Hi Team,
I am trying run ResNet101 pytorch model through TVM. This model I took from InsightFace-v2, and in the parser stage I hit with below error from the error message looks like the issue is related to dimension conflicts in PRELU layer.
Error message:
TVMError: Error(s) have occurred. The program has been annotated with them:
In main:
v0.0.4
fn (%input0: Tensor[(1, 3, 112, 112), float32], %v39: Tensor[(64, 3, 1, 1), float32], %v62: Tensor[(64), float32], %v61: Tensor[(64), float32], %v60: Tensor[(64), float32], %v59: Tensor[(64), float32], %v68: Tensor[(64, 1, 1, 3), float32], %v91: Tensor[(64), float32], %v90: Tensor[(64), float32], %v89: Tensor[(64), float32], %v88: Tensor[(64), float32], %v97: Tensor[(64, 1, 3, 1), float32], %v120: Tensor[(64), float32], %v119: Tensor[(64), float32], %v118: Tensor[(64), float32], %v117: Tensor[(64), float32], %v126: Tensor[(1), float32]) {
%0 = nn.conv2d(%input0, %v39, padding=[0, 0, 0, 0], channels=64, kernel_size=[1, 1]);
%1 = nn.batch_norm(%0, %v62, %v61, %v60, %v59);
%2 = %1.0;
%3 = reshape(%v68, meta[relay.Constant][0], newshape=[64, 1, 1, 3]);
%4 = nn.conv2d(%2, %3, padding=[0, 0, 0, 0], groups=64, channels=64, kernel_size=[1, 3]);
%5 = nn.batch_norm(%4, %v91, %v90, %v89, %v88);
%6 = %5.0;
%7 = reshape(%v97, meta[relay.Constant][1], newshape=[64, 1, 3, 1]);
%8 = nn.conv2d(%6, %7, padding=[0, 0, 0, 0], groups=64, channels=64, kernel_size=[3, 1]);
%9 = nn.batch_norm(%8, %v120, %v119, %v118, %v117);
%10 = %9.0;
%11 = nn.prelu(%10, %v126) in particular dimension 0 conflicts 64 does not match 1; unable to unify: Tensor[(64), float32] and Tensor[(1), float32]; ;
** nn.max_pool2d(%11, pool_size=[2, 2], strides=[2, 2], padding=[0, 0, 0, 0])**
}
// meta data omitted. you can use show_meta_data=True to include meta data
Is it known issue? am I missing something here.
Kindly help me to solve this issue.
Note:
- I am able to run this model in standalone pytorch framework
- My TVM Commit ID: commit e369c5a9cbacb926ca7b95ebc4ae01a6de33c6cd
Thanks and Regards, Raju