[DEPRECATED] Convert ONNX* DLRM to the Intermediate Representation


These instructions are currently deprecated. Since OpenVINO™ 2020.4 version, no specific steps are needed to convert ONNX* DLRM models. For general instructions on converting ONNX models, please refer to Converting a ONNX* Model topic.

These instructions are applicable only to the DLRM converted to the ONNX* file format from the facebookresearch/dlrm model.

Step 1. Save trained Pytorch* model to ONNX* format or download pretrained ONNX* from MLCommons/inference/recommendation/dlrm repository. If you train the model using the script provided in model repository, just add the --save-onnx flag to the command line parameters and you’ll get the dlrm_s_pytorch.onnx file containing the model serialized in ONNX* format.

Step 2. To generate the Intermediate Representation (IR) of the model, change your current working directory to the Model Optimizer installation directory and run the Model Optimizer with the following parameters:

python3 mo.py --input_model dlrm_s_pytorch.onnx
mo --input_model dlrm_s_pytorch.onnx

Note that Pytorch model uses operation torch.nn.EmbeddingBag. This operation converts to onnx as custom ATen layer and not directly supported by OpenVINO*, but it is possible to convert this operation to:

  • Gather if each “bag” consists of exactly one index. In this case offsets input becomes obsolete and not needed. They will be removed during conversion.

  • ExperimentalSparseWeightedSum if “bags” contain not just one index. In this case Model Optimizer will print warning that pre-process of offsets is needed, because ExperimentalSparseWeightedSum and torch.nn.EmbeddingBag have different format of inputs. For example if you have indices input of shape [indices_shape] and offsets input of shape [num_bags] you need to get offsets of shape [indices_shape, 2]. To do that you may use the following code snippet:

    import numpy as np
    new_offsets = np.zeros((indices.shape[-1], 2), dtype=np.int32)
    new_offsets[:, 1] = np.arange(indices.shape[-1])
    bag_index = 0
    for i in range(offsets.shape[-1] - 1):
        new_offsets[offsets[i]:offsets[i + 1], 0] = bag_index
        bag_index += 1
    new_offsets[offsets[-1]:, 0] = bag_index

    If you have more than one torch.nn.EmbeddingBag operation you’ll need to do that for every offset input. If your offsets have same shape they will be merged into one input of shape [num_embedding_bags, indices_shape, 2].