Aborted (core dumped) tflite model maker int8 quantised model

TensorFlow General Discussion
,
1

Hello,

I tried to train EfficientDetLite2 detection model on custom data following the tutorial. I changed some parameters as follows:

train_data, validation_data, test_data = object_detector.DataLoader.from_csv('/home/smurf/efficientDetTrainingDataFinal.csv')


# spec = object_detector.EfficientDetSpec(
#   tflite_max_detections = 301,
#   model_name='efficientdet-lite2', 
#   uri='https://tfhub.dev/tensorflow/efficientdet/lite2/feature-vector/1', 
#   hparams={'max_instances_per_image': 301, 'autoaugment_policy' : None, 'optimizer':'adam' , 'learning_rate' : 0.008, 'lr_warmup_init': 0.0008},
#   epochs = 120)
# model = object_detector.create(train_data, model_spec=spec, batch_size=16, train_whole_model=True, validation_data=validation_data)

# model.export(export_dir='.', tflite_filename = "lite2.tflite")
# model.export(export_dir='.', tflite_filename = "lite2_fp16.tflite", quantization_config = config.QuantizationConfig.for_float16())
# model.export(export_dir='.', tflite_filename = "lite2_dynamic.tflite", quantization_config = config.QuantizationConfig.for_dynamic())

When the training is done, i wanted to test the models on an image both “lite2_fp16.tflite” and “lite2_dynamic.tflite” work fine and give me the desired results. but int8 quantized model crashes.

it crashes right when I call interpreter.allocate_tensors() and the error is

Aborted (core dumped)

I tried loading again from checkpoint and training again for couple of epochs and saving, it saves without any error, but crashes when I want to do inference.

The funny part is when I train the model for few epochs like 1 to 8 the conversion is done correctly and the model works

3

Hi @ilyas_aroui ,

Int8 quantized model only crashing while the other two quantized models are working, might be due to the model weights were updating to lower values. So please inspect the weights at each layer, also check the quantization range and correctness of representative data. For further assistance, please provide us reproducible code to resolve the issue.

Thank You