|
don't ask me about deploying mask r-cnn models to production systems with no quantifiable validation metrics or model versioning
|
#
¿
Jun 1, 2019 15:26
|
|
|
|
| # ¿ May 11, 2024 19:37 |
|
|
lancemantis posted:bonus if it’s just the Facebook repo matterport, actually (detectron apparently doesn't run in a vm with cpu and 2gb of ram. still want to try maskrcnn-benchmark though because it's in pytorch.)
|
|
#
¿
Jun 1, 2019 16:08
|
|
|
code:
|
|
#
¿
Jun 8, 2020 15:12
|
|
|
Arcteryx Anarchist posted:m8 wtf are you doing trying to train an autoencoder-ish thing based on resnet-50 with ms coco. the total_loss is just the mse. i started training it on friday and basically it spent ~31k batches of images oscillating around learning nothing. time to start randomly changing various hyperparameters!
|
|
#
¿
Jun 8, 2020 16:57
|
|
|
initial guess with 1e-5, detectron2 does some lr warmup which is why it starts out lower than that for the beginning batches. besides double checking all the code, lowering the learning rate was experiment #1 on the "randomly change hyperparams" list.
|
|
#
¿
Jun 8, 2020 18:16
|
|
|
|
| # ¿ May 11, 2024 19:37 |
|
|
Max Facetime posted:it won’t have actually learned anything until it’s able to explain what it has learned this network has learned something but no animal ever has 
|
|
#
¿
Jun 9, 2020 03:55
|
|