[JAVA] If_0_then_branch__Inline_0__/decoder/rnn/LSTM' Status Message: Input X must have 3 dimensions only. Actual:{1,1,1,128,4}

[Furtherking]

❓ Questions and Help

Error code - ORT_INVALID_ARGUMENT - message: Non-zero status code returned while running If node. Name:'If_0' Status Message: Non-zero status code returned while running If node. Name:'If_0_then_branch__Inline_0__/decoder/If_1' Status Message: Non-zero status code returned while running LSTM node. Name:'If_0_then_branch__Inline_0__/decoder/rnn/LSTM' Status Message: Input X must have 3 dimensions only. Actual:{1,1,1,128,4}

The error occurred when using the code in example/java.I have already modified part of the code in the SlieroVadOnnxModel by referring to wavjava to make the input compatible with V5, but I still encountered this error. Has anyone else encountered this issue, and what should I do?

[snakers4]

A possible step-by-step solution:

• Launch old JAVA example;
• Print tensor dimensions on each step of the ONNX wrapper;
• Launch v5 python ONNX example and print the same dimensions;
• Amend the new JAVA example accordingly;

From the error it may seem that an extra dimension was passed at some point somewhere.

Also comparing the probabilities is useful. Also the current JAVA example was not updated for v5.

[Furtherking]

Thank you so much for your email with the step-by-step solution. It's extremely helpful and much appreciated! I'll follow your advice.Thanks again for your guidance!

…

------------------ 原始邮件 ------------------ 发件人: "Alexander ***@***.***>; 发送时间: 2024年10月11日(星期五) 晚上8:24 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [snakers4/silero-vad] [JAVA] If_0_then_branch__Inline_0__/decoder/rnn/LSTM' Status Message: Input X must have 3 dimensions only. Actual:{1,1,1,128,4} (Issue #550) A possible step-by-step solution: Launch old JAVA example; Print tensor dimensions on each step of the ONNX wrapper; Launch v5 python ONNX example and print the same dimensions; Amend the new JAVA example accordingly; From the error it may seem that an extra dimension was passed at some point somewhere. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Furtherking]

I tried it out, but strangely, despite my current code having the same input dimensions as the python/javawav example and maintaining consistent dimensions, the javawav runs normally, while the javaexample for reading microphone audio encounters errors. Could there be a problem with the original part that reads the microphone audio?

[snakers4]

All other factors being equal if python and java output similar probabilities for each chunk in ideal case, then it is data preprocessing most likely

In this case look for the data format differences (sample rate, channels, dimensions, chunk size)

[Furtherking]

I have resolved the bug in the model.call() method by modifying the multiplication symbol (*) to a division symbol (/) in the array initialization: byte[] data = new byte[WINDOW_SIZE_SAMPLES / 2];.
But there still some problems,It seems that the change in input has led to the audio input captured from the microphone no longer providing accurate results. The SpeechProb values are consistently very low, even reaching levels as low as 0.0001, which makes it impossible to determine whether there is speech activity based on SpeechProb. Is this due to modifications in the v5 model?