search

yusuketomoto / chainer-fast-neuralstyle

Chainer implementation of "Perceptual Losses for Real-Time Style Transfer and Super-Resolution".
MIT License
803 stars 229 forks source link

bigger image size #66

Open artnose opened 7 years ago

artnose commented 7 years ago

I get this error epoch 0 Traceback (most recent call last): File "train.py", line 135, in L_feat = lambda_f * F.mean_squared_error(Variable(feature[2].data), feature_hat[2]) # compute for only the output of layer conv3_3 File "/usr/lib64/python2.7/site-packages/chainer/functions/loss/mean_squared_error.py", line 44, in mean_squared_error return MeanSquaredError()(x0, x1) File "/usr/lib64/python2.7/site-packages/chainer/function.py", line 190, in call self._check_data_type_forward(in_data) File "/usr/lib64/python2.7/site-packages/chainer/function.py", line 273, in _check_data_type_forward type_check.InvalidType(e.expect, e.actual, msg=msg), None) File "/usr/lib/python2.7/site-packages/six.py", line 718, in raise_from raise value chainer.utils.type_check.InvalidType: Invalid operation is performed in: MeanSquaredError (Forward)

Expect: in_types[0].shape == in_types[1].shape Actual: (1, 256, 119, 119) != (1, 256, 118, 118)

what are the restrictions on image size?

6o6o commented 7 years ago

The network subsamples your image twice during transformation, that is, the dimensions shrink 4x at a certain point in the process. So if your resolution is not divisible by 4 it won't necessarily turn out the same size. In fact, it will always be less.

If you really need to use 119, you can explicitly pass outsize argument to Deconvolution2D upon initialization, but it makes the operation less convenient.