model parameters

[mglezer]

I am trying to understand the values in models/deploy.prototxt. Per your paper, for the first two phases of training, you train at 384x384 input resolution, and then increase to 768x768 during the last phase of training. It is my understanding that the step values for the priorboxes are essentially the 384/(width of input feature map). By my calculations, the feature map inputs for the priorbox layers for input resolution 384x384 are (48x48, 24x24, 12x12, 6x6, 4x4, 2x2). Accordingly, I would think the first four "step" sizes are (8, 16, 32, 64, 96, 192), but in deploy.prototxt they are (8, 16, 32, 64, 100, 300). Where did 100 and 300 come from?

Also, why is min_size 30 for both of the first two priorbox layers? I would think min_size for the second priorbox layer would be 60, following the pattern of the rest...

One final question: what parameters did you use for the Adam optimizer, in terms of beta1, beta2, and epsilon, when training on SynthText?

Thank you so much for your help!

[zxDeepDiver]

My thought for the first question: The step values of 100 and 300 are consistent with the original SSD, which is for better receptive field aligning and may be less important for the detector.

[MhLiao]

@mglezer The step parameter follows the original SSD, just as @zxDeepDiver said. For Adam, beta1 is the momentum (0.9), beta2 and epsilon are the default values (0.999 and 10e-8 respectively)

[mglezer]

@MhLiao thanks so much for the quick response!! One more quick training question: during training Stage 1 and Stage 2, after pre-training on SynthText, do you train all of the layers? Or do you freeze some?

[MhLiao]

I do not freeze extra layers. It's the same between pre-training and training.

[qnkhuat]

Hi. Seems like u are successfull installed. Can u give me ur os system, torch ,cudnn,cuda versions. I'm struggling with install crnn. Many Thanks

[mglezer]

I haven't tried installing crnn--I am actually trying to implement TextBoxes++ in TensorFlow