sergeyk / vislab

Set of modules and datasets for visual recognition.

Please check out the documentation here.

vislab's Issues

Speed considerations for loading Caffe features

It seems the main method to load Caffe's CNN features is from the caffe function in vislab / features / misc.py. I have several questions on performance:

1. Why does the Caffe method in misc.py use CPU instead of GPU?
2. Does your Caffe method do the 256x256 resizing or should I do that as a pre-proccessing step?
3. What is the downside to using extract_features.bin from the Caffe project to extract the features for vislab?
4. Just for reference to see if I'm on track - what is the expected time to extract features, from ~50,000 images?

Thank you Sergey!

Max Across Image Crops? [Just a Question - Not a Bug]

Hi Sergey,

A basic question here. In /vislab/features/misc.py, line 60+, you wrote:

       # First, run the network fully forward by calling predict.
       # Then, for whatever blob we want, max across image crops.
        net.predict([caffe.io.load_image(image_filename)])
        feats.append(net.blobs[layer].data.max(0).flatten())

I'd like to understand why you chose the "max" crop. Firstly, what does it mean to be the "max across image crops"? In the Caffe feature visualization tutorial it indicates one can just select the center crop. In other discussions people have mentioned that one should "average the crops." Would you mind to compare your "max across image crops" with the aforementioned alternatives and why is it the best way to go and / or pros/cons?

Thanks!

run_DARTS fails on predict stage

Hi, Sergey.

I have one question regarding this part of code in run_DARTS.m:

[~, ~, val_dec_values] = predict(val_labels, val_data.betas, model, '-b 0', 'col');

while executing predict it fails with:

predict(val_labels, val_data.betas, model, '-b 0', 'col');
Error using predict (line 84)
Systems of double class cannot be used with the "predict" command. Convert the system to an identified model first, such
as by using the "idss" command.

model has the following structure

>> model

model = 

    Parameters: 2
      nr_class: 57
    nr_feature: 100000
          bias: -1
         Label: [57x1 double]
             w: [57x100000 double]

>> numel(val_labels)

ans =

        2850
>> numel(val_data.betas)

ans =

   285000000

What exactly should be done to avoid this issue?

Thanks in advance. Regards. Taras.

ImportError: No module named vislab

I try to download ava dataset by "python datasets/ava.py", but the error information is shown as "ImportError: No module named vislab". Where can i find the vislab module.
Thanks!

How to load the features from a dataset end-to-end?

Hi Sergey - I noticed in /vislab/feature.py there is a compute command which appears to load the features. I'd just like to confirm the end-to-end procedure of adding my own dataset and getting features into the db.

So - follow the "Adding your own" section of the dataset documentation:

1. Add the file vislab/datasets/your_dataset.py that will contain a function to load a pandas.DataFrame with:
2. unique string-based index, with name image_id
3. image_url or image_filename in columns
4. a column for whatever boolean label you care about (does this matter if I am just extracting neural codes?)
5. Modify DATASETS in vislab/dataset.py to map a name to your new function.
6. Then finally, call feature.py compute from command line? (is feature.py the main entry point?)