07-17-2019 11:42 PM
I try to use decent_q to quantize faster rcnn.And the input node is Placeholder,the output nodes are resnet_v1_50/cls_score/weights and so on.
But when I finished the calibration,there was the mistake in quantize_utils.cc line 357.But I can't find the file,what's the problem? If the input node was wrong?
07-19-2019 08:31 AM
@tongtyr I am surprised to see this message for op: "Placeholder" as typically "Placeholder" is the input node to the graph. Are you able to attached your pb file or an excerpt of the first few layers of the graph from a pbtxt?
One thing that you can do if you haven't already is run decent_q in inspect mode. This will give you an indication of what decent_q believes are the correct input/output nodes. Example:
decent_q inspect --input_frozen_graph rcnngraph.pb
In general the message that you are seeing appears when one or more layers in the graph is not supported by the quantizer. I generally modify the output node for the quantization script progressively in an effort to determine which layer(s) are causing problems. In this case, the behaviour seems odd if indeed Placeholder is the first layer as you would expect it to be.
07-19-2019 07:07 PM