vitruvianscience / opendeep Goto Github PK

On line 729 of opendeep/models/model.py the default file name is set to "config":

def save_args(self, args_file="config"):

A call to file_ops.get_extension_type() in the method logs a warning "Didn't recognize file extension..." which is spurious because save_args() later appends ".pkl" to the args_file in this case. Suggested fix is to make the default file name "config.pkl" in the first place. This will have exactly the same default behaviour without the warning.

there is not any opendeep.models.single_layer.autoencoder to import DenoisingAutoencoder from :/

I try to use my custom dataset with the given AutoEncoder example but when it starts it raises following error

aining)
    225                      str(type(self.model)), func_i+1, len(self.train_functions), self.n_epoch, str(continue_training))
    226 
--> 227             log.debug("Train dataset size is: %s", self.dataset.getDataShape(datasets.TRAIN))
    228             if self.dataset.hasSubset(datasets.VALID):
    229                 log.debug("Valid dataset size is: %s", self.dataset.getDataShape(datasets.VALID))

/home/retina18/Downloads/opendeep/opendeep/data/dataset.py in getDataShape(self, subset)
    111         if subset is TRAIN:
    112             log.error("No training shape implemented")
--> 113             raise NotImplementedError("No training shape implemented")
    114         elif subset is VALID:
    115             log.error("No valid shape implemented")

NotImplementedError: No training shape implemented

When installing Theano on a 64 bit computer, it uses float64 as default. Running through the hello world example, this causes an error where the training assumes the input is float32.

from opendeep.models.container import Prototype
from opendeep.models.single_layer.basic import BasicLayer, SoftmaxLayer
from opendeep.optimization.adadelta import AdaDelta
from opendeep.data.standard_datasets.image.mnist import MNIST

mlp = Prototype()
mlp.add(BasicLayer(input_size=28*28, output_size=512, activation='rectifier', noise='dropout'))
mlp.add(BasicLayer(output_size=512, activation='rectifier', noise='dropout'))
mlp.add(SoftmaxLayer(output_size=10))

trainer = AdaDelta(model=mlp, dataset=MNIST())
trainer.train()

This raises:

TypeError: Cannot convert Type TensorType(float64, matrix) (of Variable Subtensor{int32:int32:}.0) into Type TensorType(float32, matrix). You can try to manually convert Subtensor{int32:int32:}.0 into a TensorType(float32, matrix).

Perhaps rename convolutional_network.py to alex_net.py and remove alex_net.py from futures/ to make it more clear?

On line 18 of build/lib/opendeep/monitor/plot.py, symbol cursession is imported from bokeh.plotting:

from bokeh.plotting import (curdoc, cursession, figure, output_server, push, show)

The symbol cursession no longer exists in the latest bokeh version (0.11) available via pip. Work around: install version 0.10.

If I init MemoryDataset it raise

OD_DATA = datasets.MemoryDataset(train_np, valid_X=val_np)

ValueError Traceback (most recent call last)
in ()
----> 1 OD_DATA = datasets.MemoryDataset(train_np, valid_X=val_np)

/home/retina18/Downloads/opendeep/opendeep/data/dataset.pyc in init(self, train_X, train_Y, valid_X, valid_Y, test_X, test_Y)
283 self.train_Y = sharedX(numpy.array(train_Y))
284
--> 285 if valid_X:
286 valid_X = numpy.array(valid_X)
287 self._valid_shape = valid_X.shape

Trivial fix here, the import can be done as:

Softmax as SoftmaxLayer

in opendeep/models/multi_layer/convolutional_network.py

or the name change could be reverted, as all of the documentation uses the SoftmaxLayer name.

HTH

Howdy! OpenDeep looks great and I'm keen to get it working on Kaggle Scripts. I had a go at building it in our Python Docker container, but hit a string/bytes error in the setup script:

File "setup.py", line 49, in <module>
  long_description=read('README.rst'),
File "setup.py", line 28, in read
  return sep.join(buf)
TypeError: sequence item 0: expected str instance, bytes found

Tried a few hacky workarounds, but couldn't get it to land. Is this a Python3 issue? Any ideas on how to fix it?

Hi!
I just installed OpenDeep for a workshop at the OSDC, but I am getting some importing errors:

ImportError: cannot import name DenoisingAutoencoder
ImportError: cannot import name SoftmaxLayer

Do you have any suggestions as to what I am doing wrong?

Thanks,
Rick

Google's release of Tensorflow (https://github.com/tensorflow/tensorflow, http://tensorflow.org/, http://googleresearch.blogspot.com/2015/11/tensorflow-googles-latest-machine_9.html), invites a comparison to OpenDeep, as they seem to be aimed at a similar niche (eg. both research purposes and production deployments).

Some comparison between the two should be added to the README and docs.

I tried your sample code provided under the title "Tutorial: Your First Model". However, it seems that there is no such file called "cost" in the directory "utils". As a result, I fail to import binary_crossentropy from opendeep.utils.cost.

Hi!
Great library!!
Was wondering if there were plans for a multivariate time series example.
For a multi sequence vectors to predict a single sequence output.
ie Time series prediction with multiple sequences input mapped to a single output.
Many thanks,
Best,
Andrew

On line 111 of opendeep/models/container/prototype.py, current inputs is calculated as follows:

current_inputs = zip(previous_out_sizes, previous_outs)

In Python2, zip() returns a list, but in Python3 it returns an iterable zip object. This causes a failure later, when calling code tries to index the result. A simple fix is to wrap the expression in list().

Would you provide an example to use simply a numpy array as a input to predefined model?