keithm-xmos
commented
4 years ago This is caused by the builtin CONV_2D operator calling AllocatePersistentBuffer() in Prepare(). The next call to RequestScratchBufferInArena() will raise the error here:
Closed keithm-xmos closed 4 years ago
keithm-xmos
commented
4 years ago This is caused by the builtin CONV_2D operator calling AllocatePersistentBuffer() in Prepare(). The next call to RequestScratchBufferInArena() will raise the error here:
keithm-xmos
commented
4 years ago I have inquired about this being a bug or design issue in the builtin CONV2D & CONV_2D_DEPTHWISE operators. If this is not a bug or design issue, then I do have a workaround. We will need to override these builtin ops with an implementation that does not allocate persistent buffers in Prepare. This is not a horrible solution. And, it is an easy fix.
keithm-xmos
commented
4 years ago According to Pete Warden, this is a bug and they have a fix pending review. Look for it to be merged in a few days.
keithm-xmos
commented
4 years ago TFLu fix has been merged, see this commit: https://github.com/tensorflow/tensorflow/commit/59d177d9acabe8e70bc33e554a364d2620bc6999
keithm-xmos
commented
4 years ago Still have lingering TFLu bug. See: https://github.com/tensorflow/tensorflow/issues/42964
keithm-xmos
commented
4 years ago TensorFlow devs are on it. See PR: https://github.com/tensorflow/tensorflow/pull/43109
keithm-xmos
commented
4 years ago
Traceback (most recent call last): File "./xformer.py", line 71, in
analyze.print_report(tflite_output_path)
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/analyze.py", line 106, in print_report
tensor_arena_size = calc_arena_size(model_content)
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/analyze.py", line 82, in calc_arena_size
[logger.info(line) for line in interpreter.get_allocations().split("\n")]
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/xcore_interpreter.py", line 245, in get_allocations
self._verify_allocated()
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/xcore_interpreter.py", line 232, in _verify_allocated
self.allocate_tensors()
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/xcore_interpreter.py", line 255, in allocate_tensors
self._check_status(lib.allocate_tensors(self.obj))
File "/home/kmoulton/repos/hotdog/ai_tools/tflite2xcore/tflite2xcore/xcore_interpreter.py", line 237, in _check_status
raise RuntimeError(self._error_msg.value.decode("utf-8"))
RuntimeError: Internal error: AllocateFromTail can not be called between two RequestScratchBufferInArena calls.
Node XC_conv2d_deep (number 49f) failed to prepare with status 1