Distinct samples of training data

[leowe]

Is it possible to train the ffn with several distinct pieces of data with each having its own bounding box?

[subeeshvasu]

I am also trying to do the same. After inspecting the code, I figured out the following solution, to train with more than one data volumes. Let's assume that we have two training volumes at hand.

• First, generate the coordinates for both volumes, but with different tags. i.e., while using 'build_coordinates.py' we could use "validation1" as the tag for volume 1 and "validation2" as the tag for volume 2. eg. python build_coordinates.py --partition_volumes validation1:address_to_volume1_af.h5:af_file,validation2:address_to_volume2_af.h5:af_file --coordinate_output ..
• Next, change the arguments corresponding to input volume and corresponding labels: data_volumes from validation1:address_to_volume_1.h5:raw to validation1:address_to_volume_1.h5:raw,validation2:address_to_volume_2.h5:raw

Repeat the above step for label_volumes too

• Now run the training code as usual.

[mk123qwe]

How to save the model according to the set conditions, not periodically. I noticed that MonitoredSession is a very inefficient method.

[mk123qwe]

I am also trying to do the same. After inspecting the code, I figured out the following solution, to train with more than one data volumes. Let's assume that we have two training volumes at hand.

• First, generate the coordinates for both volumes, but with different tags. i.e., while using 'build_coordinates.py' we could use "validation1" as the tag for volume 1 and "validation2" as the tag for volume 2. eg. python build_coordinates.py --partition_volumes validation1:address_to_volume1_af.h5:af_file,validation2:address_to_volume2_af.h5:af_file --coordinate_output ..
• Next, change the arguments corresponding to input volume and corresponding labels: data_volumes from validation1:address_to_volume_1.h5:raw to validation1:address_to_volume_1.h5:raw,validation2:address_to_volume_2.h5:raw

Repeat the above step for label_volumes too

• Now run the training code as usual.

Hello, I'm confused that this program has no validation set.

[subeeshvasu]

I am also trying to do the same. After inspecting the code, I figured out the following solution, to train with more than one data volumes. Let's assume that we have two training volumes at hand.

• First, generate the coordinates for both volumes, but with different tags. i.e., while using 'build_coordinates.py' we could use "validation1" as the tag for volume 1 and "validation2" as the tag for volume 2. eg. python build_coordinates.py --partition_volumes validation1:address_to_volume1_af.h5:af_file,validation2:address_to_volume2_af.h5:af_file --coordinate_output ..
• Next, change the arguments corresponding to input volume and corresponding labels: data_volumes from validation1:address_to_volume_1.h5:raw to validation1:address_to_volume_1.h5:raw,validation2:address_to_volume_2.h5:raw

Repeat the above step for label_volumes too

• Now run the training code as usual.

Hello, I'm confused that this program has no verification set. I didn't understand. You meant: "No setup/data to check monitor the improvement in segmentation?". If yes, I agree. I would also like to figure out a way to save the network weights based on some sort of quantitative evaluation. I haven't explored that direction yet. Let me know if you figure out a way to do it. Thanks.

[mk123qwe]

I am also trying to do the same. After inspecting the code, I figured out the following solution, to train with more than one data volumes. Let's assume that we have two training volumes at hand.

• First, generate the coordinates for both volumes, but with different tags. i.e., while using 'build_coordinates.py' we could use "validation1" as the tag for volume 1 and "validation2" as the tag for volume 2. eg. python build_coordinates.py --partition_volumes validation1:address_to_volume1_af.h5:af_file,validation2:address_to_volume2_af.h5:af_file --coordinate_output ..
• Next, change the arguments corresponding to input volume and corresponding labels: data_volumes from validation1:address_to_volume_1.h5:raw to validation1:address_to_volume_1.h5:raw,validation2:address_to_volume_2.h5:raw

Repeat the above step for label_volumes too

• Now run the training code as usual.

Hello, I'm confused that this program has no verification set. I didn't understand. You meant: "No setup/data to check monitor the improvement in segmentation?". If yes, I agree. I would also like to figure out a way to save the network weights based on some sort of quantitative evaluation. I haven't explored that direction yet. Let me know if you figure out a way to do it. Thanks.

I don't understand enough about FFN training process, which is different from what I know about segmentation tasks. "MonitoredTrainingSession" is different from Session.You can try "Supervisor".