Non square images representation

[ivandirtech]

Hi, Input blob in my deploy prototxt has 224x288 dimensions (bvlc-googlenet). 2 things are bothering me:

1. By default the image is resized to a square one -> 224x224 and it influences the classification results of the network. After hitting '0' to cancel reshaping, things work properly.
2. Even after I disable input image reshaping, in GUI it is still shown as a square -> crop happens. Is there any way the GUI can be fixed to show the whole image and not the cropped one?

[arikpoz]

@ivandirtech

I've substantially refactored the tool, fixed many bugs, among them the tool ability to work on rectangular inputs instead of squares

Latest version is in https://github.com/arikpoz/deep-visualization-toolbox

[ivandirtech]

@arikpoz I glad you answered, because I am already using your improved version (thx for the great job BTW!) so the issue still happens.

[arikpoz]

I recall working with non rectangular without issues, other than the ones I've fixed

[FredaZhang338]

@arikpoz Thanks for your sharing. And I have a question that when I am trying to visualize the VGG16 trained on the Faster R-CNN, I modified the mean.npy, prototxt,labels.txt and weights( I did not rename these files, I put them in the demo's folder and keep these the same name with the demo) And then I got the error like this: got module <module 'caffevis.app' from '/home/manning/mn/deep-visualization-toolbox/caffevis/app.pyc'> got app <class 'caffevis.app.CaffeVisApp'> Got settings <module 'settings' from '/home/manning/mn/deep-visualization-toolbox/settings.pyc'> Loaded caffe in GPU mode, using device 0 [libprotobuf ERROR google/protobuf/text_format.cc:274] Error parsing text-format caffe.NetParameter: 471:24: Message type "caffe.LayerParameter" has no field named "smooth_l1_loss_param". E1228 20:01:20.436556 22212 upgrade_net_proto_text.cpp:30] Failed to parse input text file as NetParameter: /home/manning/mn/deep-visualization-toolbox/./models/caffenet-yos/caffenet-yos-deploy.prototxt.processed_by_deepvis Traceback (most recent call last): File "./run_toolbox.py", line 34, in <module> main() File "./run_toolbox.py", line 24, in main lv = LiveVis(settings) File "/home/manning/mn/deep-visualization-toolbox/live_vis.py", line 61, in __init__ app = app_class(settings, self.bindings) File "/home/manning/mn/deep-visualization-toolbox/caffevis/app.py", line 48, in __init__ self.net, self._data_mean = load_network(settings) File "/home/manning/mn/deep-visualization-toolbox/settings_misc.py", line 120, in load_network deduce_calculated_settings_without_network(settings) File "/home/manning/mn/deep-visualization-toolbox/settings_misc.py", line 15, in deduce_calculated_settings_without_network read_network_dag(settings) File "/home/manning/mn/deep-visualization-toolbox/settings_misc.py", line 207, in read_network_dag text_format.Merge(str(proto_file.read()), network_def) File "/usr/local/lib/python2.7/dist-packages/google/protobuf/text_format.py", line 265, in Merge return MergeLines(text.split('\n'), message) File "/usr/local/lib/python2.7/dist-packages/google/protobuf/text_format.py", line 298, in MergeLines _ParseOrMerge(lines, message, True) File "/usr/local/lib/python2.7/dist-packages/google/protobuf/text_format.py", line 229, in _ParseOrMerge _MergeField(tokenizer, message, allow_multiple_scalars) File "/usr/local/lib/python2.7/dist-packages/google/protobuf/text_format.py", line 382, in _MergeField _MergeField(tokenizer, sub_message, allow_multiple_scalars) File "/usr/local/lib/python2.7/dist-packages/google/protobuf/text_format.py", line 356, in _MergeField message_descriptor.full_name, name)) google.protobuf.text_format.ParseError: 471:3 : Message type "caffe.LayerParameter" has no field named "smooth_l1_loss_param". The smooth_l1_loss_param is part of the train.prototxt of my own data, I do not know how to fix this problem , could you please give me some instructions? thanks for your help.

[arikpoz]

Are you using a custom caffe version? Was your model trained with custom caffe?

[FredaZhang338]

@arikpoz My model was not trained here, and the caffe was not installed by me, I am not sure the version. So you mean the problem was caused by the caffe version? Is there any way I can fix the problem? Thanks ~

[FredaZhang338]

@arikpoz I learned that the Faster-RCNN requires its own branch of caffe. This branch includes roi_pooling_layer and its associated parameters.
When I was training a model just followed the installation instructions of Faster RCNN to get the correct branch of caffe. So if I want to visualize the VGG16 trained on the Faster R-CNN, that is I need to add this additional layers and associated parameters? Could you please give me some ideas? Thanks for your help!

[arikpoz]

Either add the changes from your custom caffe to the caffe DeepVis tool uses, or if possible, retrain your faster-rcnn using latest caffe from master, if it contains the relevant roi-pooling layer