Selecting hypertext parameters for training API analysis of tensor flow objects

Question

I am creating an object discovery pipeline based on the recently released tensor flow object discovery API. I am using arXiv as a guide. I want to understand below for training on my own dataset.

• It is unclear how they chose the training speed graphs and how this will change depending on the number of GPUs available for training. How does the learning rate schedule change depending on the number of GPUs available for training? The document mentions 9 GPUs. How to change the learning speed if I want to use only 1 GPU?

• The released training sample configuration file for Pascal VOC using Faster R-CNN has an initial training rate = 0.0001. This is 10 times smaller than the original Faster-RCNN page . Is this due to the assumption of the number of GPUs available for training, or for another reason?

• When I start training at the COCO checkpoint, how should training losses be reduced? Looking at the tensogram, my lessons are low - from 0.8 to 1.2 per iteration (with a lot size of 1). The various tensor losses are shown below. Is this the expected behavior?
  

Answer 1

1 2: , SGD ~ 10 GPU. ( , Cloud ML Engine, ). , . GPU, , , , , .

3: , , . , , eval, . mAP , , , "".

, .