[TOPI] add dilation operators #316
Conversation
|
thanks! can we add a unittest case to verify this? |
|
To be consistent with existing interface, (e.g. mxnet), we should take in a Array of Expr(can be constant) as argument so we can pass tuple of int directly as dilate into it. |
|
It is also helpful to add a padding stage, as per #294 |
|
To verify logical correctness of the compute declaration, we don't need cuda schedule, just a llvm cpu schedule of small workload will suffice |
|
Actually, I think it is helpful to directly pass in a ndim Array of Expr as dilate argument, to specify dilate on each dimension. This will avoid the need of specifying layout |
topi/python/topi/nn/dilate.py
Outdated
| Output = tvm.compute( | ||
| (N, (H-1)*stride_h+1, (W-1)*stride_w+1, C), | ||
| lambda n, h, w, c: tvm.select( | ||
| tvm.all(tvm.make.EQ(h%stride_h, 0), tvm.make.EQ(w%stride_w, 0)), |
There was a problem hiding this comment.
we can use (h%stride_h).equal(0)
topi/python/topi/nn/dilate.py
Outdated
| (N, (H-1)*stride_h+1, (W-1)*stride_w+1, C), | ||
| lambda n, h, w, c: tvm.select( | ||
| tvm.all(tvm.make.EQ(h%stride_h, 0), tvm.make.EQ(w%stride_w, 0)), | ||
| Input(n, h/stride_h, w/stride_w, c), tvm.const(0.0)), |
There was a problem hiding this comment.
This allows the expression to work for other data types
topi/python/topi/nn/dilate.py
Outdated
| Output : tvm.Tensor | ||
| Output tensor, layout is NCHW. | ||
| """ | ||
| N, C, H, W = get_const_tuple(Input.shape) |
There was a problem hiding this comment.
For this expression to work, we might not need const input shapes. The new dimension calculation can be done symbolically, then follows a tvm.ir_pass.Simplfy.
topi/python/topi/nn/dilate.py
Outdated
| Parameters | ||
| ---------- | ||
| Input : tvm.Tensor | ||
| 4-D, can be any layout. |
There was a problem hiding this comment.
We can directly ask user to pass in nd tensor
topi/python/topi/nn/dilate.py
Outdated
| 4-D, the same layout as Input. | ||
| """ | ||
| A, B, C, D = Input.shape | ||
| sa, sb, sc, sd = strides |
There was a problem hiding this comment.
Use python's list unpack tricks to support nd tensor
function(*list)
|
I convert caffe model to mxnet model, I use mxnet as frontend, but I get a error: FILE nnvm/python/nnvm/top/nn.py, in compute_conv2d. Can I get some help ? |
|
@sgxu You can compose dilated conv using topi.nn.dilate and topi.nn.conv2d It won't be difficult, just not implemented now. |
…pache#316) This PR removes the `global_symbol` linkage added by Relay Translator. It also fixes unaddressed comments of apache#262. All tests can pass locally and I believe it is safe to merge this PR directly.
…pache#316) This PR removes the `global_symbol` linkage added by Relay Translator. It also fixes unaddressed comments of apache#262. All tests can pass locally and I believe it is safe to merge this PR directly.
…pache#316) This PR removes the `global_symbol` linkage added by Relay Translator. It also fixes unaddressed comments of apache#262. All tests can pass locally and I believe it is safe to merge this PR directly.
can be (has been proven) useful in implementing backward propagation of depthwise convolution #294