ndrplz / small_norb Goto Github PK

I was wondering if you were able to verify the azimuth and elevation angles.
I changed the export routine to include the azimuth and elevation integers in the filename.

image_lt_path = join( split_dir, '{:06d}_{}_{:02d}_{}_{}_lt.jpg'.format(
i, category, instance,norb_example.elevation,norb_example.azimuth))

Then I looked at all images with the same azimuth and elevation and eye.

feh *airplane_*_5_30_lt.jpg

There seemed to be quite a lot of variation in pose both within class and across classes.
Animal seemed more consistent. Airplane seemed to have large inconsistencies...
For example, these are all labeled elevation 5, azimuth 30 eye left:

Hi,

first of all thank you for your repository. Now also TF Datasets offers smallnorb dataset, but due to the poor results I obtained, I ended up with your work.

I have two very simple questions:

• I made the following processing function to create X and y tensors. Is it correct or there is a better way?

def create_dataset_tensors(split):
       X_train = np.empty((dataset.n_examples * 2, 96, 96))
       y_train = np.empty((dataset.n_examples * 2))
       index = 0
       for i in tqdm(range(dataset.n_examples)):
           X_train[index] = dataset.data[split][i].image_lt
            y_train[index] = dataset.data[split][i].category
            index += 1
            X_train[index] = dataset.data[split][i].image_rt
            y_train[index] = dataset.data[split][i].category       
            index += 1
       return X_train[...,None], y_train

• I use standardization, normalization, data augmentation, random cropping, resizing and more, but also with small models I have a huge overfitting problem. Do you have any sugggestions?

Andrea, thank you for this code, it is incredibly useful for my project. However, scipy.misc.imsave is now deprecated and instead I think we should use either scipy.misc.imwrite or imageio.imwrite.