def run_training():
if not os.path.exists(os.path.dirname(FLAGS.checkpoints_dir)):
os.mkdir(os.path.dirname(FLAGS.checkpoints_dir))
if not os.path.exists(FLAGS.checkpoints_dir):
os.mkdir(FLAGS.checkpoints_dir)
poems_vector, word_to_int, vocabularies,out_vector = process_poems(FLAGS.file_path)
x_batches, y_batches,length,outlength = generate_batch(FLAGS.batch_size, poems_vector, word_to_int,out_vector)
input_data = [tf.placeholder(tf.int32, (None,)) for t in range(length)]
labels = [tf.placeholder(tf.int32, (None,)) for t in range(length)]
output_targets = [tf.placeholder(tf.int32, (None,)) for t in range(outlength)]
rnn_model("gru",input_data,output_targets,labels,len(vocabularies))
for i in range(len(x_batches)):
print("training epoch:"+str(i))
xb = x_batches[i]
yb = y_batches[i]
for j in range(len(xb)):
X = xb[j]
Y = yb[j]
feed_dict = {input_data[t]: X[t] for t in range(length)}
feed_dict.update({labels[t]: Y[t] for t in range(length)})
_, loss_t, summary = sess.run([train_op, loss, summary_op], feed_dict)
# loss_t,summary = train_batch(x_batches[i],y_batches[i])
summary_writer.add_summary(summary,i)
summary_writer.flush()
def create_dataloader(dataset,
trans_fn=None,
mode='train',
batch_size=1,
pad_token_id=0):
"""
Creats dataloader.
Args:
dataset(obj:`paddle.io.Dataset`): Dataset instance.
mode(obj:`str`, optional, defaults to obj:`train`): If mode is 'train', it will shuffle the dataset randomly.
batch_size(obj:`int`, optional, defaults to 1): The sample number of a mini-batch.
pad_token_id(obj:`int`, optional, defaults to 0): The pad token index.
Returns:
dataloader(obj:`paddle.io.DataLoader`): The dataloader which generates batches.
"""
if trans_fn:
dataset = dataset.map(trans_fn, lazy=True)
shuffle = True if mode == 'train' else False
sampler = paddle.io.BatchSampler(
dataset=dataset, batch_size=batch_size, shuffle=shuffle)
dataloader = paddle.io.DataLoader(
dataset,
batch_sampler=sampler,
return_list=True,
collate_fn=lambda batch: data.generate_batch(batch, pad_token_id=pad_token_id))
return dataloader
def test_cache(composers):
all_pieces = dict()
for composer in composers:
start = time.time()
pieces = data.getpices(path="../midis", composer=composer)
print("Loading {} {} pieces took {} seconds".format(
composer, len(pieces),
time.time() - start))
all_pieces.update(pieces)
midi_cache = data.initialize_cache(all_pieces)
gen = data.generate_batch(midi_cache, batch_size=10)
for i in range(5):
pre_start = time.time()
batch_in, batch_out = next(gen)
print("One batch took {}".format(time.time() - pre_start))
print("Batch iput is size: ({}, {}, {})".format(
len(batch_in), len(batch_in[0]), len(batch_in[0][0])))
print("Batch output is size: ({}, {}, {})".format(
len(batch_out), len(batch_out[0]), len(batch_out[0][0])))
def run_episode(data_type, sumvar):
'''run over batches
return different losses
type: 'train', 'val' or 'test'
'''
activation_images = []
depth_predictions = []
endpoint_activations = []
start_time = time.time()
data_loading_time = 0
calculation_time = 0
start_data_time = time.time()
tot_loss = []
ctr_loss = []
dep_loss = []
odo_loss = []
q_loss = []
for index, ok, batch in data.generate_batch(data_type):
data_loading_time += (time.time() - start_data_time)
start_calc_time = time.time()
if ok:
inputs = np.array([_['img'] for _ in batch])
state = []
targets = np.array([[_['ctr']] for _ in batch])
# target_depth = np.array([_['depth'] for _ in batch]).reshape((-1,55,74,FLAGS.n_frames if FLAGS.n_fc else 1)) if FLAGS.auxiliary_depth or FLAGS.rl else []
target_depth = np.array([_['depth'] for _ in batch]).reshape(
