Uncaught (in promise) Error: Argument tensors passed to stack must be a Tensor[] or TensorLike[]

TensorFlow.js version

2.0.0

Browser version

Microsoft Edge Version 83.0.478.58 (Official build) (64-bit)

Describe the problem or feature request

Uncaught (in promise) Error: Argument tensors passed to stack must be a Tensor[] or TensorLike[] Error occurs every time the model is fit with tensor arrays or by zipping as dataset.

Error Stack : tensor_util_env.js:125 Uncaught (in promise) Error: Argument tensors passed to stack must be a Tensor[] or TensorLike[] at jg (tensor_util_env.js:125) at stack_ (array_ops.js:134) at stack (operation.js:45) at Object.x (array_ops.js:186) at tape.js:167 at engine.js:425 at t.e.scopedRun (engine.js:436) at t.e.tidy (engine.js:423) at engine.js:1005 at s (tape.js:167)

Code to reproduce the bug / link to feature request

<script>
    async function trainmodel() {
        const model = tf.sequential({
            layers:[tf.layers.lstm({units:100,activation:'relu',returnSequences:true,inputShape:[1,1]}),
                    tf.layers.lstm({units:10,activation:'relu',returnSequences:true}),
                    tf.layers.dense({units:1})
            ]
        });

        const X = tf.tensor3d([0.1,0.4,0.7,0.5],[4,1,1]);
        const y = tf.tensor3d([0.05,0.5,0.95,0.8],[4,1,1]);

        function onBatchEnd(batch, logs) {
            console.log('Loss', logs.acc);
        }

        model.compile({
            optimizer: tf.train.sgd(0.01),
            loss: 'meanSquaredError',
            metrics: ['accuracy']
        });

        model.fit(X, y, {
            epochs: 3,
            callbacks: {onBatchEnd}
        }).then(info => {
            console.log('Final accuracy', info.history.acc);
        });
       
    }
    trainmodel();
</script>

[caisq]

More info: convertToTensorArray(), which is a part of the gradient tape mechanism, expects its argument to be an array of Tensor or TensorLike. But for some reason, when input shape of the LSTM is [1, 1], i.e., only one temporal step, convertToTensorArray() receives a single Tensor as argument. If the argument is an Array<Tensor> of length 1, then everything works.

[Artod]

Ran into the same issue. It seems that the RNN layer expects an input with at least 2 time steps when training. If that is the case, I think the documentation should mention that.

Why would we need to pass into the RNN layer a sequence with just one element? From my understanding, when using stateful RNN, that might make sense (correct me if I’m wrong).