patrykchrabaszcz / imagenet32_scripts Goto Github PK

Hi,
What batch size did you set on these experiments? Thanks

So when I download the data from the website, I get a tarfile back. I've tried your unpicle code and that won't work when I provide the tar file path. How exactly should I process this data in order to get to the images and labels? I am not following the scripts as I don't see any code where you process or unzip tarfiles.

Hi there,

I would like to use the downsampled version of ImageNet. I have downloaded the val+train files from https://image-net.org/download-images.php, the "npz format" versions. Then I ran "unzip Imagenet64_val_npz.zip" which resulted in a file called "val_data.npz".

I use python 3.6.8, and trying to load the file with the provided code results in the following error:

def load_data(input_file):

    d = unpickle(input_file)
    x = d['data']
    y = d['labels']

    x = np.dstack((x[:, :4096], x[:, 4096:8192], x[:, 8192:]))
    x = x.reshape((x.shape[0], 64, 64, 3))

    return x, y

def unpickle(file):
    with open(file, 'rb') as fo:
        dict = pickle.load(fo)
    return dict

x, y = load_data("./val64/Imagenet64_val_npz/val_data.npz")
Error: "UnpicklingError: A load persistent id instruction was encountered, but no persistent_load function was specified."

Using x, y = np.load(input_file) works, but the labels seem to be wrong, and the files are also unordered which is fine though. The shapes of x and y are (50000, 64, 64, 3) and (50000,), respectively, so the file downloading, unzipping and loading seems to work (with np.load not with pickle).

These are the first five images in x with the labels y (and the human readable meanings taken from here):

-> Does one have to manipulate the labels y to make them "correct" again?

Thank you for your help :)

Hi,

are you aware that the validation set of 49999 images downloaded from here:
http://image-net.org/small/valid_32x32.tar

has a lot of duplicate images? Reproduction of a few examples:
02273.png and 42263.png
04990.png and 45295.png

overall, there are only ~45047 unique images in the validation set - about 5k of them occur twice, and a few even three times. Is that intended to give some examples more weight for validation score, or rather a bug? + wondering if it also applies to 64x64 version - haven't tested that yet

Thanks

Will the author kindly release pretrained weights for the Imagenet64?

Dear PatrykChrabaszcz:
I am using your dataset <ImageNet6464> , I want to re-implement WRN-36-k(1,2) on ImageNet6464 in the Table 1 at page5 . I read your paper and find that at page 3, you describe how to make the WRN-arcitecture. As a result, I draw this two figurs, please help me is it correct??

This is a different that in page 4, the writers said that he changed the order of BN, ReLU, Conv. He changed from conv-BN-ReLU to BN-ReLU-conv. Did you do the same procedure in your paper??

Thanks for your great work.
I've found that aspect ratio is changed in this work.
Is there a reason of why this is different to general ImageNet preprocessing? (center crop to keep the aspect ratio)

AttributeError: module 'scipy.misc' has no attribute 'imread'

imread is depreciated after Scipy version 1.2.0