Created
April 9, 2012 13:23
-
-
Save benanne/2343378 to your computer and use it in GitHub Desktop.
theano GPU convolution error reproduction
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| import theano | |
| import theano.tensor as T | |
| import numpy as np | |
| def local_feature_extractor(x, W, b, shape_info=None): | |
| xr = x.dimshuffle(0, 2, 'x', 1) # in: (num_examples, num_input_features, 1, num_timesteps) | |
| Wr = W.dimshuffle(0, 1, 'x', 2) # filters: (num_output_features, num_input_features, 1, width) | |
| # the output of the convolution should be equal in length to the input | |
| len_left = (Wr.shape[3] - 1) / 2 | |
| len_right = Wr.shape[3] - 1 - len_left | |
| zeros_buffer = T.zeros((xr.shape[0], xr.shape[1], xr.shape[2], xr.shape[3] + Wr.shape[3] - 1)) | |
| xr_padded = T.set_subtensor(zeros_buffer[:, :, :, len_left:-len_right], xr) | |
| if shape_info is None: | |
| image_shape = None | |
| filter_shape = None | |
| else: | |
| image_shape = (shape_info['num_examples'], shape_info['num_input_features'], 1, shape_info['num_timesteps'] + shape_info['width'] - 1) | |
| filter_shape = (shape_info['num_output_features'], shape_info['num_input_features'], 1, shape_info['width']) | |
| outr = T.nnet.conv.conv2d(xr_padded, Wr, border_mode='valid', image_shape=image_shape, filter_shape=filter_shape) | |
| out = outr[:,:,0,:].dimshuffle(0, 2, 1) | |
| return T.nnet.sigmoid(out + b.dimshuffle('x', 'x', 0)) | |
| # the following always fails: | |
| num_examples = 100 | |
| num_input_features = 513 | |
| num_timesteps = 1280 | |
| width = 8 | |
| num_output_features = 200 | |
| # this works: | |
| """ | |
| num_examples = 100 | |
| num_input_features = 513 | |
| num_timesteps = 500 | |
| width = 8 | |
| num_output_features = 10 | |
| """ | |
| shape_info = { | |
| 'num_examples': num_examples, | |
| 'num_input_features': num_input_features, | |
| 'num_timesteps': num_timesteps, | |
| 'width': width, | |
| 'num_output_features': num_output_features, | |
| } | |
| x = T.tensor3('x') | |
| W = theano.shared(value=np.random.normal(0, 1, (num_output_features, num_input_features, width)).astype(theano.config.floatX), name='W') | |
| b = theano.shared(value=np.random.normal(0, 1, (num_output_features,)).astype(theano.config.floatX), name='b') | |
| feats = local_feature_extractor(x, W, b, shape_info) | |
| print "compiling" | |
| f = theano.function([x], feats) | |
| print "evaluating" | |
| v = np.random.normal(0, 1, (num_examples, num_timesteps, num_input_features)).astype(theano.config.floatX) | |
| out = f(v) # BOOM |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment