Skip to content

Home

Jump to bottom
bl166 edited this page Jan 23, 2019 · 1 revision

Welcome to the Efficient_Person_ReID wiki!

  1. We are using ResNet and DenseNet for ReID (inherited from the algorithm provided by Dr. Wang), so you probably want to be familiar with those architectures.

  2. In ResNet and DenseNet, channel pruning may result in inconsistency in feature map addition/concatenation. This is a problem so I was doing "pseudo pruning", meaning that the channels to be pruned were merely set to all-zero (as opposed to actually removed from the model). This should make a big difference if the pruning were actually executed.

  3. Investigate parameter quantization implementations. (Our framework is currently in PyTorch, but, last time I checked, PyTorch did not have good support to quantization while Tensorflow did. If that's still true, we may consider move the thing to TF.)

  4. Other interesting random stuff: https://github.com/pytorch/QNNPACK https://www.tensorflow.org/lite/performance/post_training_quantization https://hanlab.mit.edu/projects/amc/papers/amc_eccv.pdf

Clone this wiki locally