AndrewZhaoLuo opened a new issue #8058:
URL: https://github.com/apache/tvm/issues/8058
The main example I have is here where avg pool expects an output with
negative dimensions in the shape:
```
from tvm.relay.testing import densenet
mod, mod_params = densenet.get_workload(classes=5, batch_size=1,
image_shape=(1, 224, 224))
mod_params["data"] = np.random.uniform(-10, 10, (1, 1, 224,
224)).astype("float32")
infer_type_pass = InferType()
out = infer_type_pass(mod)
body_typed = out["main"].body
print(body_typed)
```
Which produces output including:
`%487 = nn.avg_pool2d(%486, pool_size=[7, 7], padding=[0, 0, 0, 0]) /*
ty=Tensor[(1, 1, -5, -5), float32] */;`
*What is expected*:
The type checker will yell at you if it detects nonsensical shapes.
*What can be done*:
This can be done by having the shape relation function for each op check for
negative shapes. Alternatively we can determine that negative shapes are
inherently nonsensical for every operation and put the check more in the infer
type pass.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
