why the batch_size is a parameter of the layer

[menglin0320]

why the batch_size is a parameter of the layer

[junhyukoh]

Lstm layer assumes that the input (0-th bottom) looks like [N x T][...] where N is batch_size, T is length, and [...] is the feature dimension.
By default, batch_size = 1, so the layer assumes that the bottom shape is [T][...].
If you want to use batch size larger than 1, you should 1) provide input in [N x T][...] shape and 2) specify "batch_size: N" into your prototxt. Otherwise, lstm layer will assume that your input is NxT-long sequence with batch size of 1.

Providing input in [NxT][...] shape is a little bit confusing.
My suggestion is to fill your input data like [N][T][...] and use "Reshape" layer to make it [NxT][...] before lstm layer.

I hope this is clear to you.

[menglin0320]

I changed my question. Can you please help?? @junhyukoh

[menglin0320]

Also, what is clip_threshold used for... Also I'm confused about clip,isn't that similar to different batches, what is the difference