Network Notes



In [1]:

    
# good 160 unit *3 network
network_model = "func-28-1-3"
iter1 = 27830
iter2 = 23550
codec = 3200
model_type = 'functional'

output_fn_postfix1 = "output_" + str(0) + "_" + str(iter1)
output_fn_postfix1_mid = "output_" + str(1) + "_" + str(iter1)

output_fn_postfix2 = "output_" + str(0) + "_" + str(iter2)
output_fn_postfix2_mid = "output_" + str(1) + "_" + str(iter2)



from subprocess import check_output, call

import os
import json

home = os.environ.get('HOME')
if codec==3200:
    codec_sub = '-3200'
else:
    codec_sub = ''

os.chdir(home + "/store/c2gen/notebooks")


cb = home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix1)
call(["bash", home + "/store/c2gen/c2towav.sh", cb ])
cb = home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix1_mid)
call(["bash", home + "/store/c2gen/c2towav.sh", cb ])
cb = home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(iter2)
call(["bash", home + "/store/c2gen/c2towav.sh", cb ])

import network_data as nd
from IPython.display import display, Markdown, Audio, Image
nh = home + "/store/c2gen/out/"+network_model+"/log"
    
if os.path.isfile(nh):
  with open(nh) as f:
      res = f.readlines()
      print(str.join("",res[-12:]))

  try:
    res = check_output(["tail", '-n 1', home + "/store/c2gen/nohup/"+network_model+".out"])
  except:
    try:
      res = check_output(["tail", '-n 1', home + "/store/c2gen/nohup-"+network_model+".out"])
    except:
      res = ''
  if res != '':  
    res = str(res).replace('\\n', "\n")
    res = str(res).replace('\\r', "\n")
    res = str(res).replace('\\x08', "")
    res = res.split("\n")[-1]
    print(res)









    



generating encoded output 
Training Iteration 27837 
using full set of frames 
generating encoded output 
Training Iteration 27838 
using full set of frames 
generating encoded output 
Training Iteration 27839 
using full set of frames 
generating encoded output 
Interrupt signal caught. Closing gracefully. Iteration: 27839 
saving config 

'



In [2]:

    
with open(home + "/store/c2gen/out/"+network_model+"/config.json") as f:
    config = j = json.load(f)
    for k,v in j.items():
        print(k,": ",v)









    



start_iteration :  27839
num_iterations :  30000
fit_batch_size :  5
learn_next_step :  False
gen_every_nth :  10
generate_num_outputs :  2
save_model_every_nth :  10
framelen :  13
frame_seq_len :  200
overlap_sequence :  1
seed_seq_len :  200
seed_start_index :  60
seq_step :  198
test_data_fn :  test/critiquepracticalreason_00_kant_64kb.c2cb-3200
frame_prop_orig_scale :  [1, 127, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31]
frame_prop_loss_scale :  [1, 127, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31]
stateful :  False
shuffle :  False
limit_frames :  0
optimizer :  {'name': 'RMSprop', 'params': {}}
generate_len :  200
model_filename :  out/func-28-1-3/model-27839.h5
frame_len_ms :  20



In [3]:

    
try:
  with open(home + "/store/c2gen/out/" + network_model + '/notes.md', 'r') as fh:
    display(Markdown(fh.read()))
    
except:
    print("no markdown file found")









    




func-28-1-3 Notes
2018-01-29 15:11:06
Run with arguments test/critiquepracticalreason_00_kant_64kb.c2cb-3200 --load-weights=out/func-28-1-2/weights-15000.h5
Description
Added in two new LSTMs and reloaded weights from last run.
Disabled training on the 3 initial LSTMs and on encoder and decoder
Going well!
Restarted at 3000, having reenabled trainable flag for 
the initial LSTMs, allowing all LSTMs to refine themselves.
Restarted at 5450. Added 2 more LSTMs on the end, disabling the training 
of existing early and mid LSTMs.
Restarted at 9290 with a final wide LSTM on the end after disabling the training of all others
Didn't work due to a shape mismatch. Trying two new lstms instead of one large
Restarted at 12940 setting all LSTMs to trainable
At 18130 added two more LSTMs, disabling all others from training.
At 20290 enabled training on all the LSTMs.



In [4]:

    
if model_type != 'functional':
  nd.model_config(network_model)



In [5]:

    
nd.plot_training_loss(network_model, 'mean abs', columns=['generator_TD_Dense_0_loss', 'model_2_loss_2', 'model_2_loss_1', 'loss'], weights=[20,1,1,10], start_index=0, y_max=4)









    



Columns: ['generator_TD_Dense_0_loss', 'model_2_loss_2', 'model_2_loss_1', 'loss']



In [6]:

    
with open(home + "/store/c2gen/out/"+network_model+"/training.log") as f:
    rl = f.readlines()
    print(str.join("",rl[0:1]))
    print(str.join("",rl[-10:]))
    

with open(home + "/store/c2gen/out/"+network_model+"/gen_counter") as f:
    latest_iter = (int(f.readlines()[0]))
    print(latest_iter)









    



epoch,generator_TD_Dense_0_loss,loss,model_1_loss,model_2_loss_1,model_2_loss_2

0,0.0228834829662,0.0228834829662,0.403349318315,1.48986675234,1.28469138259
0,0.0228853921744,0.0228853921744,0.403349318315,1.4896422835,1.28469138189
0,0.023087434378,0.023087434378,0.403349318315,1.49151063209,1.28469138304
0,0.0229652069392,0.0229652069392,0.403349318315,1.49016173996,1.28469138127
0,0.023130009408,0.023130009408,0.403349318315,1.49165277819,1.28469138262
0,0.0230927647217,0.0230927647217,0.403349318315,1.49181504394,1.28469138241
0,0.0229178732956,0.0229178732956,0.403349318315,1.48996953423,1.28469138271
0,0.0230939418238,0.0230939418238,0.403349318315,1.4909953994,1.28469138184
0,0.0229480034374,0.0229480034374,0.403349318315,1.48991555011,1.28469138211
0,0.0230319671685,0.0230319671685,0.403349318315,1.49201913814,1.28469138284

27830



In [7]:

    
try:
  print("Latest Iteration:", latest_iter)
  
  output_fn_postfix = "output_" + str(0) + "_" + str(latest_iter)
  output_fn_postfix_mid = "output_" + str(1) + "_" + str(latest_iter)
  
  chart_postfix = output_fn_postfix

  nd.plot_codec_params(network_model,chart_postfix, scale_up='orig')
  nd.plot_codec_params(network_model,chart_postfix, scale_up='full', test_data_fn=config['test_data_fn'], test_seed_start=config['seed_start_index'])
  nd.plot_spec_params(network_model,chart_postfix, params='Voicing')
  nd.plot_spec_params(network_model,chart_postfix, params='Wo')
  nd.plot_spec_params(network_model,chart_postfix, params='E')
  nd.plot_spec_params(network_model,chart_postfix, params='LSPs', lsp_param=[0], test_data_fn=config['test_data_fn'], test_seed_start=config['seed_start_index'])
except FileNotFoundError:
    print("File not found")









    



Latest Iteration: 27830
3200 rate codec







    












    



3200 rate codec

Calculated diff for each parameter



In [8]:

    
cb = home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix)
call(["bash", home + "/store/c2gen/c2towav.sh", cb ])

try:
  nd.plot_audio_waveform(network_model, output_fn_postfix)
  display(Audio(filename=home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix)+codec_sub+".wav"))
except FileNotFoundError:
  print("file not found")
print()



In [9]:

    
cb = home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix_mid)
call(["bash", home + "/store/c2gen/c2towav.sh", cb ])

try:
  nd.plot_audio_waveform(network_model, output_fn_postfix_mid)
  display(Audio(filename=home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix_mid)+codec_sub+".wav"))
except FileNotFoundError:
  print("file not found")
print()

Early Iterations



In [10]:

    
print(output_fn_postfix1)
nd.plot_codec_params(network_model,output_fn_postfix1, scale_up='orig')
nd.plot_spec_params(network_model,output_fn_postfix1, params='Voicing')
nd.plot_spec_params(network_model,output_fn_postfix1, params='Wo')
nd.plot_spec_params(network_model,output_fn_postfix1, params='E')
nd.plot_spec_params(network_model,output_fn_postfix1, params='LSPs')









    



output_0_27830
3200 rate codec







    












    












    












    












    



---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-10-8c8f3d539460> in <module>()
      4 nd.plot_spec_params(network_model,output_fn_postfix1, params='Wo')
      5 nd.plot_spec_params(network_model,output_fn_postfix1, params='E')
----> 6 nd.plot_spec_params(network_model,output_fn_postfix1, params='LSPs')

/media/phil/Data/projects/babble-rnn/notebooks/network_data.py in plot_spec_params(network_tag, iteration, params, loc, lsp_param, test_data_fn, test_seed_start)
    156     elif params == 'LSPs':
    157       if len(lsp_param) == 2:
--> 158         data = data[:, 3+lsp_param[0]:3+lsp_param[1] ]
    159       else:
    160         data = data[:, 3+lsp_param[0] ]

TypeError: unsupported operand type(s) for +: 'int' and 'NoneType'



In [ ]:

    
nd.plot_audio_waveform(network_model, output_fn_postfix1)

Listen to the WAV file



In [ ]:

    
display(Audio(filename=home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix1)+codec_sub+".wav"))



In [ ]:

    
nd.plot_audio_waveform(network_model, output_fn_postfix1_mid)



In [ ]:

    
display(Audio(filename=home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix1_mid)+codec_sub+".wav"))

Later Iterations



In [ ]:

    
nd.plot_codec_params(network_model,output_fn_postfix2, scale_up=True)
nd.plot_spec_params(network_model,output_fn_postfix2, params='Voicing')
nd.plot_spec_params(network_model,output_fn_postfix2, params='Wo')
nd.plot_spec_params(network_model,output_fn_postfix2, params='E')
nd.plot_spec_params(network_model,output_fn_postfix2, params='LSPs')



In [ ]:

    
nd.plot_audio_waveform(network_model, output_fn_postfix2)

Listen to the WAV file



In [ ]:

    
display(Audio(filename=home + "/store/c2gen/out/"+network_model+"/out-c2cb-"+str(output_fn_postfix2)+codec_sub+".wav"))