I download the VCTK datasets, there are many folds for each speaker. I just copy the data of one speaker into the "train" & "test" dir (should be mkdir train & mkdir test)
And then i run like this:
python wavenet.py with 'data_dir=/vol/vssp/datasets/audio/dcase2016/vctk_data/VCTK-Corpus/wav48' small adam
(Before that: change this:
def small(desired_sample_rate):
def small():
desired_sample_rate = 4000)
And then (but i can not generate the wanted wav during test. Any wrong with my trianing process ???)
WARNING - root - Changed type of config entry "optimizer.epsilon" from NoneType to float
INFO - wavenet - Running command 'main'
WARNING - wavenet - No observers have been added to this run
INFO - wavenet - Started
Configuration (modified, added, typechanged):
batch_size = 16
data_dir = '/vol/vssp/datasets/audio/dcase2016/vctk_data/VCTK-Corpus/wav48'
data_dir_structure = 'flat'
debug = False
desired_sample_rate = 4000
dilation_depth = 8
early_stopping_patience = 20
final_l2 = 0
fragment_length = 640
fragment_stride = 400
keras_verbose = 1
learn_all_outputs = True
nb_epoch = 1000
nb_filters = 16
nb_output_bins = 256
nb_stacks = 1
predict_initial_input = ''
predict_seconds = 1
predict_use_softmax_as_input = False
random_train_batches = False
randomize_batch_order = True
res_l2 = 0
run_dir = None
sample_argmax = False
sample_temperature = 1.0
seed = 348190421
test_factor = 0.1
train_only_in_receptive_field = True
train_with_soft_target_stdev = None
use_bias = False
use_skip_connections = True
use_ulaw = True
optimizer:
decay = 0.0
epsilon = 1e-08
lr = 0.001
momentum = 0.9
nesterov = True
optimizer = 'adam'
INFO - main - Running with seed 348190421
INFO - main - Loading data...
INFO - main - Building model...
INFO - build_model - Receptive Field: 512 (128ms)
................................................................(model.summary).............................................
INFO - main - None
INFO - main - Compiling Model...
INFO - main - Starting Training...
Epoch 1/1000
9632/9632 [==============================] - 196s - loss: 4.3700 - categorical_accuracy: 0.0332 - categorical_mean_squared_error: 2048.7681 - val_loss: 4.3286 - val_categorical_accuracy: 0.0515 - val_categorical_mean_squared_error: 3280.6926
Epoch 2/1000
9632/9632 [==============================] - 198s - loss: 3.8852 - categorical_accuracy: 0.0808 - categorical_mean_squared_error: 2505.4959 - val_loss: 4.2130 - val_categorical_accuracy: 0.0652 - val_categorical_mean_squared_error: 3624.2373
Epoch 3/1000
I download the VCTK datasets, there are many folds for each speaker. I just copy the data of one speaker into the "train" & "test" dir (should be mkdir train & mkdir test)
And then i run like this: python wavenet.py with 'data_dir=/vol/vssp/datasets/audio/dcase2016/vctk_data/VCTK-Corpus/wav48' small adam
(Before that: change this:
def small(desired_sample_rate):
def small(): desired_sample_rate = 4000)
And then (but i can not generate the wanted wav during test. Any wrong with my trianing process ???) WARNING - root - Changed type of config entry "optimizer.epsilon" from NoneType to float INFO - wavenet - Running command 'main' WARNING - wavenet - No observers have been added to this run INFO - wavenet - Started Configuration (modified, added, typechanged): batch_size = 16 data_dir = '/vol/vssp/datasets/audio/dcase2016/vctk_data/VCTK-Corpus/wav48' data_dir_structure = 'flat' debug = False desired_sample_rate = 4000 dilation_depth = 8 early_stopping_patience = 20 final_l2 = 0 fragment_length = 640 fragment_stride = 400 keras_verbose = 1 learn_all_outputs = True nb_epoch = 1000 nb_filters = 16 nb_output_bins = 256 nb_stacks = 1 predict_initial_input = '' predict_seconds = 1 predict_use_softmax_as_input = False random_train_batches = False randomize_batch_order = True res_l2 = 0 run_dir = None sample_argmax = False sample_temperature = 1.0 seed = 348190421 test_factor = 0.1 train_only_in_receptive_field = True train_with_soft_target_stdev = None use_bias = False use_skip_connections = True use_ulaw = True optimizer: decay = 0.0 epsilon = 1e-08 lr = 0.001 momentum = 0.9 nesterov = True optimizer = 'adam' INFO - main - Running with seed 348190421 INFO - main - Loading data... INFO - main - Building model... INFO - build_model - Receptive Field: 512 (128ms) ................................................................(model.summary)............................................. INFO - main - None INFO - main - Compiling Model... INFO - main - Starting Training... Epoch 1/1000 9632/9632 [==============================] - 196s - loss: 4.3700 - categorical_accuracy: 0.0332 - categorical_mean_squared_error: 2048.7681 - val_loss: 4.3286 - val_categorical_accuracy: 0.0515 - val_categorical_mean_squared_error: 3280.6926 Epoch 2/1000 9632/9632 [==============================] - 198s - loss: 3.8852 - categorical_accuracy: 0.0808 - categorical_mean_squared_error: 2505.4959 - val_loss: 4.2130 - val_categorical_accuracy: 0.0652 - val_categorical_mean_squared_error: 3624.2373 Epoch 3/1000