def train_model(model_class,input_dimensions, hidden_size, batch_size, truncated_len, num_epochs, model_name,
print_period=50, save_period=50, log_period=50, n_files_per_epoch=5, sparsify_epochs = [], sparsity_level=1, wav_fnames=None, align_fnames=None):
if model_name not in os.listdir(DIRS['MODELS']):
os.mkdir(DIRS['MODELS']+model_name)
tf.reset_default_graph()
model = model_class(input_dimensions, hidden_size)
init_variables = tf.global_variables_initializer()
saver = tf.train.Saver()
if wav_fnames is None:
wav_fnames = Path(DIRS['RAW_DATA']).rglob("*.wav")
epochs_per_files_last = 0
# Initialize the losses
train_losses = []
validation_losses = []
with tf.Session() as sess:
sess.run(init_variables)
# Perform all the iterations
for epoch in tqdm_notebook(range(num_epochs)):
if epochs_per_files_last==0:
X = sl.load_data(wav_fnames, n_files_per_epoch)
if align_fnames is not None:
toh = sl.load_text_oh(align_fnames, n_files_per_epoch)
total_series_length = int(X.shape[1])
epochs_per_files_last = total_series_length//batch_size//truncated_len
epochs_per_files_last-=1
if epoch in sparsify_epochs:
k = model.calc_sparsity_level(epoch, sparsify_epochs, sparsity_level)
model.sparsify(k, sess)
X_train, Y_train, X_test, Y_test = sl.get_train_test(X, batch_size, truncated_len, sess, text_oh=toh)
train_loss, validation_loss = model.train(X_train, Y_train, X_test, Y_test, sess)
# Log the losses
train_losses.append(train_loss)
validation_losses.append(validation_loss)
msg = f'Iteration: {epoch}, train loss: {train_loss:.4f}, val loss: {validation_loss:.4f}'
# Display an update every 50 iterations
if epoch % print_period == 0 and epoch!=0:
print(msg)
if epoch % print_period == 0 and epoch!=0:
sl.plot_losses(train_losses, validation_losses, title=msg)
plt.show()
if epoch % save_period == 0:
saver.save(sess, DIRS['MODELS']+model_name+'/checkpoint',global_step=epoch,write_meta_graph=True)
sl.plot_losses(train_losses, validation_losses, title=msg)
plt.show()
saver.save(sess, DIRS['MODELS']+model_name+'/final')
return train_losses, validation_losses, model
def train_model(input_dimensions, hidden_size, batch_size, truncated_len, num_epochs, model_name,
print_period=50, save_period=50, log_period=50):
if model_name not in os.listdir(DIRS['MODELS']):
os.mkdir(DIRS['MODELS']+model_name)
tf.reset_default_graph()
model = WaveGRU(input_dimensions, hidden_size)
init_variables = tf.global_variables_initializer()
saver = tf.train.Saver()
wav_fnames = Path(DIRS['RAW_DATA']).rglob("*.wav")
epochs_per_files_last = 0
# Initialize the losses
train_losses = []
validation_losses = []
with tf.Session() as sess:
sess.run(init_variables)
# Perform all the iterations
for epoch in tqdm_notebook(range(num_epochs)):
if epochs_per_files_last==0:
X = sl.load_data(wav_fnames, 5)
total_series_length = int(X.shape[1])
epochs_per_files_last = total_series_length//batch_size//truncated_len
epochs_per_files_last-=1
X_train, Y_train, X_test, Y_test = sl.get_train_test(X, batch_size, truncated_len, sess)
train_loss, validation_loss = model.train(X_train, Y_train, X_test, Y_test, sess)
# Log the losses
train_losses.append(train_loss)
validation_losses.append(validation_loss)
# Display an update every 50 iterations
if epoch % print_period == 0 and epoch!=0:
print(f'Iteration: {epoch}, train loss: {train_loss:.4f}, val loss: {validation_loss:.4f}')
if epoch % print_period == 0 and epoch!=0:
sl.plot_losses(train_losses, validation_losses,
title=f'Iteration: {epoch}, train loss: {train_loss:.4f}, val loss: {validation_loss:.4f}')
plt.show()
if epoch % save_period == 0:
saver.save(sess, DIRS['MODELS']+model_name+'/checkpoint',global_step=epoch,write_meta_graph=True)
sl.plot_losses(train_losses, validation_losses,
title='Iteration: %d, train loss: %.4f, test loss: %.4f' % (epoch, train_loss, validation_loss))
plt.show()
saver.save(sess, DIRS['MODELS']+model_name+'/final')
return train_losses, validation_losses, model
plt.tight_layout()
plt.show()
for idx, m in enumerate([Ru, Rr, Re]):
plt.subplot(1,3,idx+1)
plt.imshow(m)
plt.title(['Ru','Rr','Re'][idx])
