c2board

A hacked-up visualization tool for caffe2. Specifically, it dumps the computation graph and the training statistics of caffe2 into a tensorboard compatible format. Once it starts dumping, you can use tensorboard to visualize the results.

Prerequisites

• Caffe2.
• Tensorboard. The code is meant to be standalone, so that you do not need to import both tensorflow and caffe2 at the same time -- unexpected behaviors can occur. However, tensorboard needs to be installed somewhere (like in another conda environment) for visualization, otherwise the dumped information is not useful by itself.
• Python libraries:

  # for conda users
  conda install protobuf
  # or for ubuntu: sudo apt-get install protobuf-compiler libprotobuf-dev
  pip install json numpy Pillow six threading

Installation

1. Clone the repository

   git clone https://github.com/endernewton/c2board.git

2. Make and install it locally

   cd c2board
   make

Usage

• For graph visualization, you can follow demo_graph.py. The main function to call is writer.write_graph, which accepts either a CNNModelHelper, or Net, or NetDef object to visualize.
• Training statistics can be divided into two types:
  • First, for scalars, they are usually loss, current iteration pre-computed, therefore we can directly store them without any additional effort. The relevant function to call is writer.write_scalars(dict, iter) where dict is the dictionary of the scalars, and iter the current iteration.
  • Second, for histograms and images, we need to call the workspace.FetchBlob function to fetch the values. For those, we first use writer.append_histogram(name) or writer.append_image(name) to build the list of blobs we are interested in when building up the graph. Then, during training we only need to call writer.write_summaries(iter) and the underlying method will take care of fetching blobs, computing histograms, etc.
• When it starts dumping, you can just use the normal tensorboard command line to visualize things.

Screen shots

These screen shots are taken when training a detector with detectron.

References

• tensorboard-pytorch to make tensors flow without tensorflow.
• caffe2-native tensorboard to visualize graphs.