tensorflow-workspace to support K210model
tensorflow-workspaceuses a .c file for model parameters and recent
kmodeloutput. Is there any plan to update
tensorflow-workspaceto support kmodel? Thanks.
@nathan Awesome, thanks for the reply! Keep up the good work!
@nullspace Different kinds of operations will be recognized and handled automatically. If an operation can be handled by KPU, then KPU will handle it. Otherwise, the operation will be handled by CPU. You don't bother to differentiate these operations, they will be handled properly as long as can be converted by
@nathan Hi, thoughts? Thanks.
@nathan Thanks a lot!
One question. Since the chip does not support 7x7 pooling and mobilenet v1 has a 7x7 pooling at the end (https://github.com/kendryte/tensorflow-workspace/blob/master/mobilenetv1/models/mobilenet_v1.py#L148-L149), how does this work with current kmodel?
More details. In the old method with .c model file, it runs the last layer (7x7 pooling) out side of kpu like https://github.com/kendryte/tensorflow-workspace/blob/master/mobilenetv1/models/mobilenet_v1.py#L148-L149. Now with kmodel,
ai_stepis going, so how does this last 7x7 pooling work with kmodel?
Thanks a lot!
@nullspace We have just made a new commit and in
tensorflow-workspace/K210codekmodel is used instead of .c file. Please git pull to update.