plt.tight_layout()
plt.show()
for idx, m in enumerate([Iu, Ir, Ie]):
plt.subplot(1,3,idx+1)
sns.heatmap(m)
plt.title(['Iu','Ir','Ie'][idx])
plt.tight_layout()
plt.show()audio = sl.load_audio_not_one_hot(DIRS['RAW_DATA']+'cv_corpus_v1/cv-other-train/sample-052026.wav')X = sl.load_data([DIRS['RAW_DATA']+'cv_corpus_v1/cv-other-train/sample-052026.wav'])with tf.Session() as sess:
init_variables = tf.global_variables_initializer()
sess.run(init_variables)
O1, O2, O3, O4 = sess.run([gru.O1, gru.O2, gru.O3, gru.O4])
Iu, Ir, Ie = sess.run([gru.Iu, gru.Ir, gru.Ie])
Ru, Rr, Re = sess.run([gru.Ru, gru.Rr, gru.Re])
audio_eval = sess.run(audio)
X_eval = sess.run(X)
X_train, Y_train, X_test, Y_test = sl.get_train_test(X, 10, 5000)
Y_train_audio = ((Y_train*128+128)[:,:,0])*256+(Y_train*128+128)[:,:,1]
Y_train_audio_eval = sess.run(Y_train_audio)
os.listdir(DIRS['RAW_DATA']+'cv_corpus_v1/cv-other-train')plt.plot(audio_eval, color='blue', label='audio')
for i in Y_train_audio_eval[:1]:
plt.plot(i)
plt.legend()
plt.show()with tf.Session() as sess:
plt.plot(train_losses, label='Train')
plt.plot(validation_losses, label='Validation')
plt.title(model_name)
plt.show()
# # Restoring model
model_name = f'Sparse_Develop'sorted(os.listdir(DIRS['MODELS']+model_name))tf.reset_default_graph()
#saver = tf.train.import_meta_graph(DIRS['MODELS']+model_name+'/final.meta')
saver = tf.train.import_meta_graph(DIRS['MODELS']+model_name+'/checkpoint-900.meta')
with tf.Session() as sess:
saver.restore(sess,tf.train.latest_checkpoint(DIRS['MODELS']+model_name))
restored_variables = {x.name:x.eval(session=sess) for x in tf.global_variables()[:13]}
tf.reset_default_graph()
gru = WaveGRU(input_dimensions, hidden_size, variables_values_dict=restored_variables)X = sl.load_data(wav_fnames, 3)
# # Sound generation
with tf.Session() as sess:
init_variables = tf.global_variables_initializer()
sess.run(init_variables)
gen_to_wav = gru.generate_sound(num_pieces=1, n_seconds=2, session=sess, sample_rate=M_PARAMS['SAMPLE_RATE'])with tf.Session() as sess:
sess.run(init_variables)
#plt.plot(audio.eval(session=sess), label='real')
plt.plot(gen_to_wav[0].eval(session=sess), label='generated')
plt.plot(np.int32([np.sin(x/1000)*16000+32256 for x in range(gen_to_wav.shape[1])]))with tf.Session() as sess:
init_variables = tf.global_variables_initializer()
sess.run(init_variables)
O1, O2, O3, O4 = sess.run([gru.O1, gru.O2, gru.O3, gru.O4])
Iu, Ir, Ie = sess.run([gru.Iu, gru.Ir, gru.Ie])
Ru, Rr, Re = sess.run([gru.Ru, gru.Rr, gru.Re])for idx, m in enumerate([O1, O2, O3, O4]):
plt.subplot(1,4,idx+1)
# In[13]:
fname = DIRS['RAW_DATA']+'/rus/voxforge_ru/0/00/78d77cdb75be'
# In[14]:
os.path.isfile(fname+'.wav')
# In[120]:
audio_data = sl.load_data([fname+'.wav']).eval(session=sess)
audio_data
# In[127]:
audio_data.shape
audio_data = np.concatenate([audio_data,np.ones((1,30-1-int(audio_data.shape[1]%30)+30,2))],1)
audio_data = np.concatenate([-np.ones((1,30*2,2)),audio_data],1)n_batches=30for i in range(n_batches-1):
audio_data = np.concatenate([audio_data[:,1:,:],audio_data[:,:-1,-2:]],2)audio_data.shapeaudio_data_resh = audio_data.reshape((1,30,-1,2))audio_data_resh.shape
# # New model with new sparse
# In[163]:
# In[8]: tf.reset_default_graph() gru = WaveGRU(input_dimensions, hidden_size) # In[9]: init_variables = tf.global_variables_initializer() # In[10]: saver = tf.train.Saver() # In[11]: X = sl.load_data(wav_fnames, 3) # In[12]: batch_size = 10 truncated_len = M_PARAMS['SAMPLE_RATE'] // 128 total_series_length = int(X.shape[1]) num_epochs = 50 #total_series_length//batch_size//truncated_len print(batch_size, truncated_len, total_series_length, num_epochs) # In[13]: n_early_stopping = 50 # In[13]:
char_to_int = {k: idx for idx,k in enumerate('\0 абвгдежзийклмнопрстуфхцчшщъыьэюяё')}
# In[70]:
fname_wav, fname_txt, fname_align = text_ds.loc[0,['wav','txt','align']]
fname_wav
# In[71]:
data = sl.load_data([fname_wav])
# In[72]:
oh = sl.load_text_oh([fname_align])
# In[73]:
sl.get_train_test(data, batch_size=10, truncated_len=100, text_oh=oh)
# In[65]: