Sequential dataset errors

[Yves Sibony]

Hello,

I am training a LSTM on sequential data where each sample is a sequence of 250 numbers. Each one of these numbers has a label, so my dataset file looks like this:

x:samples,y:labels

./LSTM_sample_0.csv,./LSTM_label_0.csv

./LSTM_sample_1.csv,./LSTM_label_1.csv

./LSTM_sample_2.csv,./LSTM_label_2.csv

./LSTM_sample_3.csv,./LSTM_label_3.csv

...

each sample/label file has no header, 250 rows and 1 column, each cell containing one number (between 0 and 20 for the samples, then converted to a one-hot vector in the network; between 0 and 8 for the labels). 

I've been trying different shapes for the input and loss (SigmoidCrossEntropy) layers but I always get a "Errors in dataset" error.

Can you help me please?

Do not hesitate to ask if you need more information.

[Yoshiyuki Kobayashi]

Could you tell me the exact error message?

To take variable x with 250 rows and 1 cols, set the Size property of Input layer to "250,1".
And use Embed layer instead of creating one-hot vector.

To compute categorical cross entropy loss between network output (250,9) and labels  (250,1), use SoftmaxCrossEntropy layer instead of SigmoidCrossEntropy layer.

[Yves Sibony]

The exact error message is "Errors in dataset /n Linked dataset might have been deleted."

I have tried your suggestions, but the error message persists

[Yoshiyuki Kobayashi]

Are you using cloud version?
The error message occurs when the specified dataset in the DATASET tab is deleted.
Could you check that the dataset is linked correctly in the DATASET tab?

[Yves Sibony]

Yes, I'm using the cloud version.

I did check that my datasets (training and validation) are properly linked in the dataset tab.