Implementation about kernel activation

[lifuguan]

Hello, Sorry to disturb you. I'm trying to visualize the kernels (called object_feats in your code). It've been illustrated in your paper. Here is my code, which aims to save and add them on kernels.npy during the inference phrase.

"""kernel_iter_update.py line:296"""
                results.append(single_result)
        from debugger import save_test_info
        save_test_info(img_metas, scores_per_img, masks_per_img, object_feats)
        return results

def save_test_info(img_metas:list, 
    cls_score:torch.Tensor, 
    scaled_mask_preds:torch.Tensor, obj_feats:torch.Tensor):
    ...
    # kernels
    if obj_feats is not None:
        kernels_old = np.load("work_dirs/tmp/kernels.npy")
        kernels_new = obj_feats.to('cpu').detach().numpy()
        kernels = kernels_new+kernels_old
        np.save("work_dirs/tmp/kernels.npy", kernels)

"""after inference phrase"""
fig,a =  plt.subplots(10,10)
kernels_2dim = kernels.reshape((100,16,16))
for i in range(100):
    # a[int(i / 10)][i % 10].set_title(i)
    a[int(i / 10)][i % 10].set_xticks([])
    a[int(i / 10)][i % 10].set_yticks([])
    a[int(i / 10)][i % 10].imshow(kernels_2dim[i], cmap = plt.cm.hot_r)
plt.savefig('work_dirs/tmp/class_80_ins_2/kernel_2dim.png', bbox_inches='tight')
plt.show()

However, the result is completely different from your figures:

It will be appreciated if anyone can show me the way to visualize kernel correctly.

[ZwwWayne]

You should not save kernels, but the masks. The paper says "... the average of mask activations of the 100 instance kernels over the 5000 images in the val split. " Thus, the mask predictions after sigmoid are saved and averaged, rather than the kernels.