Getting this working on Edge TPU
sinclairnick opened this issue · comments
Hey Sergio,
Good work on the research.
I've been trying to port this to Edge TPU using the TFLite converter and compiler. Although you may not be familiar with this in particular, you may be able to help.
Firstly, did you end up getting the MobileNet version to converge? As I think this would work on Edge TPU.
If not, given my current converter settings,
converter = tf.lite.TFLiteConverter.from_keras_model(model)
# required for edgetpu compatibility
converter.representative_dataset = representative_dataset_gen
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.experimental_new_converter = True
converter.experimental_new_quantizer = True
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8
tflite_model = converter.convert()
I get the error:
RuntimeError: Failed to quantize: <unknown>:0: error: loc("model_1/conv2d_25/Exp"): 'tfl.exp' op requires the same type for all operands and results
when I try to compile the model.
Given your familiarity with the model you created, do you have any ideas as to why this might be occurring?
Thanks in advance