Inquiry About TensorflowPredict2D Support in Essentia 2.1-beta6-dev

[alice-gn-open]

Dear Essentia Community,

I am currently using Essentia for research purposes and aiming to extract mood-related features from audio data. However, I have encountered frequent errors when attempting to use the TensorflowPredict2D algorithm, and I would appreciate your guidance on the following questions:

1️⃣ Does Essentia 2.1-beta6-dev include TensorflowPredict2D?

My installed version is 2.1-beta6-dev, but when I check the available algorithms using dir(essentia.standard), TensorflowPredict2D is not listed.
Has this function been removed or is it included in a different version?

2️⃣ How can I use TensorflowPredict2D in my current setup?

If it is not included in Essentia 2.1-beta6-dev, are there alternative ways to use it?
Would upgrading to a different Essentia version or building from source be required?

3️⃣ Alternative Models and Methods

Currently, my installation supports the following models and algorithms:
['TensorflowInputFSDSINet', 'TensorflowInputMusiCNN', 'TensorflowInputTempoCNN', 'TensorflowInputVGGish', 'MusicExtractor', ...]
Given my available models, is there another way to extract mood information (e.g., using mood_sad-openl3-music-mel128-emb512-1.pb)?
Any insights or guidance on these issues would be greatly appreciated. Thank you in advance for your support!

Best regards,
Alice

[palonso]

hi @alice-gn-open,
the latest essentia-tensorflow version (essentia_tensorflow-2.1b6.dev1177) comes with TensorflowPredict2D.

Can you upgrade to that one?

[alice-gn-open]

hi @alice-gn-open, the latest essentia-tensorflow version (essentia_tensorflow-2.1b6.dev1177) comes with TensorflowPredict2D.

Can you upgrade to that one?

Dear palonso,

Thank you for your response and for suggesting that I upgrade to the latest Essentia-TensorFlow version (essentia_tensorflow-2.1b6.dev1177). I have followed the upgrade instructions, but I am still encountering issues when attempting to use the TensorflowPredict2D algorithm.

I am a psychological researcher, and I am not very familiar with coding, so I may need further assistance. Here are the specific errors I am facing:

1. When I run the following code:
   output_pool = mood_model(pool)

I get the following error:
RuntimeError: In TensorflowPredict.compute: Descriptor name 'model/Placeholder' of type Eigen::Tensor<float, 4, 1, long> not found

This suggests that the model is unable to find the input tensor Placeholder in the TensorFlow model, which may indicate an issue with the model's compatibility or the input dimensions.

2. When I try to access the predictions from the output:
   mood_predictions = output_pool["model/Softmax"]

I encounter the error:
NameError: name 'output_pool' is not defined

This is likely because the previous step failed, and thus output_pool was never properly defined.

Could you kindly provide any insights into the following:

a. How can I ensure that the input data and TensorFlow model are properly aligned?

b. Are there specific steps I should follow to confirm the TensorFlow model's compatibility with Essentia's TensorflowPredict2D algorithm?

c. Should I verify the input tensor names and dimensions within the model? If so, could you suggest the best way to do this?

Thank you very much for your continued support. I look forward to hearing from you.

Best regards,
Alice