bob.learn.tensorflow issueshttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues2017-06-29T21:52:23Zhttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/16Tensorboard graph does not look nice2017-06-29T21:52:23ZTiago de Freitas PereiraTensorboard graph does not look niceTensorboard is building a box for each element of the graph (variables and macro elements (loss, operations, etc...)) making the graph extremely confusing to analyse.Tensorboard is building a box for each element of the graph (variables and macro elements (loss, operations, etc...)) making the graph extremely confusing to analyse.https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/15Strange behaviour when a script is finished2018-05-19T15:16:19ZTiago de Freitas PereiraStrange behaviour when a script is finishedFor every script that is finished I got this error
```
Exception AttributeError: AttributeError("'NoneType' object has no attribute 'raise_exception_on_not_ok_status'",) in <bound method Session.__del__ of <tensorflow.python.client.sess...For every script that is finished I got this error
```
Exception AttributeError: AttributeError("'NoneType' object has no attribute 'raise_exception_on_not_ok_status'",) in <bound method Session.__del__ of <tensorflow.python.client.session.Session object at 0x7f22fdb730d0>> ignored
```
Some links to track the problem
- https://github.com/tensorflow/tensorflow/issues/3473
- https://github.com/tensorflow/tensorflow/issues/3388
The proposed way to reproduce the problem (issue #3388) only reproduces the issue with tensorflow < 0.9
```
import tensorflow as tf
a = tf.constant(123)
b = tf.constant(456)
c = a * b
session = tf.Session()
# A slightly different error is produced if this is removed.
session.run(tf.initialize_all_variables())
result = session.run(c)
print(result)
session.close() # The error is produced regardless of this.
#quit() # This produces the error.
import sys
sys.exit() # This also produces the error.
```
The issue is still open in the tensorflow bug tracker.
Let's track it..https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/14Memory issues2017-09-07T15:13:27ZTiago de Freitas PereiraMemory issuesIt seems that the memory management of tensorflow sucks.
A simple VGG16 with a batch of 128 makes the application crash by OOM (with 11GB of memory available)
https://github.com/tensorflow/tensorflow/issues/492It seems that the memory management of tensorflow sucks.
A simple VGG16 with a batch of 128 makes the application crash by OOM (with 11GB of memory available)
https://github.com/tensorflow/tensorflow/issues/492https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/13Move the creation of placeholdres from the datashufflers to trainers2016-11-03T11:02:53ZTiago de Freitas PereiraMove the creation of placeholdres from the datashufflers to trainersThis will make it easier to recovery of this informationThis will make it easier to recovery of this informationhttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/12My pooling layer does not have a stride2016-10-27T12:26:20ZTiago de Freitas PereiraMy pooling layer does not have a stridehttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/11CUDA control devices2016-11-28T13:11:34ZTiago de Freitas PereiraCUDA control devicesThink about how to organize the visible GPU devices using:
`os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, ..."`Think about how to organize the visible GPU devices using:
`os.environ["CUDA_VISIBLE_DEVICES"] = "0, 1, ..."`https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/10Implement a mechanism to turn on and off the GPU.....2016-11-14T13:53:05ZTiago de Freitas PereiraImplement a mechanism to turn on and off the GPU......... when you load a NN.... when you load a NNhttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/9Implement batch normalization2016-12-06T17:00:00ZTiago de Freitas PereiraImplement batch normalizationSome references
- https://youtu.be/gYpoJMlgyXA?list=PLkt2uSq6rBVctENoVBg1TpCC7OQi31AlC&t=3078
- https://github.com/davidsandberg/facenet/Some references
- https://youtu.be/gYpoJMlgyXA?list=PLkt2uSq6rBVctENoVBg1TpCC7OQi31AlC&t=3078
- https://github.com/davidsandberg/facenet/https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/8Datashufflers are not generic enough2017-10-21T15:11:56ZTiago de Freitas PereiraDatashufflers are not generic enoughSome reescaling is done inside of the datashuffler which makes it not generic enough for other kind of signals.
Maybe is a good idea to move this reescaling to another component.
Something similar what I did with the `ImageDataAugment...Some reescaling is done inside of the datashuffler which makes it not generic enough for other kind of signals.
Maybe is a good idea to move this reescaling to another component.
Something similar what I did with the `ImageDataAugmentation`
@pkorshunov keep track of thathttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/7The learning rate is not updated2016-10-27T14:12:48ZTiago de Freitas PereiraThe learning rate is not updatedFor some reason the learning rate update via weight decay is not workingFor some reason the learning rate update via weight decay is not workinghttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/6problem with GPU selection2016-12-06T17:00:42ZTiago de Freitas Pereiraproblem with GPU selectionI'm having a similar issue
https://github.com/tensorflow/tensorflow/issues/4674I'm having a similar issue
https://github.com/tensorflow/tensorflow/issues/4674https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/5Triplet selection2016-10-26T10:23:14ZTiago de Freitas PereiraTriplet selectionImplements the triplet selection of the google netImplements the triplet selection of the google nethttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/4http://tflearn.org/2017-10-12T15:44:39ZAmir MOHAMMADIhttp://tflearn.org/Have you seen this?Have you seen this?https://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/3Implement a class to do data augmentation2016-10-17T05:55:30ZTiago de Freitas PereiraImplement a class to do data augmentationImplement a class to do data augmentation
- [x] Mirroring
- [x] Salt and peper
- [x] Small translations
- [x] Small rotationsImplement a class to do data augmentation
- [x] Mirroring
- [x] Salt and peper
- [x] Small translations
- [x] Small rotationshttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/2Analizers2017-10-21T15:12:36ZTiago de Freitas PereiraAnalizersThe analizers are tools that are supposed to do 2 things
1 - Analize the network state
2 - Send a stop flag when the stop criteria is reached
I'm thinking on two types of analizers:
- [x] Experiment like Analizer. This one will ru...The analizers are tools that are supposed to do 2 things
1 - Analize the network state
2 - Send a stop flag when the stop criteria is reached
I'm thinking on two types of analizers:
- [x] Experiment like Analizer. This one will run a full experiment at each `N` iteration in the validation set and will do some plots, such as
- EER
- FAR10
- FAR1000
- Loss
- LR
- [ ] Simple analizer that just plot, basically, the loss of the train and validationhttps://gitlab.idiap.ch/bob/bob.learn.tensorflow/-/issues/1Improvements load function2016-10-27T07:32:27ZTiago de Freitas PereiraImprovements load functionToday the load function has to know before hand with network to load.
For example, if you want to load a Lenet, you have to do
```
lenet = Lenet(...)
lenet.load(hdf5_file)
```
If you don't know before hand the network it is impos...Today the load function has to know before hand with network to load.
For example, if you want to load a Lenet, you have to do
```
lenet = Lenet(...)
lenet.load(hdf5_file)
```
If you don't know before hand the network it is impossible to load it.
A possible solution would be:
```
net = SequenceNet()
net.load(hdf5_file)
print net __Lenet__
```