Hi, Community!
I’m new to the torch-mlir, and I’m trying it with a simple op, but met such error:
error: unsupported by backend lowering:
torch.operatorop
note: see current operation: %41 = “torch.operator”(%38, %40) {name = “aten.maximum”} : (!torch.tensor<[?,?],f32>, !torch.tensor<[?,?],f32>) → !torch.tensor
note: this is likely due to a missing op that needs to be generated by torch_ods_gen.py
Traceback (most recent call last):
File “/home/SENSETIME/mupei/Projects/Coder_tvm/torch-mlir/examples/iou_of.py”, line 119, in
compiled = compile_module(mlp_module)
File “/home/SENSETIME/mupei/Projects/Coder_tvm/torch-mlir/examples/iou_of.py”, line 99, in compile_module
pm.run(mb.module)
RuntimeError: Failure while executing pass pipeline.
As far as I’m concerned, such an error means there’s something wrong in the pass manager of “torch-backend-to-linalg-on-tensors-backend-pipeline”. But I have no idea how to fix that or is there an error with my code?
Thanks!
Here’s the python operator code:
def iou_of(bbox1, bbox2):
"""get iou of two boxes.
Args:
bbox1 array(n, 4): (x1, y1, x2, y2)
bbox2 array(n, 4): (x1, y1, x2, y2)
return: the value of iou
"""
area1 = (bbox1[:, 2] - bbox1[:, 0]) * (bbox1[:, 3] - bbox1[:, 1])
area2 = (bbox2[:, 2] - bbox2[:, 0]) * (bbox2[:, 3] - bbox2[:, 1])
lt = np.maximum(bbox1[:, :2], bbox2[:, :2])
rb = np.minimum(bbox1[:, 2:], bbox2[:, 2:])
overlap_coord = (rb - lt).clip(0)
overlap = overlap_coord[:, 0] * overlap_coord[:, 1]
union = area1 + area2 - overlap
return overlap / union
Follow with the examples, I have such whole python code(sorry for the long code):
import torch
import torchvision
import numpy as np
from torch_mlir.dialects.torch.importer.jit_ir import ClassAnnotator, ModuleBuilder
from torch_mlir.dialects.torch.importer.jit_ir.torchscript_annotations import extract_annotations
from torch_mlir_e2e_test.torchscript.annotations import annotate_args, export
from torch_mlir.passmanager import PassManager
from torch_mlir_e2e_test.linalg_on_tensors_backends.refbackend import RefBackendLinalgOnTensorsBackend
def iou_of(bbox1, bbox2):
"""get iou of two boxes.
Args:
bbox1 array(n, 4): (x1, y1, x2, y2)
bbox2 array(n, 4): (x1, y1, x2, y2)
return: value of iou
"""
area1 = (bbox1[:, 2] - bbox1[:, 0]) * (bbox1[:, 3] - bbox1[:, 1])
area2 = (bbox2[:, 2] - bbox2[:, 0]) * (bbox2[:, 3] - bbox2[:, 1])
lt = np.maximum(bbox1[:, :2], bbox2[:, :2])
rb = np.minimum(bbox1[:, 2:], bbox2[:, 2:])
overlap_coord = (rb - lt).clip(0)
overlap = overlap_coord[:, 0] * overlap_coord[:, 1]
union = area1 + area2 - overlap
return overlap / union
class IouOfModule(torch.nn.Module):
def __init__(self):
super().__init__()
# Reset seed to make model deterministic.
torch.manual_seed(0)
self.fc0 = torch.nn.Linear(3, 5)
self.tanh0 = torch.nn.Tanh()
@export
@annotate_args([
None,
([1024, 4], torch.float32, True),
([1024, 4], torch.float32, True),
])
def forward(self, bbox1, bbox2):
area1 = (bbox1[:, 2] - bbox1[:, 0]) * (bbox1[:, 3] - bbox1[:, 1])
area2 = (bbox2[:, 2] - bbox2[:, 0]) * (bbox2[:, 3] - bbox2[:, 1])
lt = torch.maximum(bbox1[:, :2], bbox2[:, :2])
rb = torch.minimum(bbox1[:, 2:], bbox2[:, 2:])
overlap_coord = (rb - lt).clip(0)
overlap = overlap_coord[:, 0] * overlap_coord[:, 1]
union = area1 + area2 - overlap
return overlap / union
# BACKEND = LinalgOnTensorsTosaBackend()
BACKEND = RefBackendLinalgOnTensorsBackend()
def compile_module(program: torch.nn.Module):
"""Compiles a torch.nn.Module into an compiled artifact.
This artifact is suitable for inclusion in a user's application. It only
depends on the rebackend runtime.
"""
## Script the program.
scripted = torch.jit.script(program)
print("-------------------------------graph-------------------------------")
print(scripted.graph)
print("-------------------------------code-------------------------------")
print(scripted.code)
## Extract annotations.
class_annotator = ClassAnnotator()
extract_annotations(program, scripted, class_annotator)
## Import the TorchScript module into MLIR.
mb = ModuleBuilder()
mb.import_module(scripted._c, class_annotator)
print("-------------------------------init IR-------------------------------")
mb.module.dump()
## Lower the MLIR from TorchScript to RefBackend, passing through linalg-on-tensors.
pm = PassManager.parse('torchscript-module-to-torch-backend-pipeline', mb.module.context)
pm.run(mb.module)
print("-------------------------------torch-back-pipeline-------------------------------")
mb.module.dump()
pm = PassManager.parse('torch-backend-to-linalg-on-tensors-backend-pipeline', mb.module.context)
pm.run(mb.module)
print("-------------------------------linalg-------------------------------")
mb.module.dump()
# pm = PassManager.parse('builtin.func(convert-torch-to-linalg)', mb.module.context)
# pm.run(mb.module)
# print("-------------------------------test2-------------------------------")
# mb.module.dump()
## Invoke RefBackend to compile to compiled artifact form.
compiled_module = BACKEND.compile(mb.module)
print("-------------------------------after-compile-------------------------------")
compiled_module.dump()
return compiled_module
if __name__ == '__main__':
print("---------------main-------------------")
mlp_module = IouOfModule()
# Create the module and compile it.
compiled = compile_module(mlp_module)
# compiled.dump()
# Loads the compiled artifact into the runtime
jit_module = BACKEND.load(compiled)
# Run it!
jit_module.forward(torch.rand(1024, 4).numpy(), torch.rand(1024, 4).numpy())
