Xuan Lin

Results 63 comments of Xuan Lin

how to use gen_hard_example.py to generate R-net training data?

#4

facial landmark regression

@Rainy000 Try clip_gradient while initializing optimizer in train.py

facial landmark regression

@Rainy000 Sorry, I mistook deviation for derivative, so clip gradient is not a solution. Did you follow the ratio the author mentioned in the paper? I think raising the importance...

facial landmark regression

@Rainy000 You can test the celebA dataset using your Rnet model and compute normalized error of predicted landmarks, then choose those with big deviations to be the training set of...

facial landmark regression

@Rainy000 Exactly.

wider face annotations format (x1,y1, width, height)?

@hkdqliu That's right, but I had it transformed. The scripts are located in prepare_data/wider_annotations

where block back propagate?

core/negativemining.py, line 25 and line 42 find out the valid indexes for cls and bbox. In backward, only gradient on valid indexes are assigned 1 otherwise 0.

How to save the trained model

No, it's done in epoch end callback inside mod.fit()

Where is input size of Pnet specified?

@huynhthedang PNet is fully convolutional so there's no restriction on the input size

I only find pnet-0016.params. Where is the pnet-0000.params?

I trained 16 epochs so pnet-0016.params is the final model. The codes to init the params are in example/train.py