Problems with MLP and DenseDesignMatrixPyTables

20 views

Skip to first unread message

xser...@gmail.com

unread,

Sep 15, 2016, 12:27:04 PM9/15/16

to pylearn-users

Dear pylearn2 friends,

I am trying to feed a large dataset (600000 rgb images) to a MLP using SGD as trainer. I have created a subclass of DenseDesignMatrixPyTables, and stored the data using init_hdf5 and fill_hdf5. However, when feeding the resulting dataset to the trainer the following error arises:

Traceback (most recent call last):
File "individual_classification.py", line 157, in <module>
results = trainAndTest(datasets, path_models, path_img, path_pairs, experiment_params, first_time)
File "/home/workspace/xserraa/face_identification/scripts/libraries/Experiment.py", line 76, in trainAndTest
global_model.trainCNNModel(train_dataset,validation_dataset,validation_dataset,paths)
File "/home/workspace/xserraa/face_identification/scripts/libraries/GlobalModel.py", line 160, in trainCNNModel
trainer.train(dataset=train_dataset)
File "/home/workspace/xserraa/pylearn/pylearn2/training_algorithms/sgd.py", line 453, in train
for batch in iterator:
File "/home/workspace/xserraa/pylearn/pylearn2/utils/iteration.py", line 984, in next
rval = self._fallback_next(next_index)
File "/home/workspace/xserraa/pylearn/pylearn2/utils/iteration.py", line 1002, in _fallback_next
for data, fn in safe_izip(self._raw_data, self._convert)
File "/home/workspace/xserraa/pylearn/pylearn2/utils/iteration.py", line 1002, in <genexpr>
for data, fn in safe_izip(self._raw_data, self._convert)
File "/usr/local/lib/python2.7/dist-packages/tables/array.py", line 656, in __getitem__
coords = self._point_selection(key)
File "/usr/local/lib/python2.7/dist-packages/tables/leaf.py", line 561, in _point_selection
coords[idx] = (coords + self.shape)[idx]
ValueError: operands could not be broadcast together with shapes (51,) (2,)

Being 51 the number of instances (I have used a small dataset to speed testing the code). I do not know which could be the error, any help would be really appreciated, and please feel free to ask for any clarification.