timzaman / dagstudio Goto Github PK

This is needed for caffe.
For example the 'Accuracy' node needs input data and a label as reference. These are two input connections, that need to go to input ports [0] and [1] repectively. This might be problematic to obtain.

From a caffe point of view, this is easy to obtain if this can be solved: BVLC/caffe#4712 . If this cannot be solved, I will have to crawl through the caffe code and obtain and input all the top and bottom i/o's by hand. This will also mess up the dynamic .proto parsing, as I will add things that are static.. sigh..

Tim: Any interest to convert this into a workflow studio - including conditions and such?

Technically I guess that makes that part of the graph cyclical, but the graph probably couldn't care less.
In a Caffe networks this looks like:

(...)
layer {
  name: "relu1"
  type: "ReLU"
  bottom: "ip1"
  top: "ip1"
}
(...)

[asas]

Although this cyclical behavior is supported since #1, this ^ isn't pretty.

Be able to switch frameworks with a (examplatory) toggle.

The layers and their args should preferably be formatted to .json.

Listing Caffe Layers

• Through python and BVLC/caffe#2930 (from facebookarchive/caffe2#42 (comment)); Example .json output: https://github.com/TimZaman/dagstudio/blob/master/tests/caffe-all-layers.json

Listing Caffe args

• By protobuf introspection, f.e. https://developers.google.com/protocol-buffers/docs/reference/python/google.protobuf.descriptor-module from (facebookarchive/caffe2#42 (comment))
• By parsing the .proto with nlf/node-proto2json#1 ;
  Example .json output: https://github.com/TimZaman/dagstudio/blob/master/tests/caffe.json

Listing torch layers is pretty easy by listing the contents of nn for example.
Any recommendations on getting the arguments of the constructor of the layers?

Should be straightforward using the work done in NVIDIA/DIGITS#1009

Requires #4, automatic node and edge placement because the .prototxt does not have visual (x,y) coordinate system information.

This is automatically formatted:

While lines are running behind other lines:

See title.

Should structure the graph and snap the nodes in a coordinate system to some kind of grid so that the DAG looks clean.
This is especially useful when a graph is imported that has no 'x,y' structure yet.

Interesting discussion about this on d3/d3#349

Should be straightforward using the work done in NVIDIA/DIGITS#1009

Requires #4, automatic node and edge placement because the .lua does not have visual (x,y) coordinate system information.

Especially useful when you've created something by hand and things get messy.
Can be done with ease after #4. Looks like: