Converting a PyTorch YOLACT Model

You Only Look At CoefficienTs (YOLACT) is a simple, fully convolutional model for real-time instance segmentation. The PyTorch implementation is publicly available in this GitHub repository. The YOLACT++ model is not supported, because it uses deformable convolutional layers that cannot be represented in ONNX format.

Creating a Patch File

Before converting the model, create a patch file for the repository. The patch modifies the framework code by adding a special command-line argument to the framework options. The argument enables inference graph dumping:

  1. Go to a writable directory and create a YOLACT_onnx_export.patch file.

  2. Copy the following diff code to the file:

    From 76deb67d4f09f29feda1a633358caa18335d9e9f Mon Sep 17 00:00:00 2001
    From: "OpenVINO" <openvino@intel.com>
    Date: Fri, 12 Mar 2021 00:27:35 +0300
    Subject: [PATCH] Add export to ONNX
    
    ---
     eval.py                |  5 ++++-
     utils/augmentations.py |  7 +++++--
     yolact.py              | 29 +++++++++++++++++++----------
     3 files changed, 28 insertions(+), 13 deletions(-)
    
    diff --git a/eval.py b/eval.py
    index 547bc0a..bde0680 100644
    --- a/eval.py
    +++ b/eval.py
    @@ -593,9 +593,12 @@ def badhash(x):
         return x
    
     def evalimage(net:Yolact, path:str, save_path:str=None):
    -    frame = torch.from_numpy(cv2.imread(path)).cuda().float()
    +    frame = torch.from_numpy(cv2.imread(path)).float()
    +    if torch.cuda.is_available():
    +        frame = frame.cuda()
         batch = FastBaseTransform()(frame.unsqueeze(0))
         preds = net(batch)
    +    torch.onnx.export(net, batch, "yolact.onnx", opset_version=11)
    
         img_numpy = prep_display(preds, frame, None, None, undo_transform=False)
    
    diff --git a/utils/augmentations.py b/utils/augmentations.py
    index cc7a73a..2420603 100644
    --- a/utils/augmentations.py
    +++ b/utils/augmentations.py
    @@ -623,8 +623,11 @@ class FastBaseTransform(torch.nn.Module):
         def __init__(self):
             super().__init__()
    
    -        self.mean = torch.Tensor(MEANS).float().cuda()[None, :, None, None]
    -        self.std  = torch.Tensor( STD ).float().cuda()[None, :, None, None]
    +        self.mean = torch.Tensor(MEANS).float()[None, :, None, None]
    +        self.std  = torch.Tensor( STD ).float()[None, :, None, None]
    +        if torch.cuda.is_available():
    +            self.mean.cuda()
    +            self.std.cuda()
             self.transform = cfg.backbone.transform
    
         def forward(self, img):
    diff --git a/yolact.py b/yolact.py
    index d83703b..f8c787c 100644
    --- a/yolact.py
    +++ b/yolact.py
    @@ -17,19 +17,22 @@ import torch.backends.cudnn as cudnn
     from utils import timer
     from utils.functions import MovingAverage, make_net
    
    -# This is required for Pytorch 1.0.1 on Windows to initialize Cuda on some driver versions.
    -# See the bug report here: https://github.com/pytorch/pytorch/issues/17108
    -torch.cuda.current_device()
    -
    -# As of March 10, 2019, Pytorch DataParallel still doesn't support JIT Script Modules
    -use_jit = torch.cuda.device_count() <= 1
    -if not use_jit:
    -    print('Multiple GPUs detected! Turning off JIT.')
    +use_jit = False
    
     ScriptModuleWrapper = torch.jit.ScriptModule if use_jit else nn.Module
     script_method_wrapper = torch.jit.script_method if use_jit else lambda fn, _rcn=None: fn
    
    
    +def decode(loc, priors):
    +    variances = [0.1, 0.2]
    +    boxes = torch.cat((priors[:, :2] + loc[:, :, :2] \* variances[0] \* priors[:, 2:], priors[:, 2:] \* torch.exp(loc[:, :, 2:] \* variances[1])), 2)
    +
    +    boxes_result1 = boxes[:, :, :2] - boxes[:, :, 2:] / 2
    +    boxes_result2 = boxes[:, :, 2:] + boxes_result1
    +    boxes_result = torch.cat((boxes_result1, boxes_result2), 2)
    +
    +    return boxes_result
    +
    
     class Concat(nn.Module):
         def __init__(self, nets, extra_params):
    @@ -476,7 +479,10 @@ class Yolact(nn.Module):
    
         def load_weights(self, path):
             """ Loads weights from a compressed save file. """
    -        state_dict = torch.load(path)
    +        if torch.cuda.is_available():
    +            state_dict = torch.load(path)
    +        else:
    +            state_dict = torch.load(path, map_location=torch.device('cpu'))
    
             # For backward compatability, remove these (the new variable is called layers)
             for key in list(state_dict.keys()):
    @@ -673,8 +679,11 @@ class Yolact(nn.Module):
                     else:
                         pred_outs['conf'] = F.softmax(pred_outs['conf'], -1)
    
    -            return self.detect(pred_outs, self)
    +            pred_outs['boxes'] = decode(pred_outs['loc'], pred_outs['priors']) # decode output boxes
    
    +            pred_outs.pop('priors') # remove unused in postprocessing layers
    +            pred_outs.pop('loc') # remove unused in postprocessing layers
    +            return pred_outs
    
    
    
    --
  3. Save and close the file.

Converting a YOLACT Model to the OpenVINO IR format

Step 1. Clone the GitHub repository and check out the commit:

  1. Clone the YOLACT repository:

    git clone https://github.com/dbolya/yolact
  2. Check out the necessary commit:

    git checkout 57b8f2d95e62e2e649b382f516ab41f949b57239
  3. Set up the environment as described in README.md.

Step 2. Download a pretrained model from the list attached in the Evaluation section of README.md document, for example yolact_base_54_800000.pth.

Step 3. Export the model to ONNX format.

  1. Apply the YOLACT_onnx_export.patch patch to the repository. Refer to the Create a Patch File instructions if you do not have it:

    git apply /path/to/patch/YOLACT_onnx_export.patch
  2. Evaluate the YOLACT model to export it to ONNX format:

python3 eval.py \
    --trained_model=/path/to/yolact_base_54_800000.pth \
    --score_threshold=0.3 \
    --top_k=10 \
    --image=/path/to/image.jpg \
    --cuda=False
  1. The script may fail, but you should get yolact.onnx file.

Step 4. Convert the model to the IR:

mo --input_model /path/to/yolact.onnx

Step 4. Embed input preprocessing into the IR:

To get performance gain by offloading to the OpenVINO application of mean/scale values and RGB->BGR conversion, use the following options of the Model Optimizer (MO):

  • If the backbone of the model is Resnet50-FPN or Resnet101-FPN, use the following MO command line:

    mo \
        --input_model /path/to/yolact.onnx \
        --reverse_input_channels \
        --mean_values "[123.68, 116.78, 103.94]" \
        --scale_values "[58.40, 57.12, 57.38]"
  • If the backbone of the model is Darknet53-FPN, use the following MO command line:

    mo \
        --input_model /path/to/yolact.onnx \
        --reverse_input_channels \
        --scale 255