params_,
&observer, params_.num_ranks) != kTfLiteOk) { &observer, params_.num_ranks) != kTfLiteOk) {
all_okay = all_okay && false; all_okay = false;
} }
}; };
thread_pool.push_back(std::thread(func)); thread_pool.push_back(std::thread(func));
@ -274,7 +274,7 @@ TfLiteStatus ImagenetModelEvaluator::EvaluateModel() const {
thread.join(); thread.join();
} }
return kTfLiteOk; return all_okay ? kTfLiteOk : kTfLiteError;
} }
} // namespace metrics } // namespace metrics

5
tensorflow/lite/tools/evaluation/stages/tflite_inference_stage.cc
View File

@ -140,7 +140,10 @@ TfLiteStatus TfliteInferenceStage::Run() {
auto& params = config_.specification().tflite_inference_params(); auto& params = config_.specification().tflite_inference_params();
for (int i = 0; i < params.invocations_per_run(); ++i) { for (int i = 0; i < params.invocations_per_run(); ++i) {
int64_t start_us = profiling::time::NowMicros(); int64_t start_us = profiling::time::NowMicros();
interpreter_->Invoke(); if (interpreter_->Invoke() != kTfLiteOk) {
LOG(ERROR) << "TFLite interpreter failed to invoke at run " << i;
return kTfLiteError;
}
latency_stats_.UpdateStat(profiling::time::NowMicros() - start_us); latency_stats_.UpdateStat(profiling::time::NowMicros() - start_us);
} }