Predict on batch #195

Activity

1. de-code commented on Mar 17, 2022

   

   de-code

   on Mar 17, 2022

   Last edited by de-code

   Owner

   More actions

   It wouldn't be difficult to add a predict on batch function.
   Internally (except for TF Lite I believe), it is using a batch already.
   I think were it becomes a bit tricky is to follow it through with the post processing.

   e.g. currently with a single image it looks like this:

   result = bodypix_model.predict_single(image_array)
   
   # simple mask
   mask = result.get_mask(threshold=0.75)
   
   # colored mask (separate colour for each body part)
   colored_mask = result.get_colored_part_mask(mask)

   For a batch it would then be something like:

   batch_result = bodypix_model.predict_batch(image_array_batch)
   
   # simple mask
   mask_batch = batch_result.get_mask_batch(threshold=0.75)
   
   # colored mask (separate colour for each body part)
   colored_mask_batch = batch_result.get_colored_part_mask_batch(mask_batch)

   I also wonder whether there would be any noticable speed improvement.

   What have you observed when using the JS version?
   And what is your use-case?
2. LKchemposer commented on Mar 21, 2022

   

   LKchemposer

   on Mar 21, 2022

   Author

   More actions

   Thank you for the reply.

   My current use case is measuring BodyPix's performance over a dataset of 10000+ images, but given the high image resolution (1080p), it's taking ~1.3-1.5 sec/image on Colab's default CPU. Any improvement on inference time is useful (without having to downsample).

   I have not tried with the JS version yet, but it seems that there is "a large performance difference" according to tensorflow/tfjs#2197
3. de-code commented on Mar 31, 2022

   

   de-code

   on Mar 31, 2022

   Owner

   More actions

   Would you be happy to see whether you want to submit a PR to add batch support?
4. de-code

   added

   enhancementNew feature or requestNew feature or request

   on Aug 8, 2022