search

mortenjust / trainer-mac

Trains a model, then generates a complete Xcode project that uses it - no code necessary
125 stars 9 forks source link

Retraining Data #1

Open BFMarks opened 7 years ago

BFMarks commented 7 years ago

I'm trying to optimize my model but when I change the training data - the Trainer doesn't seem to retrain based on my new input data. What other files should I change besides the Images/Originals/ folder? Also, I want to adjust the variance and bias and I can't seem to find the best way to do that.

Anyway - love the app! Very cool and I'd like to keep working with it!

Cheers,

mortenjust commented 7 years ago

Good pointers for the tutorial here, thanks.

Adjusting parameters: Should definitely be a part of the UI, so that's coming up next. Retraining: Adding images to the Originals folder should indeed add those to the model and retrain.

There's a resizing and bottleneck cache, so retraining will be a lot faster. If you're only adding, say 10 new images, it will be many times faster than the initial training. When it's done, all you have to do is to delete the .pb file from the xcode project and drag the newly trained model pb file (check the timestamp to verify it's been retrained) into the xcode project.

If you think there may be a bug, try and paste in the content of the console output panel here.

mortenjust commented 7 years ago

Oh, and can you paste in the shell command where you'd normally change variance and bias while retraining?

BFMarks commented 7 years ago

Hey Morten,

The training logs say:

grep: /tf_files/log-resized-images.txt: Transport endpoint is not connected
grep: /tf_files/log-resized-images.txt: Transport endpoint is not connected
Resizing 2632855.jpg for label test2 into /tf_files/images/resized/test2
mkdir: cannot create directory '/tf_files': Transport endpoint is not connected
Error response from daemon: Bad response from Docker engine

It maybe the file space on my Macbook - I'll try again on a different system. I'll also get back to you on how I change the variance and bias. The docker engine is still running, will restart.