Reduced vs basic model

[barrypitman]

Thanks for the great project! The LPRNet paper talks about the "basic" and "reduced" model:

To improve runtime performance we also modified LPRNet basic by using 2 × 2 strides for all pooling layers. This modification (the LPRNet reduced model) reduces the size of intermediate feature maps and total inference computational cost significantly

Is your model using the basic or reduced architecture? From what I can see, it is the basic architecture, I'm interested to know if you tried the reduced architecture? If I try modifying your maxpool3d layers to maxpool2d, and use 2x2 strides instead of 1x1, 1x2 and 1x2, I can't get the shapes to match up at the end. Just curious to know if this is something that you tried?

[kHarshit]

@barrypitman how did you solve it then?

[123Vincent2018]

@barrypitman can you share the "reduced" Net.py?

[barrypitman]

My changes are here: https://github.com/barrypitman/tensorflow_LPRnet.

Edit - sorry I'm getting my repos mixed up. I don't think I solved this issue (maxpool3d layers to maxpool2d)