DeformConv2d Cannot be Quantized

🐛 Describe the bug

Operator Metatype DeformConv2dOp is mentioned in nncf/nncf/torch/graph/operator_metatypes.py which uses the namespace torch.nn.functional whereas the function deform_conv2d belongs to torchvision.ops.deform_conv2d. As seen in the code output attached below, this deformable convolution was not quantized.

Environment

about-time==4.2.1 absl-py==2.1.0 accelerate==0.28.0 accuracy_checker @ git+https://github.com/openvinotoolkit/open_model_zoo.git@37f60eb7fe1dcdedc552b2fb184d646723ed5e80#subdirectory=tools/accuracy_checker addict==2.4.0 aiohttp==3.9.5 aiosignal==1.3.1 alive-progress==3.1.5 async-timeout==4.0.3 attrs==23.2.0 autograd==1.6.2 certifi==2024.7.4 cfgv==3.4.0 charset-normalizer==3.3.2 cma==3.2.2 coloredlogs==15.0.1 contourpy==1.2.1 coverage==7.5.4 cycler==0.12.1 datasets==2.14.7 defusedxml==0.7.1 Deprecated==1.2.14 dill==0.3.7 distlib==0.3.8 efficientnet-pytorch==0.7.1 evaluate==0.3.0 exceptiongroup==1.2.1 execnet==2.1.1 fastcore==1.5.48 fastdownload==0.0.7 fastprogress==1.0.3 filelock==3.15.4 flatbuffers==24.3.25 fonttools==4.53.0 frozenlist==1.4.1 fsspec==2023.10.0 future==1.0.0 grapheme==0.6.0 grpcio==1.64.1 huggingface-hub==0.23.4 humanfriendly==10.0 identify==2.6.0 idna==3.7 iniconfig==2.0.0 Jinja2==3.1.4 joblib==1.4.2 jsonschema==4.22.0 jsonschema-specifications==2023.12.1 jstyleson==0.0.2 kiwisolver==1.4.5 lightning-utilities==0.11.3.post0 Markdown==3.6 markdown-it-py==3.0.0 MarkupSafe==2.1.5 matplotlib==3.9.1 mdurl==0.1.2 mpmath==1.3.0 multidict==6.0.5 multiprocess==0.70.15 natsort==8.4.0 networkx==3.3 ninja==1.11.1.1 -e git+https://github.com/anzr299/nncf.git@bfc94b9d1078024b04246f8fc41106582c227f7b#egg=nncf nodeenv==1.9.1 numpy==1.26.4 nvidia-cublas-cu12==12.1.3.1 nvidia-cuda-cupti-cu12==12.1.105 nvidia-cuda-nvrtc-cu12==12.1.105 nvidia-cuda-runtime-cu12==12.1.105 nvidia-cudnn-cu12==8.9.2.26 nvidia-cufft-cu12==11.0.2.54 nvidia-curand-cu12==10.3.2.106 nvidia-cusolver-cu12==11.4.5.107 nvidia-cusparse-cu12==12.1.0.106 nvidia-nccl-cu12==2.20.5 nvidia-nvjitlink-cu12==12.5.82 nvidia-nvtx-cu12==12.1.105 onnx==1.16.0 onnxruntime==1.17.1 opencv-python==4.10.0.84 openvino==2024.2.0 openvino-telemetry==2024.1.0 packaging==24.1 pandas==2.2.2 pillow==10.4.0 platformdirs==4.2.2 pluggy==1.5.0 pre-commit==3.2.2 protobuf==4.25.3 psutil==6.0.0 pyarrow==16.1.0 pyarrow-hotfix==0.6 pycocotools==2.0.7 pydot==2.0.0 Pygments==2.18.0 pymoo==0.6.1.1 pyparsing==3.1.2 pytest==8.0.2 pytest-cov==4.1.0 pytest-dependency==0.6.0 pytest-mock==3.12.0 pytest-xdist==3.5.0 python-dateutil==2.9.0.post0 pytz==2024.1 PyYAML==6.0.1 referencing==0.35.1 regex==2024.5.15 requests==2.32.3 responses==0.18.0 rich==13.7.1 rpds-py==0.18.1 safetensors==0.4.3 scikit-learn==1.5.1 scipy==1.14.0 six==1.16.0 sympy==1.12.1 tabulate==0.9.0 tensorboard==2.17.0 tensorboard-data-server==0.7.2 threadpoolctl==3.5.0 timm==0.9.2 tokenizers==0.15.2 tomli==2.0.1 torch==2.3.0 torchmetrics==1.0.1 torchvision==0.18.0 tqdm==4.66.4 transformers==4.38.2 triton==2.3.0 typing_extensions==4.12.2 tzdata==2024.1 urllib3==2.2.2 virtualenv==20.26.3 Werkzeug==3.0.3 wrapt==1.16.0 xxhash==3.4.1 yarl==1.9.4

Minimal Reproducible Example


#Sample Deformable Convolution Network Model Definition which utilizes torchvision.ops.DeformConv2d
class DCNV2(nn.Module):
    def __init__(self):
        super(DCNV2, self).__init__()
        self.deform_conv1 = DeformConv2d(in_channels=3, out_channels=32, kernel_size=3, stride=1, padding=1)
        self.offset_conv1 = nn.Conv2d(3, 18, kernel_size=3, stride=1, padding=1)

        self.deform_conv2 = DeformConv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=1)
        self.offset_conv2 = nn.Conv2d(32, 18, kernel_size=3, stride=1, padding=1)

        self.fc = nn.Linear(64 * 8 * 8, 10)

    def forward(self, x):
        offset1 = self.offset_conv1(x)
        x = self.deform_conv1(x, offset1)
        x = nn.ReLU()(x)
        x = nn.MaxPool2d(kernel_size=2, stride=2)(x)

        offset2 = self.offset_conv2(x)
        x = self.deform_conv2(x, offset2)
        x = nn.ReLU()(x)
        x = nn.MaxPool2d(kernel_size=2, stride=2)(x)

        x = x.view(x.size(0), -1)
        x = self.fc(x)
        return x

model = DCNV2()

#Dummy data for our model
class RandomDataset(torch.utils.data.Dataset):
    def __getitem__(self, index):
        return torch.randn(3, 32, 32), torch.tensor(0)

    def __len__(self):
        return 1000

data_loader = torch.utils.data.DataLoader(RandomDataset(), batch_size=32)

#transform function for the calibration dataset
def transform_fn(data_item):
    images, _ = data_item
    return images

calibration_dataset = nncf.Dataset(data_loader, transform_fn)
model.eval()
quantized_model = nncf.quantize(model, calibration_dataset)
print(quantized_model)
'''
OUTPUT: 

DCNV2(
  (deform_conv1): DeformConv2d(3, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (offset_conv1): Conv2d(3, 18, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (deform_conv2): DeformConv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (offset_conv2): Conv2d(32, 18, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  (fc): Linear(in_features=4096, out_features=10, bias=True)
  (_nncf): NNCFNetworkInterface(
    (external_quantizers): ModuleDict(
      (/nncf_model_input_0|OUTPUT): SymmetricQuantizer(bit=8, ch=False)
      (DCNV2/Conv2d[offset_conv1]/conv2d_0|INPUT1): SymmetricQuantizer(bit=8, ch=True)
      (DCNV2/Conv2d[offset_conv2]/conv2d_0|INPUT1): SymmetricQuantizer(bit=8, ch=True)
      (DCNV2/Linear[fc]/linear_0|INPUT1): SymmetricQuantizer(bit=8, ch=True)
    )
  )
)
'''

Are you going to submit a PR?

[ ] Yes I'd like to help by submitting a PR!

openvinotoolkit / nncf