mean file from caffe model

[anguoyang]

Hi all,
I converted a caffe model to onnx successfully, and want to infer it with onnxruntime, but the problem is when I used caffe, I have to use the mean file:mean.binaryproto, but there is no document on how to use the mean file when infer with onnxruntime:
sess = nxrun.InferenceSession("./deploy.onnx")
to e.g: sess = nxrun.InferenceSession("./deploy.onnx", "mean.binaryproto")?
how can I add the mean file when infer the model? thank you

[ybrnathan]

Please refer to this article: https://stackoverflow.com/questions/41512845/what-does-caffe-do-with-the-mean-binary-file

Mean file seems like is a way to do normalization, if it is not part of your ONNX file, then you need to do it in your pre-processing step; if it's already in your converted ONNX file, then you do not need the file anymore.

[dashesy]

@anguoyang I use this to read the mean files and export in ONNX:

def read_blob(meanmodel):
    """Read blob
    :param meanmodel: path to mean blob
    """
    import caffe
    import caffe.proto.caffe_pb2 as caffe_pb2
    mean_blob = caffe_pb2.BlobProto()
    with open(meanmodel, 'rb') as fp:
        mean_blob.ParseFromString(fp.read())
    return mean_blob

class FloatSubtract(nn.Module):
    def __init__(self, dc):
        """Subtract a constant from the tensor
        dc could be anything that can be subtracted,
            it could be per-pixel mean (like the landmark example)
            or per-channel mean (like the celeb) example.
        This is a simple subtraction, fewer operations means smaller network (no need to divide by 1.0 for example)
        :param dc: the constant to subtract
        """
        super(FloatSubtract, self).__init__()
        self.dc = dc

    def extra_repr(self):
        """Extra information
        """
        return 'dc.shape={}'.format(self.dc.shape)

    def forward(self, x):
        return x.float() - self.dc.to(x.device)

mean_blob = read_blob(meanfile)
pixel_mean = np.array(mean_blob.data).astype('float32')
pixel_mean.resize(mean_blob.channels, mean_blob.height, mean_blob.width)[np.newaxis]
dc = torch.from_numpy(pixel_mean)

# Just make sure dimensions can broadcast in `dc` with the input image

dep_model = nn.Sequential(FloatSubtract(dc), model)

[anguoyang]

Hi@dashesy , thank you so much, this is what I want!