Session fail to run with error: indices = 1

TensorFlow.js version

1.5

Browser version

NA

Describe the problem or feature request

tf.saved_model with embeddings fails on model.predict in tfjs-node.

Code to reproduce the bug / link to feature request

const model = await tf.node.loadSavedModel(savedModelPath);
const input = {
        "f1": tf.tensor(['dfsdfsdf']),
        "f2": tf.tensor([1719.0]),
        "f3": tf.tensor(['ouipoio']),
        "f4": tf.tensor(['trterte']),
        "f5": tf.tensor(['sdsddsf']),
        "f6": tf.tensor(['tretert']),
        "f7": tf.tensor(['sdfsdf']),
        "f8": tf.tensor([12.0]),
        "f9": tf.tensor(['defg']),
        "f10": tf.tensor(['abcd'])
}
const output = model.predict(input);

All the features other than f8 and f2 numerical columns have embeddings. My Python code looks very similar to https://www.tensorflow.org/tutorials/structured_data/feature_columns. All my embedding feature columns have num_oov_buckets=1 if that helps in any way. After training I saved it using model.save("some_path", save_format='tf').

And here is the full stack trace,

2020-02-12 01:14:46.522724: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-02-12 01:14:46.562096: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x1072cbfc0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-02-12 01:14:46.562129: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
Loading model from <PROJECT_HOME>/saved_model/keras_saved_model
2020-02-12 01:14:46.923646: I tensorflow/cc/saved_model/reader.cc:31] Reading SavedModel from: <PROJECT_HOME>/saved_model/keras_saved_model
2020-02-12 01:14:46.943722: I tensorflow/cc/saved_model/reader.cc:54] Reading meta graph with tags { serve }
2020-02-12 01:14:47.070512: I tensorflow/cc/saved_model/loader.cc:202] Restoring SavedModel bundle.
2020-02-12 01:14:47.328407: I tensorflow/cc/saved_model/loader.cc:151] Running initialization op on SavedModel bundle at path: <PROJECT_HOME>/saved_model/keras_saved_model
2020-02-12 01:14:47.575025: I tensorflow/cc/saved_model/loader.cc:311] SavedModel load for tags { serve }; Status: success. Took 651378 microseconds.

(node:99068) UnhandledPromiseRejectionWarning: Error: Session fail to run with error: indices = 1 is not in [0, 1)
	 [[{{node sequential/dense_features/f9/f9_weights/GatherV2}}]]
    at NodeJSKernelBackend.runSavedModel (<PROJECT_HOME>/node_modules/@tensorflow/tfjs-node/dist/nodejs_kernel_backend.js:1556:43)
    at TFSavedModel.predict (<PROJECT_HOME>/node_modules/@tensorflow/tfjs-node/dist/saved_model.js:347:52)
    at predict (<PROJECT_HOME>/MY_FILE.js:28:26)
    at main (<PROJECT_HOME>/MY_FILE.js:47:5)
(node:99068) UnhandledPromiseRejectionWarning: Unhandled promise rejection. This error originated either by throwing inside of an async function without a catch block, or by rejecting a promise which was not handled with .catch(). (rejection id: 1)
(node:99068) [DEP0018] DeprecationWarning: Unhandled promise rejections are deprecated. In the future, promise rejections that are not handled will terminate the Node.js process with a non-zero exit code.

[dsmilkov]

• tf.tensor(['sdfsdf']) infers shape of [1]
• tf.tensor(['sdfsdf'], [1, 1], 'string') gives shape of [1,1].

Both are working as intended. If you want to infer shape of [1,1], you have to do tf.tensor([['sdfsdf']])

[Nithanaroy]

Ah makes sense. Thanks Daniel and Ping.

Any advice on a similar workflow https://stackoverflow.com/questions/60201536/how-to-avoid-64-bit-data-types-to-run-my-python-estimator-model-natively-on-tfjs