Commits

Branch selector

 master

User selector

All users

All time

Commit History

Commits on May 25, 2017

• hyperparameter tuning more robust before files are written etc.

  

  Julian Kates-Harbeck

  committed
  75ab3d4

  Copy full SHA for 75ab3d4

• hyperparameter tuning

  

  Julian Kates-Harbeck

  committed
  34cedd2

  Copy full SHA for 34cedd2

• several changes to enable hyperparameter tuning. Had to redirect some paths in the case of hyperparameter tuning to do things like copying the normalization files

  

  Julian Kates-Harbeck

  committed
  5257294

  Copy full SHA for 5257294

• early stopping works

  

  Julian Kates-Harbeck

  committed
  6f0142c

  Copy full SHA for 6f0142c

• added early stopping via keras callback

  

  Julian Kates-Harbeck

  committed
  6ad6083

  Copy full SHA for 6ad6083

• added early stopping via keras callback

  

  Julian Kates-Harbeck

  committed
  a14a5c4

  Copy full SHA for a14a5c4

• added early stopping via keras callback

  

  Julian Kates-Harbeck

  committed
  022bf3a

  Copy full SHA for 022bf3a

• changing csv logs to local folder for hyperparameter tuning

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  90a9c89

  Copy full SHA for 90a9c89

Commits on May 24, 2017

• added positive example penalization for MaxHingeTarget to conf as tunable parameter. Added Classes for hyperaparameter search

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  c4cd4f7

  Copy full SHA for c4cd4f7

• removed old signals paths, masks, and plotting strings from conf files. Moved NUM_GPUS to conf file

  

  Julian Kates-Harbeck

  committed
  29c0050

  Copy full SHA for 29c0050

• removed old signals paths, masks, and plotting strings from conf files. Moved NUM_GPUS to conf file

  

  Julian Kates-Harbeck

  committed
  af20771

  Copy full SHA for af20771

Commits on May 23, 2017

• merge

  

  Julian Kates-Harbeck

  committed
  b16209c

  Copy full SHA for b16209c

• fixed nasty bug related to copy by view vs. copy by value of numpy arrays in the buffer based prefetching. Needed to copy the array, because the view was being changed behind the scenes.

  

  Julian Kates-Harbeck

  committed
  bf4433a

  Copy full SHA for bf4433a

• validation_batch_generator -> training_batch_generator. Use the same generator for validation, but with different shotlist

  ASvyatkovskiy

  committed
  bd61b2a

  Copy full SHA for bd61b2a

• version bump to 0.2. Added multiple dimensional signals, multiple machines, and signal versatility.

  

  Julian Kates-Harbeck

  committed
  3160bb9

  Copy full SHA for 3160bb9

• Merge pull request #13 from PPPLDeepLearning/d3d_data

  Show description for d95eb03
  jnkh

  authored
  d95eb03

  Copy full SHA for d95eb03

• Merge branch 'master' into d3d_data

  jnkh

  authored
  4d91009

  Copy full SHA for 4d91009

• removed intermediate dense layers

  

  Julian Kates-Harbeck

  committed
  bb40e6a

  Copy full SHA for bb40e6a

• fixed tensorflow re-import issue

  

  Julian Kates-Harbeck

  committed
  b393aee

  Copy full SHA for b393aee

Commits on May 22, 2017

• Remove debugging setup

  ASvyatkovskiy

  committed
  d02c6c0

  Copy full SHA for d02c6c0

• Restore default signal_dirs

  ASvyatkovskiy

  committed
  f168236

  Copy full SHA for f168236

• Merge pull request #12 from PPPLDeepLearning/tensorboard_callbacks

  Show description for 1141062
  ASvyatkovskiy

  authored
  1141062

  Copy full SHA for 1141062

• Implement Tensorboard callback as a separate class (not using Keras at this time, due to issue with validation generator)

  ASvyatkovskiy

  committed
  8c7ce1b

  Copy full SHA for 8c7ce1b

• added data pre-fetching generator with separate process for prefetching. Also cleanup for this new process. This works in conjuction with buffer based data loading

  

  Julian Kates-Harbeck

  committed
  e81c826

  Copy full SHA for e81c826

Commits on May 21, 2017

• adding buffer based training to non mpi version

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  7684509

  Copy full SHA for 7684509

• added data prefetching process for training generator (via buffer) in the loader. This runs in parallel to the GPU training and outsources prefetching. Shortens training time and avoids problems wi…

  Show description for 3907ffb
  

  Julian Kates-Harbeck

  committed
  3907ffb

  Copy full SHA for 3907ffb

• debugging buffer based training

  

  Julian Kates-Harbeck

  committed
  254ec6e

  Copy full SHA for 254ec6e

• adding functions to reset states partially (within a batch)

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  23a48fa

  Copy full SHA for 23a48fa

• Merge branch 'd3d_data' of https://github.com/PPPLDeepLearning/plasma-python into d3d_data

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  dd10c51

  Copy full SHA for dd10c51

• adding buffer based batch generator (with partial state reset and without chunks. Also adding option to cut end of shot (less than T_min_warn) for training

  

  Julian Kates-Harbeck

  authored and

  Julian Kates-Harbeck

  committed
  fff3f76

  Copy full SHA for fff3f76

• added max target, only the maximum value is penalized

  

  Julian Kates-Harbeck

  committed
  5dae443

  Copy full SHA for 5dae443

Commits on May 19, 2017

• Added maximum-based target. It uses a loss function that only penalizes the maximum value out of a temporal sequence, not the mean value. This is targeted for classification tasks in which only the…

  Show description for f6a3273
  

  Julian Kates-Harbeck

  committed
  f6a3273

  Copy full SHA for f6a3273

Commits on May 18, 2017

• mpi inference bug fixed via sorting and printout removed. Sorting not necessary in non-mpi version because there is only one version of the shot list

  

  Julian Kates-Harbeck

  committed
  5e7751a

  Copy full SHA for 5e7751a

• Merge branch 'd3d_data' of https://github.com/PPPLDeepLearning/plasma-python into d3d_data

  

  Julian Kates-Harbeck

  committed
  0b52fb8

  Copy full SHA for 0b52fb8

• debug printouts

  

  Julian Kates-Harbeck

  committed
  85178c4

  Copy full SHA for 85178c4