Custom operator support

[kennu]

I'm trying to load a model trained using the MXNet DQN example. I managed to export the symbols as a JSON structure shown below. But when I try to load it into MXNet.js, I get this error:

mxnet_predict-all.cc:637: [17:17:32] mxnet_predict-all.cc:22177: Check failed: op != nullptr Operator Custom is not registered

I couldn't find any mention of custom operators in the docs. Should I add something somewhere to make them work?

{
  "nodes": [
    {
      "op": "null", 
      "name": "data", 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "conv1_weight", 
      "attr": {
        "kernel": "(8, 8)", 
        "num_filter": "32", 
        "stride": "(4, 4)"
      }, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "conv1_bias", 
      "attr": {
        "kernel": "(8, 8)", 
        "num_filter": "32", 
        "stride": "(4, 4)"
      }, 
      "inputs": []
    }, 
    {
      "op": "Convolution", 
      "name": "conv1", 
      "attr": {
        "kernel": "(8, 8)", 
        "num_filter": "32", 
        "stride": "(4, 4)"
      }, 
      "inputs": [[0, 0, 0], [1, 0, 0], [2, 0, 0]]
    }, 
    {
      "op": "Activation", 
      "name": "relu1", 
      "attr": {"act_type": "relu"}, 
      "inputs": [[3, 0, 0]]
    }, 
    {
      "op": "null", 
      "name": "conv2_weight", 
      "attr": {
        "kernel": "(4, 4)", 
        "num_filter": "64", 
        "stride": "(2, 2)"
      }, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "conv2_bias", 
      "attr": {
        "kernel": "(4, 4)", 
        "num_filter": "64", 
        "stride": "(2, 2)"
      }, 
      "inputs": []
    }, 
    {
      "op": "Convolution", 
      "name": "conv2", 
      "attr": {
        "kernel": "(4, 4)", 
        "num_filter": "64", 
        "stride": "(2, 2)"
      }, 
      "inputs": [[4, 0, 0], [5, 0, 0], [6, 0, 0]]
    }, 
    {
      "op": "Activation", 
      "name": "relu2", 
      "attr": {"act_type": "relu"}, 
      "inputs": [[7, 0, 0]]
    }, 
    {
      "op": "null", 
      "name": "conv3_weight", 
      "attr": {
        "kernel": "(3, 3)", 
        "num_filter": "64", 
        "stride": "(1, 1)"
      }, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "conv3_bias", 
      "attr": {
        "kernel": "(3, 3)", 
        "num_filter": "64", 
        "stride": "(1, 1)"
      }, 
      "inputs": []
    }, 
    {
      "op": "Convolution", 
      "name": "conv3", 
      "attr": {
        "kernel": "(3, 3)", 
        "num_filter": "64", 
        "stride": "(1, 1)"
      }, 
      "inputs": [[8, 0, 0], [9, 0, 0], [10, 0, 0]]
    }, 
    {
      "op": "Activation", 
      "name": "relu3", 
      "attr": {"act_type": "relu"}, 
      "inputs": [[11, 0, 0]]
    }, 
    {
      "op": "Flatten", 
      "name": "flatten0", 
      "inputs": [[12, 0, 0]]
    }, 
    {
      "op": "null", 
      "name": "fc4_weight", 
      "attr": {"num_hidden": "512"}, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "fc4_bias", 
      "attr": {"num_hidden": "512"}, 
      "inputs": []
    }, 
    {
      "op": "FullyConnected", 
      "name": "fc4", 
      "attr": {"num_hidden": "512"}, 
      "inputs": [[13, 0, 0], [14, 0, 0], [15, 0, 0]]
    }, 
    {
      "op": "Activation", 
      "name": "relu4", 
      "attr": {"act_type": "relu"}, 
      "inputs": [[16, 0, 0]]
    }, 
    {
      "op": "null", 
      "name": "fc5_weight", 
      "attr": {"num_hidden": "3"}, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "fc5_bias", 
      "attr": {"num_hidden": "3"}, 
      "inputs": []
    }, 
    {
      "op": "FullyConnected", 
      "name": "fc5", 
      "attr": {"num_hidden": "3"}, 
      "inputs": [[17, 0, 0], [18, 0, 0], [19, 0, 0]]
    }, 
    {
      "op": "null", 
      "name": "dqn_action", 
      "attr": {"op_type": "DQNOutput"}, 
      "inputs": []
    }, 
    {
      "op": "null", 
      "name": "dqn_reward", 
      "attr": {"op_type": "DQNOutput"}, 
      "inputs": []
    }, 
    {
      "op": "Custom", 
      "name": "dqn", 
      "attr": {"op_type": "DQNOutput"}, 
      "inputs": [[20, 0, 0], [21, 0, 0], [22, 0, 0]]
    }
  ], 
  "arg_nodes": [
    0, 
    1, 
    2, 
    5, 
    6, 
    9, 
    10, 
    14, 
    15, 
    18, 
    19, 
    21, 
    22
  ], 
  "node_row_ptr": [
    0, 
    1, 
    2, 
    3, 
    4, 
    5, 
    6, 
    7, 
    8, 
    9, 
    10, 
    11, 
    12, 
    13, 
    14, 
    15, 
    16, 
    17, 
    18, 
    19, 
    20, 
    21, 
    22, 
    23, 
    24
  ], 
  "heads": [[23, 0, 0]], 
  "attrs": {"mxnet_version": ["int", 905]}
}