Open sed0724963 opened 5 years ago
By default model_main does't have any logging,
After all the imports you can add: tf.logging.set_verbosity(tf.logging.INFO)
This will log everything that is happening.
You should also know that it will be logging each 100 steps instead of each step like train.py. That is configurable but you will need to dig into the code for that.
The other reason it may seem slow, it by default you will be doing evaluations while training (well kind of). After x number of steps, training will pause to load up an evaluation of the current checkpoint. Once that is complete, training will resume again. This is also a configurable option if you only want to train.
@Cavan09 If I only want to train and validation,(loss_1,loss_2), the mAP or anything else I don't need it ,can I delete?
@sed0724963 Delete tf.logging.set_verbosity(tf.logging.INFO)?
All that will do is remove the logging to console, all the different parameters will still be captured.
@Cavan09 I'm mean the tensorboard have many scalars about: loss_1,loss_2 ,learning_rate,DetectionBoxes_Precision/mAP...,so I only want a function that only plot the train and validation loss(loss_1,loss_2). the other function about learning_rate,DetectionBoxes_Precision/mAP ,can I close?because I assume that run a part of evalution function(only train and validation) can to be faster than run all the evalution function,I don't need the mAP or other .
@sed0724963 I'm not sure you will get a huge benefit from removing the others, but if that is something you would like to try, you will have to dig into the COCO tools and evaluation metrics.
By default model_main does't have any logging,
After all the imports you can add: tf.logging.set_verbosity(tf.logging.INFO)
This will log everything that is happening.
You should also know that it will be logging each 100 steps instead of each step like train.py. That is configurable but you will need to dig into the code for that.
The other reason it may seem slow, it by default you will be doing evaluations while training (well kind of). After x number of steps, training will pause to load up an evaluation of the current checkpoint. Once that is complete, training will resume again. This is also a configurable option if you only want to train.
Do you know how to change the config files or command lines in order to only train?
I did training using model_main.py, but loss_1 appears like loss_2 in tensorboard. They supposed to be different since one of the is for training and the other for validation.
I'm using the object detecton api and use the model_main.py to training a detection model, I use the faster-rcnn resnet101.coco and batch size =1,but it's so slow?
what can I do to let it to be faster??
System information