Keras interface and Python 3 compatability #3
Conversation
Compatability imports Print parens Fix iters Fix imports
README.md
Outdated
| ## NEW: an easy-to-use Keras interface | ||
|
|
||
| Just in time for NIPS 2017, we're releasing an easy-to-use | ||
| [Keras](https://keras.io/) interface for TANDA. Just train your TAN, load the |
There was a problem hiding this comment.
Re: "Just train your TAN...": I'd say let's lead with a one line (bolded and red?) about Keras interface + pre-trained TAN. Then mention can be used with any TAN you train / is implied
There was a problem hiding this comment.
I'd suggest something like: "Just in time for NIPS 2017, we're releasing an easy-to-use, pre-trained substitute for Keras' ImageDataGenerator data augmentation class. Just swap in TANDAImageDataGenerator and you'll be using our trained data augmentation generators!"
| @@ -1,6 +1,6 @@ | |||
| #!/bin/bash | |||
| source set_env.sh | |||
| python experiments/cifar10/train.py \ | |||
| python -m experiments.cifar10.train \ | |||
experiments/cifar10/dataset.py
Outdated
|
|
||
| def load_cifar10_batch(fpath, one_hot=True, as_float=True): | ||
| with open(fpath, 'rb') as f: | ||
| data = cPickle.load(f) | ||
| data = cPickle.load(f, encoding='latin1') |
There was a problem hiding this comment.
Many stacks overflowed here. Just let it be.
| self.tan = tan | ||
| self.session = K.get_session() | ||
|
|
||
| def random_transform(self, x, seed=None): |
There was a problem hiding this comment.
Minor: Update doc string just to clarify that this is not in fact a random transform anymore, just maintaining same interface?
| """ | ||
|
|
||
| def __init__(self, | ||
| tan, |
There was a problem hiding this comment.
Cleaner to just provide the checkpoint path I think?
There was a problem hiding this comment.
Offline: can either be checkpoint path or TAN object
| from experiments.train_scripts import GENERATORS | ||
| from experiments.utils import parse_config_str | ||
| from keras import backend as K | ||
| from tanda.generator import ( |
There was a problem hiding this comment.
A bit cleaner than redefining IMO
|
Add a comment linking to said posts?
…On Tue, Dec 5, 2017 at 2:49 PM Henry Ehrenberg ***@***.***> wrote:
***@***.**** commented on this pull request.
------------------------------
In experiments/cifar10/dataset.py
<#3 (comment)>:
>
def load_cifar10_batch(fpath, one_hot=True, as_float=True):
with open(fpath, 'rb') as f:
- data = cPickle.load(f)
+ data = cPickle.load(f, encoding='latin1')
Many stacks overflowed here. Just let it be.
—
You are receiving this because your review was requested.
Reply to this email directly, view it on GitHub
<#3 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ABgw_WGumkIOb_HmQvXMN_gVZ2IN9CPkks5s9chrgaJpZM4Q3CYd>
.
|
Compatability imports Print parens Fix iters Fix imports
Keras interface and Python 3 compatability
Also adds a pretrained CIFAR-10 model