Anant Gupta
Hi, Should I add the `resume training` feature? I guess the option to save at intervals is already present (via model_save_interval).
Currently, yes. However, it would be very easy to simply do a padding to account for dynamic length.
When convolving inputs, the zero-padding added to the top rows of input layer makes sure that a hidden state does not contain information from future words.
@ruotianluo Zero padding is used in every layer to keep the layer size same: https://github.com/anantzoid/Language-Modeling-GatedCNN/blob/master/model.py#L62 The zero padding I referred to in the above comment is the extra padding required...
Since there is no pooling, the height and width of the output layer remains the same. The depth is also kept constant for each layer, but can be modified to...
That's an interesting observation. However, it's mentioned in the paper that X (input to any hidden layer h) has the dimension Nxm, and this input could be either word embeddings...
#119 addresses some of the mentioned issues.
You can pass the activation_fn in conv2d as `relu`, `sigmoid`, `tanh` etc. `binarize` function converts pixel intensities to binary values. Try declaring an array: `ar = np.asarray([[0.12,1.2], [1.1, 0.5]])` and...
I'm not sure why the memory blows up but a quick fix is to reduce the batch size.
There's a reproducibility issue on Task 19 as well. Running the code for 53041 global steps reports: "Best step: 37440 with accuracy = 0.412", i.e., 0.6 error, while the repo...