ruohoruotsi / lstm-music-genre-classification Goto Github PK

Could you please let me know about
data[i, :, 0:13] = mfcc.T[0:self.timeseries_length, :]
data[i, :, 13:14] = spectral_center.T[0:self.timeseries_length, :]
data[i, :, 14:26] = chroma.T[0:self.timeseries_length, :]
data[i, :, 26:33] = spectral_contrast.T[0:self.timeseries_length, :]
I don't know what is 0:13?
and 13:14?
thanks

in lstm_genre_classifier_pytorch.py file
model.hidden = model.init_hidden() actually does not work.
Beacause model has no attribute 'hidden'
if you print model.hidden in every batch,you will find it's always [[0,0,...]]
model.hidden has no relation with the actual model. It's just something you define every epoch.
So it's always a stateful LSTM.

Accuracy figures for training/dev/test set in the README were generated from a run with an older version of TF (1.2) on a GPU.

Running the project freshly on a CPU, the following figures were obtained, which gives me pause. Investigate. Training accuracy is waay up and dev/validation is down significantly.

What are the (extra) implications to training on the CPU vs GPU?

420/420 [==============================] - 3s 6ms/step - loss: 0.3547 - acc: 0.8667

Validating ...
120/120 [==============================] - 0s 4ms/step
Dev loss:   1.0411598483721416
Dev accuracy:   0.6666666865348816

Testing ...
60/60 [==============================] - 0s 1ms/step
Test loss:   1.1342438459396362
Test accuracy:   0.6000000089406967```

This repo was created to facilitate learning how a simple LSTM model works on a small dataset (gtzan) solving a well understood audio/music problem (music-genre).

Add Tensorboard support to help visualize the "learning" process, debug, and optimize the code: https://www.tensorflow.org/guide/summaries_and_tensorboard

Is there a way to get the list (.csv) containing the actual value and the predicted value of the output. Here we only get the loss and accuracy?

Thank you for your contribution. I noticed that you used four audio features in the calculation of audio features. can you tell me why you used these four features?

Don't regenerate these files unless there's a force flag or they don't exist.

	data_test_input.npy
	data_test_target.npy
	data_train_input.npy
	data_train_target.npy
	data_validation_input.npy
	data_validation_target.npy

When i get all the data together, and reshuffle the dataset. I get the better accuracy in LSTM.
this is the code:

from sklearn.model_selection import train_test_split
X = torch.concatenate((train_X,test_X,dev_X),axis=0)
y = torch.concatenate((train_Y,test_Y,dev_Y),axis=0)
train_X, test_X, train_Y, test_Y = train_test_split(X, y, test_size=0.2, random_state=42)

I am training a different audio dataset on your code.
For predicting the results, I have replaced model_weights.h5 with lstm_genre_classifier_lstm.h5 (generated during lstm training) but need to know how to generate model .json file for my data?
predict_example.py

• Clean up documentation (add code to train - oneliner)
• Add graphs and or Tensorboard plots from training progress
• what else is useful to make this code approachable?

I am trying to train my dataset containing .wav files but it throws errors when I run the
lstm_genre_classifier_keras.py file
error:
self.progbar.update(self.seen, self.log_values)
AttributeError: 'ProgbarLogger' object has no attribute 'log_values'

I have changed the line 151 in GenreFeatureData.py
from:
if file.endswith(".au"):
to
if file.endswith(".wav"):