That's great! Until now, we almost have everything that cuda-convnet provided and many more that it did not.

Considering the work in #142, this is better placed in tools/extra and renamed to something like visualize_conv_layer.py.

Pinging @longjon who's been working on related things.

Please rebase on the latest dev and file this in tools/extras or work it into the imagenet example at examples/imagenet for merge. Thanks!

It would be nice to follow the docstring convention of the existing Python code (i.e., omit the first blank line), and more generally to follow PEP8 / the Google Python style guide (e.g., stick to an 80 character line width).

I'm also slightly concerned about reading the protobuf file using the Python protobuf wrapper -- even though it ought to be just as good as in C++, in practice it's slower, and can fail for very large files (including Alex-sized nets). So, I'd prefer if this code used pycaffe directly to read the net (i.e., by constructing a CaffeNet), and then used that interface to access the parameters; this would also mean this code wouldn't care about the underlying file format. (On the other hand, there is something to be said for being able to examine the net without any compiled dependency; but I think that will be more frustrating than it is worth.)

I second Jon's suggestions.

I would like to point out my earlier visualization codes at decaf:

https://github.com/UCB-ICSI-Vision-Group/decaf-release/blob/master/decaf/util/visualize.py

which should have almost everything you need. A visualization in the decaf
days is at:

http://nbviewer.ipython.org/github/UCB-ICSI-Vision-Group/decaf-release/blob/master/decaf/demos/notebooks/lena_imagenet.ipynb

Yangqing

On Wed, Feb 26, 2014 at 2:36 AM, Sergey Karayev [email protected]:

I second Jon's suggestions.

Reply to this email directly or view it on GitHubhttps://github.com//pull/158#issuecomment-36112008
.

Thanks for the input everyone. I'll pep8-ify the code and move it to tools/extra.

@Yangqing nice! I hadn't seen the decaf visualizations yet.

@longjon have you had bigger problems with the python protobuf. For me everything worked smoothly so far. That is the imagenet model loaded really quickly and I didn't experience any crashes so far.

Indeed, it's really slow for me. This is using protoc version 2.5.0. Do you see different behavior?

In [1]: from caffe.proto import caffe_pb2

In [2]: net = caffe_pb2.NetParameter()

In [3]: with open('/u/vis/jonlong/alexnet_train_iter_470000') as f:
   ...:     data = f.read()
   ...:     

In [4]: %time net.ParseFromString(data)
CPU times: user 2min 55s, sys: 1.12 s, total: 2min 56s
Wall time: 2min 56s

In [5]: from caffe import pycaffe

In [6]: %time pycaffe.CaffeNet('/u/vis/jonlong/caffe/examples/imagenet_deploy.prototxt', '/u/vis/jonlong/alexnet_train_iter_470000')
CPU times: user 1.26 s, sys: 313 ms, total: 1.57 s
Wall time: 1.83 s
Out[6]: <caffe.pycaffe.CaffeNet at 0x63da680>

I too have had terrible luck with python protobuf support.

I recommend reading the params through the python net interface: net.params()[idx].data for param blob at index idx. You might model your contribution after Yangqing's decaf example, which is a nice illustration of the filters and responses.

Python protobuf is quite slow indeed, esp with large protobuf values. So
for large files you might want to use the C interface unless it's easier to
code up in python - balance between coding time and run time :)

There is a fast python protobuf (essentially wrapper around C) but you may
need to install it separately, AFAIK it does not come with pip.

Yangqing

On Thu, Feb 27, 2014 at 4:09 PM, Evan Shelhamer [email protected]:

I too have had terrible luck with python protobuf support.

I recommend reading the params through the python net interface. You might
model your contribution after Yangqing's decaf example, which is a nice
illustration of the filters and responses.

Reply to this email directly or view it on GitHubhttps://github.com//pull/158#issuecomment-36307978
.

@tdomhan thanks for the visualization proposal, but due to the implementation details cited in this thread, @longjon will port the DeCAF visualization demo in #207.

no worries. I'm looking forward to @longjon 's visualizations then.