Cannot add images to an existing classifier - ibm-cloud

Question

I have successfully created a classifier using 20 classes, with 200 images in each class. When I try to add additional classes to this classifier, the API call fails:
Traceback (most recent call last):
File "create_classifier.py", line 34, in <module>
print visual_recognition.update_classifier('flowers_1337689264', **myhash)
File "/usr/local/lib/python2.7/dist-packages/watson_developer_cloud/visual_recognition_v3.py", line 96, in update_classifier
params=params, accept_json=True)
File "/usr/local/lib/python2.7/dist-packages/watson_developer_cloud/watson_developer_cloud_service.py", line 268, in request
raise WatsonException(error_message)
watson_developer_cloud.watson_developer_cloud_service.WatsonException: Unknown error
Here is my code:
import json
from watson_developer_cloud import VisualRecognitionV3
from glob import glob
visual_recognition = VisualRecognitionV3('2016-05-20', api_key='xxxx')
myhash = {}
for f in sorted(glob('*.zip'))[20:30]:
clas = f.replace('.zip','');
clas = clas.replace(' ', '-')
clas += '_positive_examples'
myhash[clas] = open(f, 'rb')
print myhash
#print visual_recognition.create_classifier('test', **myhash)
print visual_recognition.update_classifier('test_1337689264', **myhash)

Answer 1

Not sure if this is your problem, but there is an issue for retraining - If your training .zip file included files that are named like images (for example, *.jpg or *.png), but are not actually valid image files, the system will think it has incomplete data when trying to marshal all the vectors for retraining. This can occur frequently when using .zip files from MacOS, which might include a "hidden" __MACOSX folder with files named like this.
A solution for this issue of non-image files causing retraining to fail was deployed on Aug 10, 2016. Classifiers created after that date should not be hindered in retraining from non-image files presence in the training examples. It's not retroactive, though, so the workaround for classifiers trained before that date is to start the training process over with the same input zip files to create a new one, and then delete the old classifier.