(-1, 55, 74)) if FLAGS.auxiliary_depth or FLAGS.rl else []
target_odom = np.array([_['odom'] for _ in batch]).reshape(
(-1, 6)) if FLAGS.auxiliary_odom else []
# target_odom = np.array([_['odom'] for _ in batch]).reshape((-1,4)) if FLAGS.auxiliary_odom else []
prev_action = np.array([_['prev_act'] for _ in batch]).reshape(
(-1, 1)) if FLAGS.auxiliary_odom else []
if data_type == 'train':
losses = model.backward(inputs,
state,
targets,
depth_targets=target_depth,
odom_targets=target_odom,
prev_action=prev_action)
elif data_type == 'val' or data_type == 'test':
state, losses, aux_results = model.forward(
inputs,
state,
auxdepth=False,
auxodom=False,
prev_action=prev_action,
targets=targets,
target_depth=target_depth,
target_odom=target_odom)
tot_loss.append(losses['t'])
if not FLAGS.no_control and (not FLAGS.rl
or FLAGS.auxiliary_ctr):
ctr_loss.append(losses['c'])
if FLAGS.auxiliary_depth: dep_loss.append(losses['d'])
if FLAGS.auxiliary_odom: odo_loss.append(losses['o'])
if FLAGS.rl: q_loss.append(losses['q'])
if index == 1 and data_type == 'val':
if FLAGS.plot_activations:
activation_images = model.plot_activations(
inputs, targets)
if FLAGS.plot_depth:
depth_predictions = model.plot_depth(
inputs, target_depth)
if FLAGS.plot_histograms:
# stime = time.time()
endpoint_activations = model.get_endpoint_activations(
inputs)
# print('plot activations: {}'.format((stime-time.time())))
calculation_time += (time.time() - start_calc_time)
start_data_time = time.time()
sumvar['loss_' + data_type + '_total'] = np.mean(tot_loss)
if not FLAGS.no_control and (not FLAGS.rl or FLAGS.auxiliary_ctr):
sumvar['loss_' + data_type + '_control'] = np.mean(ctr_loss)
if FLAGS.auxiliary_depth:
sumvar['loss_' + data_type + '_depth'] = np.mean(dep_loss)
if FLAGS.auxiliary_odom:
sumvar['loss_' + data_type + '_odom'] = np.mean(odo_loss)
if FLAGS.rl: sumvar['loss_' + data_type + '_q'] = np.mean(q_loss)
if len(activation_images) != 0:
sumvar['conv_activations'] = activation_images
if len(depth_predictions) != 0:
sumvar['depth_predictions'] = depth_predictions
if FLAGS.plot_histograms:
for i, ep in enumerate(model.endpoints):
sumvar['activations_{}'.format(ep)] = endpoint_activations[i]
print(
'>>{0} [{1[2]}/{1[1]}_{1[3]:02d}:{1[4]:02d}]: data {2}; calc {3}'.
format(data_type.upper(), tuple(time.localtime()[0:5]),
print_dur(data_loading_time), print_dur(calculation_time)))
# print('losses: tot {0:.3g}; ctrl {1:.3g}; depth {2:.3g}; odom {2:.3g}; q {3:.3g}'.format(np.mean(tot_loss), np.mean(ctr_loss), np.mean(dep_loss), np.mean(odo_loss), np.mean(q_loss)))
if data_type == 'val' or data_type == 'test':
print('{}'.format(str(sumvar)))
sys.stdout.flush()
return sumvar
def main():
print("ROOT = {}".format(ROOT))
vae = models.VAE().to(device=DEVICE)
optimizer = torch.optim.Adamax(vae.parameters(), lr=2e-5, betas=(0.5, 0.9))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer,
lambda e: max(0.5**e, 0.05))
if CONTINUE is not None:
vae.load_state_dict(
torch.load(ROOT + 'dump-{:06d}.pt'.format(CONTINUE)))
# pre-advance the learning rate
for _ in range(CONTINUE // 20000):
scheduler.step()
start_iter = CONTINUE + 1
else:
start_iter = 0
# run loop
with open(ROOT + 'out.csv', 'w') as logfile:
print("batch;wi losses;z loss;x loss;ELBO", file=logfile)
for b in range(start_iter, ITERNUM):
if b % 20000 == 0:
scheduler.step()
vae.zero_grad()
bmin = 1
bmax = max(2, min(6, 1 + int(b / 2500)))
