XiaoMi / mace-models

Mobile AI Compute Engine Model Zoo
Apache License 2.0
371 stars 73 forks source link

DeepLab Test #11

Closed ilous12 closed 4 years ago

ilous12 commented 5 years ago

Hi. nice work

I will try to run my own deeplabv3plus model on mace framework. So I want to know what was model changes for mace.

I saw your description about deeplabv3plus model

  • input_tensors: sub_7 -output_tensors: ResizeBilinear_2 -input_shapes: 1,513,513,3 -output_shapes: 1,65,65,21

I understood original deeplab v3 plus was below.

  • input_tensors: ImageTensor -output_tensors: SementicPredictions -input_shapes: 1,513,513,3 -output_shapes: 1,513,513,21

What do you fix for mace framework? Can you tell me about that? or your training code.

if you use "transform_graph", please share me command.

llhe commented 5 years ago

@ilous12 Here is the description: https://github.com/XiaoMi/mace-models/tree/master/deeplab-v3-plus

Part of the model (sub_7 ~ ResizeBilinear_2) is converted for MACE.

ilous12 commented 5 years ago

Thanks for reply. Deeplab had 21 label and I tried to run your model. But I got invalid result. Is It working valid? I only saw background label.

hoonkai commented 5 years ago

@llhe Do you know why the output shapes are 1,65,65,21 rather than 1,513,513,21?

lydoc commented 5 years ago

@hoonkai We use "ResizeBilinear_2" node which has such a shape (1,65,65,21) as output.

hoonkai commented 5 years ago

@lydoc But if the output layer is 65x65x21, how should the 512x512 mask bitmap be built? Or is the result meant to be a 64x64 mask bitmap?

lydoc commented 5 years ago

@hoonkai You can resize the output size to 512 * 512.