What is the relationship between lambda and batchsize?
How do you determine the lambda?
Good question. Actually, the batch size I used is nothing special. I think the best lambda should be determined by experiment.
@goodluckcwl Thank you very much. Does the lambda remain the same? layer { name: "fc6" type: "MarginInnerProduct" bottom: "fc5" bottom: "label" top: "fc6" top: "lambda" param { lr_mult: 1 decay_mult: 1 } margin_inner_product_param { num_output: 10575 type: QUADRUPLE weight_filler { type: "xavier" } base: 3.6 gamma: 0.12 power: 1 lambda_min: 3.6 iteration: 0 } }
Note that this prototxt is used for fine-tune only. Besides, I don't known whether a larger batch size makes sense or not. You can have a try. For example, set batch_size=256.