search

facebookresearch / maskrcnn-benchmark

Fast, modular reference implementation of Instance Segmentation and Object Detection algorithms in PyTorch.
MIT License
9.29k stars 2.5k forks source link

can add tensorboardx for visualize the training status? #76

Open mmxuan18 opened 5 years ago

mmxuan18 commented 5 years ago

🚀 Feature

current training information only display in console, this is not very anschauung，can add tensorboardx to support loss or feature map 's visualize

fmassa commented 5 years ago

Adding support for TensorBoard would be a good addition, but I'm currently without bandwidth to add it.

Another alternative is to have a script to parse the logs that are printed, and plot them in a jupyter notebook for example (that's what I've been doing).

steve-goley commented 5 years ago

I'm planning to add a visdom plotter. Also the only feature I miss from TF OD API is the concurrent eval loop. That's also on my list.

fmassa commented 5 years ago

@steve-goley what do you mean by concurrent eval loop? Currently, it is possible to evaluate using multiple GPUs, is that what you meant?

steve-goley commented 5 years ago

In TF OD API the user can run an eval process that

  1. Loads the last checkpoint and evaluates validation images (or a subset of them)
  2. Computes AP for the classes
  3. Plots AP in Tensorboard
  4. Visualizes images in Tensorboard
  5. Waits until a new checkpoint is created and repeats the process

It requires that you keep one GPU for eval but it's nice to have some sort of visibility while training. FRCNN loss outputs are hard to use as a gauge for performance plateaus, IMHO, AP over the validation set is much more informative.

fmassa commented 5 years ago

@steve-goley oh, I see, yes, that would definitely be a plus to have!

mmxuan18 commented 5 years ago

@steve-goley can provide any docs describe TF OD API's visualize support mechanism，so may i can add some doc to do this

steve-goley commented 5 years ago

@mlinxiang TensorFlow uses TensorBoard. I haven't dug into the checkpoint structure, yet. I think that a similar capability could be constructed with visdom. There are some advantages and disadvantages to visdom over Tensorboard / tensorboardX. I've found visdom quite easy to use and the API, much like many things FAIR, is quite readable and concise. @fmassa do you have a strong opinion on tensorboardX vs. visdom.

soumith commented 5 years ago

given the rich nature and potential of mask-rcnn visualizations, where a lot of custom visualizations might make sense, Visdom seems better longer-term. Visdom has trivial ways to construct rich custom visualizations.

mmxuan18 commented 5 years ago

has visdom integrate into pytorch?

fmassa commented 5 years ago

I think Visdom works with PyTorch

wangg12 commented 5 years ago

I think using tensorboardX to visualize the accuracies and losses in real time is a good idea.