oviazlo opened a new issue, #13070:
URL: https://github.com/apache/tvm/issues/13070
When running unit tests from
`tests/python/frontend/pytorch/test_fx_quant.py` with `Torch 1.12.0` and
`Torchvision 0.13.0` environment, `test_ssd_vgg` test fails.
### Expected behavior
Unit test to run successfully.
### Actual behavior
Unit tests fail with the following error message:
```python
test_fx_quant.py:41: in quantize_and_build
mod, _ = relay.frontend.from_pytorch(script_module, [(input_name,
inp.shape)])
../../../../python/tvm/relay/frontend/pytorch.py:4626: in from_pytorch
outputs = converter.convert_operators(tmp, outputs, ret_name)
../../../../python/tvm/relay/frontend/pytorch.py:3998: in convert_operators
relay_out = relay_op(
../../../../python/tvm/relay/frontend/qnn_torch.py:627: in _impl
inputs[0], _expr.const(inputs[1]), _expr.const(inputs[2]),
out_dtype="uint8", axis=axis
../../../../python/tvm/relay/expr.py:508: in const
value.dtype, None
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
_ _
self = Var(model_backbone_scale_0, ty=TensorType([], float32)), name =
'dtype'
def __getattr__(self, name):
# specially check handle since
# this is required for PackedFunc calls
if name == "handle":
raise AttributeError("handle is not set")
try:
return _ffi_node_api.NodeGetAttr(self, name)
except AttributeError:
> raise AttributeError("%s has no attribute %s" %
(str(type(self)), name)) from None
E AttributeError: <class 'tvm.relay.expr.Var'> has no attribute
dtype
../../../../python/tvm/runtime/object.py:67: AttributeError
```
The problem is happening at the moment of executing the following code:
```python
relay_out = relay_op(
inputs, _get_input_types(op_node, outputs,
default_dtype=self.default_dtype)
)
```
with the following arguments:
```python
op_node = {Node} %quantize_per_tensor_14 : QUInt8(512, strides=[1],
requires_grad=0, device=cpu) = aten::quantize_per_tensor(%scale_weight,
%model_backbone_scale_0, %model_backbone_zero_point_0, %217) #
<eval_with_key>.10:25:0
inputs = {list:4} [Var(model.backbone.scale_weight, ty=TensorType([512],
float32)), Var(model_backbone_scale_0, ty=TensorType([], float32)),
Var(model_backbone_zero_point_0, ty=TensorType([], int64)), 13]
```
The problem is that scale (`model_backbone_scale_0`) and point zero
(`model_backbone_zero_point_0`) parameters for the operator
`aten::quantize_per_tensor` has to be canonicalized by function
[inline_input_quant_params_for_fx](https://github.com/apache/tvm/blob/b44f1343a10ccc908de5e65b864012c72d564a7b/python/tvm/relay/frontend/qnn_torch.py#L537).
However, it's not happening because in the implementation of
[inline_input_quant_params_for_fx](https://github.com/apache/tvm/blob/b44f1343a10ccc908de5e65b864012c72d564a7b/python/tvm/relay/frontend/qnn_torch.py#L537)
function only parameters that have `"_input_scale"` or `"_input_zero_point"`
inside their names are canonicalized:
https://github.com/apache/tvm/blob/b44f1343a10ccc908de5e65b864012c72d564a7b/python/tvm/relay/frontend/qnn_torch.py#L567-L584
which is not the case for the parameters with names `model_backbone_scale_0`
and `model_backbone_zero_point_0`.
The following problem can be solved by simply changing keywords from
`"_input_scale"`, `"_input_zero_point"` to `"_scale"`, `"_zero_point"` inside
[inline_input_quant_params_for_fx](https://github.com/apache/tvm/blob/b44f1343a10ccc908de5e65b864012c72d564a7b/python/tvm/relay/frontend/qnn_torch.py#L537)
function. I have tried this solution and the test started to pass
successfully.
### Environment
- TVM version:
[v0.10.0rc0](https://github.com/apache/tvm/releases/tag/v0.10.0rc0)
(https://github.com/apache/tvm/commit/b44f1343a10ccc908de5e65b864012c72d564a7b)
- Ubuntu 20.04.1, x86_64, GNU/Linux 5.15.0-1021-aws
- torch **1.12.0+cpu**
- torchvision **0.13.0+cpu**
### Steps to reproduce
With `Torch 1.12` and `Torchvision 0.13.0` environment, run:
`python -m pytest tests/python/frontend/pytorch/test_fx_quant.py`
### Triage
* frontend:pytorch
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
