Styjun

In this tutorial (creating new model for object detection), it is mentioned at the middle as

"We typically initialize the weights of this feature extractor using those from the Slim Resnet-101 classification checkpoint, and we know that images were preprocessed when training this checkpoint by subtracting a channel mean from each input image. Thus, we implement the preprocess function to replicate the same channel mean subtraction behavior."

Now I am trying to load pretrained model for MobileNet_v1_1.0_224 at this page.

I checked all variables from loaded checkpoint and those variables required to initialize in training FasterRcnn. Loaded checkpoint has more varaibles than those needed.

For example,
I need to initialize this variable 'FirstStageFeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/beta'.

But in loaded variables have

'MobilenetV1/Conv2d_0/BatchNorm/beta/ExponentialMovingAverage': [32]
'MobilenetV1/Conv2d_0/BatchNorm/beta/RMSProp_1': [32],
'MobilenetV1/Conv2d_0/BatchNorm/beta': [32],

My queries are

(1)So for me, it is enough to use the last one 'MobilenetV1/Conv2d_0/BatchNorm/beta' to initialize to 'FirstStageFeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/beta'.

Is it correct?

(2)What ExponentialMovingAverage and RMSProp_1 are for?

(3)Then how FirstStageFeatureExtractor and SecondStageFeatureExtractor are separated in FasterRcnn in Tensorflow?

(4)Those variables initialized use initialized weights, for those variable not initialized will use Xavier initializer according to config file, is it true?

initializer 
 variance_scaling_initializer 
 factor: 1.0
 uniform: true
 mode: FAN_AVG
 

EDIT:

Then for the variable MobilenetV1/Conv2d_12_pointwise/depthwise_weights shape=(3, 3, 512, 1), I can't find exact variable.
Those closer are

'MobilenetV1/Conv2d_12_depthwise/depthwise_weights': [3, 3, 512, 1],
'MobilenetV1/Conv2d_12_depthwise/depthwise_weights/RMSProp': [3, 3, 512, 1],
'MobilenetV1/Conv2d_12_depthwise/depthwise_weights/ExponentialMovingAverage': [3, 3, 512, 1],
'MobilenetV1/Conv2d_12_depthwise/depthwise_weights/RMSProp_1': [3, 3, 512, 1],

So I used weights of variable 'MobilenetV1/Conv2d_12_depthwise/depthwise_weights': [3, 3, 512, 1], from loaded checkpoint to assign to
MobilenetV1/Conv2d_12_pointwise/depthwise_weights shape=(3, 3, 512, 1)

Yes what I did was correct. What variables are needed to initialize and what variable are available from loaded checkpoint can be checked from variable_names_map. From there, select the variable and initialize to further fine tuning.

Need a bit of modification to the Tensorflow's code mainly at utils/variables_helper.py file.

What to be in FirstStage and SecondStage of FasterRCNN are decided at faster_rcnn_mobilenet_v1_feature_extractor.py