keums / melodyextraction_mcdnn Goto Github PK

I'm using keras 2.2.4, and the code can't run. I found the model weights file in this repository has this architecture: (take _frames_11_note_res_1_drop_out_0.2_optimizer_rms_activation_relu_output_activation_sigmoid_Melody.hdf5 for example)

<HDF5 group "/layer_0" (2 members)>
('param_0', <HDF5 dataset "param_0": shape (2816, 512), type "<f4">)
('param_1', <HDF5 dataset "param_1": shape (512,), type "<f4">)
<HDF5 group "/layer_1" (0 members)>
<HDF5 group "/layer_10" (0 members)>
<HDF5 group "/layer_2" (0 members)>
<HDF5 group "/layer_3" (2 members)>
('param_0', <HDF5 dataset "param_0": shape (512, 512), type "<f4">)
('param_1', <HDF5 dataset "param_1": shape (512,), type "<f4">)
<HDF5 group "/layer_4" (0 members)>
<HDF5 group "/layer_5" (0 members)>
<HDF5 group "/layer_6" (2 members)>
('param_0', <HDF5 dataset "param_0": shape (512, 256), type "<f4">)
('param_1', <HDF5 dataset "param_1": shape (256,), type "<f4">)
<HDF5 group "/layer_7" (0 members)>
<HDF5 group "/layer_8" (0 members)>
<HDF5 group "/layer_9" (2 members)>
('param_0', <HDF5 dataset "param_0": shape (256, 41), type "<f4">)
('param_1', <HDF5 dataset "param_1": shape (41,), type "<f4">)

but the model created by the code has the architecture:

<HDF5 group "/activation_1" (0 members)>
<HDF5 group "/activation_2" (0 members)>
<HDF5 group "/activation_3" (0 members)>
<HDF5 group "/activation_4" (0 members)>
<HDF5 group "/dense_1" (1 members)>
('dense_1', <HDF5 group "/dense_1/dense_1" (2 members)>)
<HDF5 group "/dense_2" (1 members)>
('dense_2', <HDF5 group "/dense_2/dense_2" (2 members)>)
<HDF5 group "/dense_3" (1 members)>
('dense_3', <HDF5 group "/dense_3/dense_3" (2 members)>)
<HDF5 group "/dense_4" (1 members)>
('dense_4', <HDF5 group "/dense_4/dense_4" (2 members)>)
<HDF5 group "/dropout_1" (0 members)>
<HDF5 group "/dropout_2" (0 members)>
<HDF5 group "/dropout_3" (0 members)>

when running the code, keras would raise:
ValueError: You are trying to load a weight file containing 0 layers into a model with 4 layers.

I did some modify so that the code could wrong.

• in the MelodyExtraction_MCDNN.py, I modify the line where loads the weights file:
  model.load_weights(weight_file，by_name=True)

• in the main.py, there is a bug
  TypeError: slice indices must be integers or None or have an __index__ method ,so I modify something as follow:
  `for i in range(y_predict_3rd.shape[0]):
  for j in range(y_predict_1st.shape[1]-1):
  y_predict_tmp_1_3[i,int(jratio_res_1_3):int(jratio_res_1_3+ratio_res_1_3)] = y_predict_1st[i,j]
  y_predict_tmp_1_3[i,-1] = y_predict_1st[i,-1]

  for i in range(y_predict_3rd.shape[0]):
  for j in range(y_predict_2nd.shape[1]-1):
  y_predict_tmp_2_3[i,int(jratio_res_2_3):int(jratio_res_2_3+ratio_res_2_3)] = y_predict_2nd[i,j]
  y_predict_tmp_2_3[i,-1] = y_predict_2nd[i,-1] `

After this ,the code works well. But the result txtfile looks strange:
0.000000 0.000000 0.010000 0.000000 0.020000 578.909109 0.030000 578.909109 0.040000 578.909109 0.050000 578.909109 0.060000 578.909109 0.070000 578.909109 0.080000 578.909109 0.090000 578.909109 0.100000 578.909109 0.110000 578.909109 0.120000 578.909109 0.130000 578.909109 0.140000 578.909109 0.150000 578.909109 0.160000 578.909109 0.170000 578.909109 0.180000 578.909109 0.190000 578.909109 0.200000 578.909109 0.210000 578.909109 0.220000 578.909109 0.230000 578.909109 0.240000 578.909109 0.250000 578.909109 0.260000 578.909109 0.270000 578.909109 0.280000 578.909109 0.290000 578.909109 0.300000 578.909109 0.310000 578.909109 0.320000 578.909109 0.330000 578.909109 0.340000 578.909109 0.350000 578.909109 0.360000 578.909109 0.370000 578.909109 0.380000 578.909109 0.390000 578.909109 0.400000 578.909109 0.410000 578.909109 0.420000 578.909109 0.430000 578.909109 0.440000 578.909109 0.450000 578.909109 0.460000 578.909109 0.470000 578.909109 0.480000 578.909109 0.490000 578.909109 0.500000 578.909109
the pitch only have two value: 0 or 578.909109
I don't know where goes wrong! Please help!
Thank you very much!

When running load_train_data_RWC_dnn.py, I found that:
line 37 features = np.load(PATH + 'features/'+file_name.replace('.TXT','.npy'))
there's no "features" directory in "Data"
How should I do to get "features"? And what's the difference between "features" and "pitch"?
Thanks!!

Hi @keums

Is there the code how to train the model ?. Thanks

in librosa changelog:

#627_ The following functions and features have been removed:

• real= parameter in cqt
• core.logamplitude (replaced by amplitude_to_db) <--- used in your code
• beat.estimate_tempo (replaced by beat.tempo)
• n_fft= parameter to feature.rmse
• ref_power= parameter to power_to_db

I personally recommend GPLv3, but I would rather there just be a license so the air is clear about ownership of any audio made with this software and whether it can be borrowed from in contributing to projects like Audacity or Blender. Please & Thanx~!