Do I have to make changes in the multioutput file? I ideally want to train any model, like Inception, on my training data which has multi labels. How do I do that?

This PR allows a user to specify a validation dataset that are used for early stopping (and reporting). The PR was created to address issue 85

I made changes in 3 places.

1. The trainer now takes a dictionary containing the validation data (in the same format as the output of the data feeder's get_dict_fn).
2. The fit method now takes arguments for val_X and val_y. It converts these into the correct format for the trainer.
3. The example file digits.py now uses early stopping, by supplying val_X and val_y.

I can add early stopping to other examples if this approach looks good, though their behavior should not otherwise be affected by the current PR.

Hi,

I am using skflow.ops.dnn to classify two - classes dataset (True and False). The percentage of True example is very small, so I have an imbalanced dataset.

It seems to me that one way to resolve the issue is to use weighted classes. However, when I look to the implementation of skflow.ops.dnn, I do not know how could I do weighted classes with DNN.

Is it possible to do that with skflow, or is there another technique to deal with imbalanced dataset problem in skflow?

Thanks

I added an option to control the "verbosity". For this, I added the parameter "verbose" in the init method of the init.py file and to the train function in the trainers.py file. In addition, I passed this argument to the "self._trainer.train()" call in the init file and added a condition to make the prints in the trainer.py file.

This changes the default batch size for prediction to be the same as for training, enabling efficient grid search. Previously GridSearchCV would try to make predictions in a single batch, which could take a lot of memory.

This also adds a simple example of using skflow with GridSearchCV.

It wasn't clear how to access weights using classifier.get_tensor_value('foo') syntax. This adds some examples for the CNN model. They were figured out by logging the training as though for using TensorBoard, and then running strings on the logfile to look for the right namespace.

Is there a better way to access these weights? Or to learn their names? The logging must walk through the graph and record these names. Maybe if there were a way to quickly list all the names, that'd be enough for advanced users to figure it out.

Hi,

Sorry I asked too much.

I think plotting is always a nice feature. Is it possible right now for skflow (or can we do that through tensorflow directly)?

This moves the monitor and logdir arguments for the fit method to the base estimator __init__ method so that fit better follows the sklearn API guidelines: "fit parameters should be restricted to directly data dependent variables". This should make it easier to use, e.g., GridSearchCV with skflow models.

Hi,

Thanks for making this tool. It will definitely make things easier for NN newcomers.

I just tried running your language model example and got the following exception:

Traceback (most recent call last):
  File "test.py", line 84, in <module>
    estimator.fit(X, y)
  File "/Users/aleksandar/tensorflow/lib/python3.5/site-packages/skflow/estimators/base.py", line 243, in fit
    feed_params_fn=self._data_feeder.get_feed_params)
  File "/Users/aleksandar/tensorflow/lib/python3.5/site-packages/skflow/trainer.py", line 114, in train
    feed_dict = feed_dict_fn()
  File "/Users/aleksandar/tensorflow/lib/python3.5/site-packages/skflow/io/data_feeder.py", line 307, in _feed_dict_fn
    inp[i, :] = six.next(self.X)
StopIteration

I made sure that my python distribution has the correct version of six. I tried running it both in a virtual environment and in a normal Python 3 distro. Any ideas what might be causing this?

From what I understand, the existing ValidationMonitor performs validation every [print_steps] steps, and checks for stop condition every [early_stopping_rounds] steps. I'd like to add another ValidationMonitor that performs validation once and checks for stoping condition once every epoch. Is this the recommended practice in machine learning regarding validation and early stopping? I mean I'd like to add a fit process something like this:

def fit(self, x_train, y_train, x_validate, y_validate):
    while (current_validation_loss < previous_validation_loss):
        estimator.train_one_more_epoch(x_train, y_train)
        previous_validation_loss = current_validation_loss
        current_validation_loss = some_error(y_validate, estimator.predict(x_validate))

Add an example of language model (RNN). For example character level on sheikspear book (similar to https://github.com/sherjilozair/char-rnn-tensorflow).

Travis are now recommending removing the sudo tag.

"If you currently specify sudo: false in your .travis.yml, we recommend removing that configuration"

New versions of TF have already been released since the last commit to this repo. As far as I've understood, after having read the README file of this project, you intended to close this repo. So, why hasn't it been done yet?