Dmitriy Serdyuk
Dmitriy Serdyuk
Could you clarify, which features are used for the results above? resnet152? And does `fc` stand for ShowTell?
The problem is that when NFS is heavily used, the reading time jumps up to 2-3s per batch. Furthermore, this is not healthy for NFS.
The problem is not the size, the file system sometimes is [not able to tackle many small files](https://serverfault.com/questions/31628/slow-nfs-transfer-performance-of-small-files). So, it is fine for me to read from disk.
Actually, not being able to read data in parallel is a big downside. And it seems that multiprocess access to hdf5 is very tricky. I tried to go different path:...
This should be fixed with #35.
We might consider switching to [Pastalog](https://github.com/rewonc/pastalog). As far as I understand, it doesn't have so much functionality as Bokeh. For example, it doesn't seem to be able to plot images.
It was very straightforward for me: conda is a magical tool, it has anything I needed.
Do you mean that it's painful to install node.js? Pretty much the same amount of effort is needed to install Bokeh.
Another disadvantage comparing to Bokeh is that it dumps everything on the same chart. I'd like to see cost/error rate/gradient norms at the same time on different plots since it...
I was able to install it at the lab machines and having the same issues as you on my mac. It looks pretty cool, take a look at eos4:5008 .