Does your code contain the mask mAP computation? #139
Conversation
added test.py
- bad assignment in mask.py - smoothen the cropped features from pyramid in crop.py - other visualization and minor parameter tuning fixed network setting - add relu and maxpooling pad to non-batch normalization scope
- remove batch normalization in the layers before loss Fixed some mask issues - temporary fixed inaccurate gt_mask from cv crop and resize - change gt_mask from int to float - smoothen mask by cv INTER_CUBIC Misc. - match all rois and their target (manually checked through indexs)
corrected mask losses in pyramid_network.py
changed some variable names
added color codes
fixed the issue that is_training=False causes no bounding boxes during test fixed batch normalization gradient update
change anchor back to 3x3 (5x3 creates too many anchors. anchor_encoder slows down the training significantly)
added recall and precision calculation set test dataset to val2014 remove random picking data during test
fixed some evaluation problems
revert partial sort (too buggy)
changed the layers for pyramid features C2-4 to match the original paper (endpoints of each resnet block) test.py generates results.json file pycocoEval.py evaluates AP and AR from results.json ***my max current AP 0.5 is 0.262
|
The one in fix_testing branch is included with evaluation API from coco. |
… yet) current mAP@50 IOU by FPN+Resnet50 (training from scratch) is 43.7% changed anchor scale to match with the original paper changed head parts of RPN RCNN and Mask rename and remove multiple variables
changed some hyper params to reduce overfitting (mAP@1M train:0.47 test:0.36)
removed images with extreme aspect ratio which cause OOM from tfrecords (records must be re-created)
|
With current version, I added the evaluation part using API from MSCOCO for both bounding box and mask. You can evaluate the network by After training up to 1M+ iteration, I got mAP(mask IoU=0.5) on training data = 0.47 and on testing data = 0.36 (comparing to original paper Resnet 50 mAP of 0.512). Please note that your previously trained networks and created tfrecords may not work with this version because I changed a lot of network settings. I found 2 problems but I will not have time to fix it. |
No description provided.