@Jiali this could be due to the fact that the ratio of valid schedules over an
entire search space is relatively low, so population sampling cannot find any
valid schedule. If the evolutionary search is based on a set of invalid
schedules, it is highly possible that it will be trapped in inval
@merrymercy Alternatively, if we are applying LSTM to a fixed length sequence,
we can use tvm.scan. This requires that we deal with the IterVarType correctly.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/auto-scheduling-for-lstm-operator/8158/7)
to respond.
You are receiving this b
@Jiali
For the problem of the slow compilation of LSTM, you can open another post for
it.
I imagine that you are unfolding the LSTM cells so the relay program is very
long.
Since relay supports control flow, you do not need to unfold all of them. You
can open another post to ask how to do thi
Thanks for your fix. And I run the fixed code, the above error has been fixed~
For the input shape as (2,1,240), the schedule can be generated. But for the
data input shape as (3,1,240), there is another error : **Process finished with
exit code 136 (interrupted by signal 8: SIGFPE**.
```
---
Thanks for your reply~
The reason of using large graph is that the build time is too long for small
subgraph lstm computation function. The build time of the lstm computation
declaration as small subgraph will be more than 3723 s. Because the relay graph
for this computation declaration is so
Thanks for reporting. This PR
(https://github.com/apache/incubator-tvm/pull/6683) fixed this bug.
However, typically for large graphs like this. We don't treat them as a single
search task. Treating it as a single large graph makes the search very
inefficient.
For large graphs, we use Relay
Thanks! We have not tried such ops before, this case seems interesting.
I'll try your code and figure out where the possible bug occurs.
---
[Visit
Topic](https://discuss.tvm.apache.org/t/auto-scheduling-for-lstm-operator/8158/2)
to respond.
You are receiving this because you enabled mai
Hi Experts, @Lianminzheng @jcf94
I tried to define a new operator for lstm network.The computation declaration
for lstm op has been tested and it is correct. Now I want to use
auto-scheduling to automatically generate a large search space and find a good
schedule in the space. But it can no
