Open kungkungpanda opened 5 years ago
@kungkungpanda Hi,
1.-2. Yes, I implemented the same conv-LSTM as in this paper if peephole=0
, except if peephole=1
then I use in peephole: *
- Convolution instead of o
- Element-wise-product (Hadamard product),
so convLSTM is still resizable - can be used with any network input resolution:
random=1
in the last [yolo]
-layer.
But resizing for conv-LSTM layer isn't tested, so the training can be unstable. May be it will be solved later by using special pre-trained LSTM-weights on lstm-Classifier.@AlexeyAB Thanks for your answers! 👍 I have additional questions.
What's the role of "seq_scales" and "sequential_subdivisions" in cfg file? Are the roles of "scales" and "seq_sales" different in the cfg file? Are the roles of "subdivisions" and "sequential_subdivisions" different? Sorry if I ask naive questions.
Many thanks!
@kungkungpanda
What's the role of "seq_scales" and "sequential_subdivisions" in cfg file?
Read wiki (it is in progress): https://github.com/AlexeyAB/darknet/wiki/CFG-Parameters-in-the-%5Bnet%5D-section
Are the roles of "scales" and "seq_sales" different in the cfg file?
Yes
Are the roles of "subdivisions" and "sequential_subdivisions" different?
Yes
@AlexeyAB Thanks for your work.
Can you please guide me to make yolov3-spp_lstm version in tensorflow for inference
I tried but conv_lstm weights can't load it gives error.
@AlexeyAB Great work! Thank you very much for sharing the code.
I have a three questions about your implementation.
Did you implement the same "Convolutional LSTM" proposed in the following document? I want to study Convolutional LSTM using the paper and your code. (http://arxiv.org/abs/1506.04214v1)
Is there any part that is different from the paper?
I would like to do multi-scale training on a network with conv_LSTM. However, the provided code does not resize in training process when it has an conv_LSTM layer. Which part of the code should I fix? I'd appreciate your advice.
Many thanks!