Hello!
I tried the following model structure:
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
reshape (Reshape) (None, 199, 65) 0
conv1d (Conv1D) (None, 199, 8) 2088
conv1d_1 (Conv1D) (None, 199, 16) 528
max_pooling1d (MaxPooling1D (None, 100, 16) 0
)
conv1d_2 (Conv1D) (None, 100, 16) 1040
conv1d_3 (Conv1D) (None, 100, 32) 2080
max_pooling1d_1 (MaxPooling (None, 50, 32) 0
1D)
conv1d_4 (Conv1D) (None, 50, 32) 4128
conv1d_5 (Conv1D) (None, 50, 64) 8256
global_max_pooling1d (Globa (None, 64) 0
lMaxPooling1D)
dropout (Dropout) (None, 64) 0
y_pred (Dense) (None, 2) 130
=================================================================
Total params: 18,250
Trainable params: 18,250
Non-trainable params: 0
Resulting in following error:
Saving best performing model...
Converting TensorFlow Lite float32 model...
Converting TensorFlow Lite int8 quantized model...
Calculating performance metrics...
Profiling float32 model...
Profiling float32 model (tflite)...
Traceback (most recent call last):
File "/home/tflite_find_operators.py", line 100, in <module>
raise Exception('TensorFlow Lite op "{}" has no known MicroMutableOpResolver method.'.format(key))
Exception: TensorFlow Lite op "REDUCE_MAX" has no known MicroMutableOpResolver method.
Error while finding memory:
Command '['sh', '/home/prepare_tflite.sh']' returned non-zero exit status 1.
Profiling int8 model...
Profiling int8 model (tflite)...
Traceback (most recent call last):
File "/home/tflite_find_operators.py", line 100, in <module>
raise Exception('TensorFlow Lite op "{}" has no known MicroMutableOpResolver method.'.format(key))
Exception: TensorFlow Lite op "REDUCE_MAX" has no known MicroMutableOpResolver method.
Error while finding memory:
Command '['sh', '/home/prepare_tflite.sh']' returned non-zero exit status 1.
Model training complete
Traceback (most recent call last):
File "/app/./resources/libraries/ei_tensorflow/profiling.py", line 328, in profile_tflite_model
subprocess.check_output(['sh', '/home/prepare_tflite.sh']).decode("utf-8")
File "/usr/lib/python3.8/subprocess.py", line 415, in check_output
return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
File "/usr/lib/python3.8/subprocess.py", line 516, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['sh', '/home/prepare_tflite.sh']' returned non-zero exit status 1.
Traceback (most recent call last):
File "/app/./resources/libraries/ei_tensorflow/profiling.py", line 328, in profile_tflite_model
subprocess.check_output(['sh', '/home/prepare_tflite.sh']).decode("utf-8")
File "/usr/lib/python3.8/subprocess.py", line 415, in check_output
return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
File "/usr/lib/python3.8/subprocess.py", line 516, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['sh', '/home/prepare_tflite.sh']' returned non-zero exit status 1.
Taking inspiration from https://stackoverflow.com/a/66971484, I constructed a new model:
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
reshape (Reshape) (None, 199, 1, 65) 0
conv2d (Conv2D) (None, 199, 1, 8) 2088
conv2d_1 (Conv2D) (None, 199, 1, 16) 528
max_pooling2d (MaxPooling2D (None, 100, 1, 16) 0
)
conv2d_2 (Conv2D) (None, 100, 1, 16) 1040
conv2d_3 (Conv2D) (None, 100, 1, 32) 2080
max_pooling2d_1 (MaxPooling (None, 50, 1, 32) 0
2D)
conv2d_4 (Conv2D) (None, 50, 1, 32) 4128
conv2d_5 (Conv2D) (None, 50, 1, 64) 8256
global_max_pooling2d (Globa (None, 64) 0
lMaxPooling2D)
dropout (Dropout) (None, 64) 0
y_pred (Dense) (None, 2) 130
=================================================================
Total params: 18,250
Trainable params: 18,250
Non-trainable params: 0
Which is identical to the 1D counterpart. In another use-case, I had a similar problem in my own scripts (having errors with 1D model including GlobalMaxPooling), but I have been able to full-integer quantize and convert 2D-model to .tflite. I have also been able to successfully deploy it to Nano 33 BLE Sense.
However, Edge Impulse gives the same error as for the 1D model. I guess though that the models convert successfully, but when the profiling starts (Profiling float32 model (tflite)), some operation is missing (REDUCE_MAX). The problem is probably in the GlobalMaxPooling.
Anyway, I think it would be helpful to be able to use the GlobalMaxPooling-layer in Edge Impulse. 
Well, at least because of it, I didn’t get the measures for Peak RAM and Flash Usage.