setup_logging(logging.getLogger())
data_source = []
for f in sorted(os.listdir(env.dataset())):
if f.endswith(".wav"):
data_source.append(env.dataset(f))
cm = ConvModel(
batch_size = 30000,
filter_len = 150,
filters_num = 100,
target_sr = 3000,
gamma = 1e-03,
strides = 8,
avg_window = 5,
lrate = 1e-04
)
sess = tf.Session()
dataset = cm.form_dataset(data_source, proportion = 0.1)
cm.train(sess, dataset, 10000)
cm.evaluate_and_save(sess, dataset)
cm.serialize(sess)
from matplotlib import pyplot as plt
from os.path import join as pj
from util import setup_logging
from conv_model import ConvModel
from env import current as env
setup_logging(logging.getLogger())
data_source = []
for f in sorted(os.listdir(env.dataset())):
if f.endswith(".wav"):
data_source.append(env.dataset(f))
cm = ConvModel(batch_size=30000,
filter_len=150,
filters_num=100,
target_sr=3000,
gamma=1e-03,
strides=8,
avg_window=5,
lrate=1e-04)
sess = tf.Session()
dataset = cm.form_dataset(data_source, proportion=0.1)
cm.train(sess, dataset, 10000)
cm.evaluate_and_save(sess, dataset)
cm.serialize(sess)
inputs_v = seq_out[step_id:(step_id+bptt_steps)]
elif len(seq_out)>0:
for b_id in xrange(batch_size):
if p_sub > np.random.random_sample():
for seq_id in xrange(bptt_steps):
inputs_v[seq_id][b_id, :] = seq_out[step_id + seq_id][b_id, :]
feed_dict = {k: v for k, v in zip(inputs, inputs_v) }
feed_dict[state] = state_v
feed_dict.update({k: v for k, v in zip(targets, targets_v)})
outputs_v, state_v, loss_v, _ = sess.run([outputs, finstate, loss, train_step], feed_dict)
new_seq_out += outputs_v
losses.append(loss_v)
seq_out = new_seq_out
batch_ids += forecast_step
gc.collect()
print "Epoch {}, learning rate {}".format(e, ep_lrate), "train loss {}".format(sum(losses)/float(len(losses)))
if e % 100 == 0 and e > 0:
print "Generating sample"
generated = generate()
generated = np.asarray(generated).reshape(len(generated), input_size)
cm = ConvModel.deserialize()
waveform = cm.restore_hidden(generated, sess = sess)
cm.save_waveform_as(waveform, 0, env.run("{}_generated.wav".format(e)))
print "Saving model {}".format(saver.save(sess, model_fname))
model_fname = env.run("model.ckpt")
batch_size = 30000
L = 150
filters_num = 100
target_sr = 3000
gamma = 1e-03
epochs = 2000
lrate = 1e-04
k = 8 # filter strides
avg_size = 5
sel = None
cm = ConvModel(batch_size, L, filters_num, k, avg_size, lrate, gamma)
sess = tf.Session()
saver = tf.train.Saver()
if os.path.exists(model_fname):
print "Restoring from {}".format(model_fname)
saver.restore(sess, model_fname)
epochs = 0
else:
sess.run(tf.initialize_all_variables())
def read_song(source_id):
song_data_raw, source_sr = lr.load(data_source[source_id])
song_data = lr.resample(song_data_raw, source_sr, target_sr, scale=True)
song_data = song_data[:song_data.shape[0]/10]