search

willard-yuan / CNN-for-Face-Image-Retrieval

:apple:Face image rerieval
90 stars 55 forks source link

Isn't the 36th layer of 2622 dimension? #2

Closed surlogics closed 7 years ago

surlogics commented 7 years ago

Hi, I've noticed your extractCNN.m picks up FC activations from the 36th layer. As far as I see, 36th layer is FC8 and it is of 2622 dimensions. Shouldn't you pick up FC activations from 34th layer (FC7). Or am I see it wrong? Thanks for your help.

willard-yuan commented 7 years ago

@surlogics The MatConvNet and the model may be changed a little. I tested it about a year ago. And the feature dimension is 4096D I extracted. So it's good to choose the layer whose dimension is 4096d.

surlogics commented 7 years ago

OK! Thanks for your help!