researchmm / dbtnet Goto Github PK

I don't understand the code in 'dbt.py', line 216 to 221.
How to get the divisor of 128 and 512? plz

Link for downloading pre-trained model does not work

When is open source code ?

• Thanks you for your work.

• The structure of DBTNet-50 in the paper:
  

• The structure in the source code:
  

  DBTNet/dbt.py

  Line 442 in 40877de

  for i, num_layer in enumerate(layers):

The souce code shows that when stage_index == 3 or stage_index == 4, in other words, the Block3 or Block4 two blocks only have SG(Semantic Grouping Layer) and GB(Group Bilinear Layer) modules.

• But this is different from the structure described in the origin paper.
• I hope the authors can help me to answer the confusion.
• Thank you very much again.

Will there be a pytorch version of this?

Thanks!

hi，Thanks for your share！
but, when I dubug the module GroupConv(nn.Conv2D) like this code, I get this error output: 'symbol' object has no attribute 'diag'. May u tell me, how I can use this module Independently？
`
temp_model = GroupConv(2048, kernel_size=1, use_bias=False, width=224)

temp_model.initialize(mx.init.MSRAPrelu())

temp_input = nd.random.normal(-1, 1, shape=(2,3, 224, 224))

temp_output = temp_model(temp_input)
`
I am an newer for MXNET

why set differentiable=False, will it block the gradient of weight and bias.

• I wonder why the tmp is reshaped to (-1, 128), why is it 128?
• I think it should usually be 2048.

Thanks for your sharing.
I have a problem: why train_data processing is different from val_data processing when they are got through interface of ImageRecordIter.

• Thank you for your work.
• I am very interested with your work, but confused with the position encoding in the convolutional channels.
• In the

  DBTNet/dbt.py

  Line 101 in 40877de

  self.bias = self.params.get('bias',

  I can't find out where to use the self.bias and the self.weight.
• I am looking forward to your answer.