Single path tan loading by ajratner · Pull Request #6 · HazyResearch/tanda

ajratner · 2017-12-06T02:39:40Z

Still needs to be tested

henryre

Definitely needs to be tested as well. For next rev, can use the MNIST small example script

henryre · 2017-12-06T19:19:47Z

keras/utils.py

+
+    # Load TFs
+    # Assume they are present in config['train_module'] as list called tfs
+    tfs = import_module(config['train_module']).tfs


This seems like way too strong of an assumption. Any other ideas on how to load in TFs?

Fixed in latest commit- we just pickle them using the cloud lib (a good logging / reproducibility step regardless), then just load them in; much simpler. And all handled in the train_scripts.py file

henryre · 2017-12-06T19:20:38Z

keras/utils.py

-    with open(config_path, 'r') as f:
+def load_pretrained_tan(path):
+    # Load config dictionary from run log
+    with open(os.path.join(path, 'run_log.json'), 'r') as f:


Is this where this file is always located? I think it's usually in logs, just like this for the pretrained dir

henryre · 2017-12-06T19:21:37Z

keras/tanda_keras.py


    def __init__(self,
-                 tan,
+                 tan_path,


Definitely need to allow user to pass in TAN object. Can use isinstance(tan, str)

henryre · 2017-12-06T19:28:44Z

experiments/cifar10/train.py


+    # Make sure dims and current module name is included in the run log
+    # Note: this is currently kind of hackey, should clean up...
+    FLAGS.__flags['train_module'] = re.sub(r'\/', '.',


Breaks if any experiment subdir contains "tanda" right?

Also shouldnt this be somewhere in train_scripts.py?

ajratner · 2017-12-07T19:08:07Z

@henryre ready for re-review! Can be tested when training new TANs

henryre · 2017-12-07T19:08:40Z

Ok will look soon!

henryre

Let's run a basic test here to make sure these changes work

henryre · 2017-12-07T21:09:27Z

experiments/cifar10/train.py

 from __future__ import unicode_literals

-from dataset import load_cifar10_data
+import sys


Remove these

henryre · 2017-12-07T21:09:56Z

experiments/cifar10/train.py

+import sys
+import re
+
+from .dataset import load_cifar10_data


Revert to from dataset

henryre · 2017-12-07T21:13:20Z

experiments/train_scripts.py

 import tensorflow as tf
 import tensorflow.contrib.slim as slim
+import sys
+import cloud


Needs to be added to package requirement

Also why this instead of e.g. dill?

henryre · 2017-12-07T21:14:53Z

keras/utils.py

 from __future__ import unicode_literals

 import json
+import pickle


Nit: from six import cPickle

henryre · 2017-12-07T21:15:38Z

pretrained/cifar10/logs/run_log.json

@@ -0,0 +1,61 @@
+{


henryre · 2017-12-07T21:16:46Z

experiments/train_scripts.py

 import re
 import tensorflow as tf
 import tensorflow.contrib.slim as slim
+import sys


ajratner · 2017-12-13T06:50:08Z

@henryre Addressed all inline comments

henryre · 2017-12-13T21:49:33Z

experiments/cifar10/train.py

 from __future__ import print_function
 from __future__ import unicode_literals

+import re


Still need to remove

henryre · 2017-12-13T21:49:40Z

keras/utils.py

 from __future__ import unicode_literals

 import json
+from six import pickle


It's cPickle not pickle

henryre · 2017-12-13T21:50:18Z

python-package-requirement.txt

 scipy>=0.18
 six
 tensorflow>=1.2
+cloud


Nit: Alpha order

ajratner · 2017-12-13T23:02:51Z

@henryre changes made

ajratner added 2 commits December 5, 2017 18:24

Loading pretrained TAN from single path

9e86464

Sneaking dims and module name into run_log

2282078

ajratner requested a review from henryre December 6, 2017 02:39

henryre requested changes Dec 6, 2017

View reviewed changes

ajratner added 3 commits December 6, 2017 12:39

Addressing PR comments

2f11687

Now we pickle the TFs using the cloud lib (a good logging step anyway)

c76be20

Merge branch 'master' into single-path-tan-loading

2e7ab62

henryre requested changes Dec 7, 2017

View reviewed changes

henryre reviewed Dec 7, 2017

View reviewed changes

experiments/train_scripts.py Outdated

import re

import tensorflow as tf

import tensorflow.contrib.slim as slim

import sys

Copy link

Member

henryre Dec 7, 2017

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Remove

Addressing review comments

5feb735

henryre reviewed Dec 13, 2017

View reviewed changes

Nits on nits on nits

4dbc41c

Conversation

ajratner commented Dec 6, 2017

Uh oh!

henryre left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ajratner commented Dec 7, 2017

Uh oh!

henryre commented Dec 7, 2017

Uh oh!

henryre left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ajratner commented Dec 13, 2017

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ajratner commented Dec 13, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants