problem
Since all data_entry_
's memory space is according to JSON file, different layer’s output may use a same ‘data_entry_’ item.
_run_debug
in debug_runtime.py calls run_individual
at first, then append every output from ‘data_entry_’ to _output_tensor_list
.
In short, if we call _run_debug
and get output from _output_tensor_list
, we may find some layers output has modified by other layers.
how to fix
We can alloc another ‘data_entry_’ for debug runtime, which will save all layers output, when calling run_individual
. Then append it into _output_tensor_list
.