Hi,
I have downloaded NNlib from https://source.codeaurora.org/quic/hexagon_nn/nnlib/ and compiled the test app application. I had to convert the .bmp file to .dat using the python script and finally, it is running on the DSP (Open Q820).
The issue with the testapp in nnlib is that, it is not a complete developed neural net, the results are random, the code in the `test/graphinit_small.c`:
'73 #define APPEND_CONST_NODE(ID,...) if (hexagon_nn_append_const_node(nn_id,ID,__VA_ARGS__) != 0) \
74 printf("node %d returned nonzero\n",ID); else printf("const node %d success\n",ID)
75 #define APPEND_NODE(ID,...) if (hexagon_nn_append_node(nn_id,ID,__VA_ARGS__) != 0) \
76 printf("node %d returned nonzero\n",ID); else printf("node %d success\n",ID)
77 void init_graph(int nn_id) {
78 APPEND_NODE(0x1024a,OP_INPUT,NN_PAD_ANY,inputs_for_1024a,1,outputs_for_1024a,1);
79 APPEND_NODE(0x10300,OP_Nop,NN_PAD_ANY,inputs_for_10300,1,outputs_for_10300,1);
80 APPEND_NODE(0x1044d,OP_OUTPUT,NN_PAD_ANY,inputs_for_1044d,1,NULL,0);'
You see the defined network is completely NULL, so the ImageNet classification we get by running the code is useless.
Can we have a minimun working network (any random network with 3-4 layers) that we start from? The documentation is poor on how to APPEND_NODE() and I don't figure out the 0x codes need to be used with that to connect the layers to one another. Appreciate any help.
-- Amir
Thank you!
-- Amir