Matmul op for relay

yongwww · 2019-02-01 00:41:07 UTC

Feel free to close the topic. Just wanna know the reason why we don’t implement matmul op for relay, I know we can use dense instead, is there any other reason?

junrushao1994 · 2019-02-02 23:28:29 UTC

What’s the diff between the two? bias?

yongwww · 2019-02-04 08:30:19 UTC

yup, bias. Y = XW^T + b

junrushao1994 · 2019-02-04 08:44:55 UTC

I guess that you don’t have to provide the bias term, so that dense degenerates to matmul.

github.com

dmlc/tvm/blob/3a75b13d663a3bb14e11f5b7fd91980163f9dd75/topi/include/topi/nn/dense.h#L47


auto in_dim = data->shape[1];
auto out_dim = weight->shape[0];


auto k = tvm::reduce_axis(Range(0, in_dim), "k");
auto matmul = tvm::compute(
  { batch, out_dim },
  [&](Var i, Var j) {
    return tvm::sum(data(i, k) * weight(j, k), { k });
  }, "tensor", "dense");


if (bias.defined()) {
  matmul = tvm::compute(
    { batch, out_dim },
    [&](Var i, Var j) {
      return matmul(i, j) + bias(j);
    }, "tensor", kBroadcast);
}


return matmul;
}

yongwww · 2019-02-04 18:12:23 UTC

yeah, dense will do the same thing as matmul without specifying bias. I like the name tvm.relay.matmul more than tvm.relay.nn.dense for matrix mulitiplication. Anyway, dense is good.

junrushao1994 · 2019-02-04 18:24:02 UTC

Sure. That sounds reasonable.

junrushao1994 · 2019-02-04 18:34:28 UTC

I guess I misread relay for topi…anyway…

jroesch · 2019-02-04 21:10:14 UTC

We could expose matmul as a synonym. I think this might actually be a good idea, the overloaded use of dense has confused many users (including me when I started working on TVM).

yongwww · 2019-02-04 21:26:00 UTC

thanks, I agree. Confused me, too

yongwww · 2019-02-04 22:04:51 UTC

I will send out a pr to expose matmul then