search

ssampang / im2latex

Tensorflow port of im2markup project
128 stars 23 forks source link

CNN order reversed #2

Open Randnx opened 7 years ago

Randnx commented 7 years ago

Hi, I am reading your excellent code, but find the CNN layer order is reversed from the paper author's code(https://github.com/harvardnlp/im2markup/blob/master/src/model/cnn.lua). Is this a problem?

ssampang commented 7 years ago

Yeah I misread the architecture lol. Haven't gotten around to fixing it, but the code still worked albeit reduced accuracy.

Randnx commented 7 years ago

I'm also a little confused by this comment: "Assuming a batch size of 20, about one fourth of the dataset consisted of images that were too large to fit on a Titan X GPU.". Can a smaller batch size make full dataset fit on a Titan X GPU?

ssampang commented 7 years ago

Yes - since we're dealing with recurrent nets, long equations will result in more timesteps per equation, which means the training process will require more VRAM to store the data of each timestep in memory. What I found was that if we sort the equations by length (essentially character count), then the top 25% of equations (by length) were too "large" to train with on a titan X when using a batch size of 20. If you reduce the batch size, fewer training examples will allow more VRAM per training example, which will allow the longer examples to fit in the available VRAM.

Randnx commented 7 years ago

Thanks, that is a great help (: