Training not working - Githubissues

robrao commented 6 years ago

mobilenet ckpt model and object detection model don't seem to align (may not be core issue)

robrao commented 6 years ago

using ssd_model
ensure steps above ckpt model steps
command: python object_detection/model_main.py --pipeline_config_path=models/model/ssdlite_mobilenet_v2_coco.config --train_dir=train

May work...

robrao commented 5 years ago

Need to move model.ckpt out of models/model/ssd_.. into /models/model (automate)
Loss decreasing but evaluation returning no values (investigate)
Training always killed even with unlimited evalutions (investigate)
Add tensorboard image evaluation result to training aspect of model_lib, and input image visualization
Evaluation has tensorbard scalars for classification loss, localization loss, total loss, and DetectionBoxes Recall and Precision. These should be available during training as well.
Eval metrics come from coco_evaluation.get_estimator_eval_metric_ops, after retrieved put into tf.estimator.EstimatorSpec. Do something similar for training?
Why classification loss, and localization loss decreasing over time?
Console AP and AR don't change, always 0s

robrao / text-generation