microsoft

/

Multilingual-MiniLM-L12-H384

like 58

I have successfully converted the sentencepiece.bpe.model to JSON vocabulary and text merges. See here for more details about BPE SentencePiece tokenizer.
The tokenization works fine using tokenizers library:

const { SentencePieceBPETokenizer } = require("tokenizers");

const tokenizer = await SentencePieceBPETokenizer.fromOptions({
        vocabFile: "sentencepiece-vocab.json",
        mergesFile: "sentencepiece-merges.txt",
    });
     
    let encoder = (tokenizer) => promisify(tokenizer.encode.bind(tokenizer))
    encode = encoder(tokenizer);
    
    encoded = await encode("Hello how are you?");
    console.log("ids ", encoded.getIds());
    console.log("tokens ", encoded.getTokens());
    decoded = await tokenizer.decode(encoded.getIds(), skipSpecialTokens);
    console.log(decoded); // hello how are you?

getting the expected result

ids  [ 35377, 3641, 620, 397, 31 ]
tokens  [ '▁Hello', '▁how', '▁are', '▁you', '?' ]
Hello how are you?

When passing the generated ids to the model inference in ONNX I get an error:

const ort = require('onnxruntime-node');

// onnx sesion
session = await ort.InferenceSession.create(self._options.model.path, options);

const encoded_ids = (await tokenizer.tokenize(text)).getIds();
const model_input = ONNX.create_model_input(encoded_ids);
const output = await session.run(model_input, ['output_0']);

and getting invalid expand shape errors

2022-09-29 18:59:08.039 node[63957:683481] 2022-09-29 17:59:08.039329 [E:onnxruntime:, parallel_executor.cc:210 RunNodeAsync] Non-zero status code returned while running Expand node. Name:'Expand_21' Status Message: invalid expand shape
2022-09-29 18:59:08.040 node[63957:683429] 2022-09-29 17:59:08.039978 [E:onnxruntime:, parallel_executor.cc:75 Execute] [ONNXRuntimeError] : 2 : INVALID_ARGUMENT : Non-zero status code returned while running Expand node. Name:'Expand_21' Status Message: invalid expand shape

Model input was

model_input  {
  input_ids: l {
    dims: [ 1, 604 ],
    type: 'int64',
    data: BigInt64Array(604) [
        101n,     5n,     0n,   567n,   807n,  4504n,   91n,   397n,
       6164n,    75n,  1379n,   644n,     0n, 69084n, 1365n,    86n,
      49781n,   441n,  5674n,   515n, 58967n,    69n, 4223n,  1080n,
      34270n,     0n,   283n,   106n,   397n,  3567n, 3677n,    31n,
      15900n,   397n,   220n,  1296n,    31n,     0n,  131n,   617n,
       4504n,    91n,  1671n,   441n,    15n,     0n, 8330n,    89n,
         18n,    24n,    17n, 26865n,     0n, 37838n, 2173n,   441n,
      30481n,   397n, 12318n,  4126n,  7067n,    86n, 5153n,    53n,
        441n,     0n,   567n,  2300n,    24n,    17n, 3713n,  2366n,
        397n,    24n,   106n, 89288n,    99n,     0n,  918n, 11632n,
        758n, 79196n,   256n,  3407n,  9587n,   213n,    0n, 23388n,
         24n,     6n,  5791n,  5044n, 80096n,     0n,  567n,    24n,
         38n,   705n,   190n,    75n,
      ... 504 more items
    ],
    size: 604
  },
  attention_mask: l {
    dims: [ 1, 604 ],
    type: 'int64',
    data: BigInt64Array(604) [
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n, 1n,
      1n, 1n, 1n, 1n,
      ... 504 more items
    ],
    size: 604
  },
  token_type_ids: l {
    dims: [ 1, 604 ],
    type: 'int64',
    data: BigInt64Array(604) [
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n, 0n,
      0n, 0n, 0n, 0n,
      ... 504 more items
    ],
    size: 604
  }
}