Regarding maximum total number and size of images which we can upload when training Watson VR classifier

Question

Could you tell me maximum total number and size of images which we can upload when training Watson VR classifier? I only found maximum number and size of images per API Call (10000 images, 100MB per zip).

Answer 1

Guidelines for good training and details about size limitations are available in the Watson Visual Recognition Docs:https://console.bluemix.net/docs/services/visual-recognition/customizing.html#size-limitations

Guidelines for high volume classifying is available here: https://console.bluemix.net/docs/services/visual-recognition/customizing.html#guidelines-for-high-volume-classifying

Hope this helps!

Answer 2

Could you tell me maximum total number and size of images which we can upload when training Watson VR classifier? I only found maximum number and size of images per API Call (10000 images, 100MB per zip).

A single API call is limited as you noted, however you can train a classifier with more images using the "retraining" capability. Basically you would make batches of your training data which fit in these per-API-call limits, and then train your classifier with the first batch. Then you would use the retraining capability (also called "Update a classifier" to add successive batches, by supplying the classifier_id to add that batch to. The spec for updating a classifier is here: https://console.bluemix.net/apidocs/visual-recognition#update-a-classifier

If you are training with a large number of images, you may choose to resize your training images to exactly 224x224 pixels before sending them to the Watson API. This will not affect the quality of the training or classification, but it will allow you to maximize the number of images per batch. One of the first steps the custom training and classification system takes is to resize input images to 224x224 so if there are details not visible at that resolution, they will very likely not be discernible by the learner. If you have really large images, or aspect ratios >2 or <0.5, they should be cropped into tiles before being submitted to get best results.

There is no limit on the number of images you can add via retraining to a classifier's training set. However, each additional image does count as a "training event" as defined in our pricing schedule: https://www.ibm.com/cloud/watson-visual-recognition/pricing

There is also not a strict limit on the size of an individual image inside a training set, however it is wasteful and gives no advantage to send large images. The service is not running on magic, though, so extremely large images may get skipped during the training process or cause training to fail. The service is designed to work best on photographic images. You should not encounter any problem with JPGs or PNGs produced by a wide range of standard cameras. (The /classify endpoint does limit each image to 10MB though)