loss_w = torch.zeros((), device=DEVICE)
loss_z = torch.zeros((), device=DEVICE)
loss_x = torch.zeros((), device=DEVICE)
images = []
recs = []
gens = []
for _ in range(BATCH_SPLIT):
(img, col_labels,
loc_labels) = data.generate_batch(BATCHLEN // BATCH_SPLIT,
(bmin, bmax),
device=DEVICE)
(rec_img, l_w, l_z, l_x,
K) = vae.rec_with_losses(img, (col_labels, loc_labels))
loss_w += l_w / BATCH_SPLIT
loss_z += l_z / BATCH_SPLIT
loss_x += l_x / BATCH_SPLIT
images.append(img)
recs.append(rec_img.detach())
with torch.no_grad():
gens.append(vae.generate((col_labels, loc_labels)))
elbo = loss_w + loss_z + loss_x
print("{};{:.2e};{:.2e};{:.2e};{:.2e}".format(
b, float(loss_w), float(loss_z), float(loss_x), float(elbo)),
file=logfile)
logfile.flush()
# some flukes can make the loss get incredibly big in rare occasions, ignore them
if math.isfinite(float(elbo)) and float(elbo) < 1e8:
elbo.backward()
optimizer.step()
if b % 100 == 0:
save_image(torch.cat(images, dim=0),
ROOT + "{0:06d}-orig.png".format(b),
nrow=16)
save_image(torch.cat(recs, dim=0),
ROOT + "{0:06d}-rec.png".format(b),
nrow=16)
save_image(torch.cat(gens, dim=0),
ROOT + "{0:06d}-gen.png".format(b),
nrow=16)
if b % MODEL_DUMP_PERIOD == 0:
torch.save(vae.state_dict(), ROOT + 'dump-{:06d}.pt'.format(b))
def train(self,
cache,
batch_size=32,
predict_freq=100,
show_freq=10,
save_freq=500,
max_epoch=10000,
saveto='NewSong',
step=319,
conservativity=1,
pre_trained_model=None):
'''
This is the train function for the biaxial_model. It alos predicts while trianing.
Args:
pieces: dict, containing the all note statematrixs as training set;
batch_size: how many sample in one batch;
predict_freq: int, every predict_freq we generate a new song;
save_freq: int, the frequency to save the model;
show_freq: int, the frequency to show the loss;
max_epoch: max steps we gonna train;
saveto: str, the dir we save or new songs;
step: int, the length we generate a new song. one step means on step for the time model
conservativity: The conservativity of number of notes of song we generate.
pre_trained_model: used for restore training
'''
cur_model_name = 'biaxial_rnn_{}'.format(int(time.time()))
batch_generator = data.generate_batch(cache, batch_size)
val_barch_generator = data.generate_val_batch(cache, batch_size)
loss_log = []
min_loss = np.inf
with tf.Session() as sess:
merge = tf.summary.merge_all()
writer = tf.summary.FileWriter("log/{}".format(cur_model_name),
sess.graph)
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
if pre_trained_model is not None:
try:
print("Load the model from: {}".format(pre_trained_model))
saver.restore(sess, 'model/{}'.format(pre_trained_model))
#writer = tf.summary.FileWriterCache.get('log/{}'.format(pre_trained_model))
except Exception:
print("Load model Failed!")
pass
if self.new_trainer is not None:
optimizer = self.optimizer2
else:
optimizer = self.optimizer
for i in range(max_epoch):
X_train, y_train = next(batch_generator)
_, loss, merge_result = sess.run((optimizer, self.loss, merge),
feed_dict={
self.input_mat: X_train,
self.output_mat: y_train
})
loss_log.append(loss)
pickle.dump(
loss_log,
open('model/' + cur_model_name + '_loss_log.pkl', 'wb'))
if i % show_freq == 0:
print('Step {0}: loss is {1}'.format(i, loss))
if (i + 1) % 10 == 0:
writer.add_summary(merge_result, i)
# generate a new song
if (i + 1) % predict_freq == 0:
xIpt, xOpt = map(np.array, data.getPieceSegment(cache))
new_state_matrix = sess.run(self.new_song,
feed_dict={
self.predict_seed:
xIpt[0],
self.step_to_sumulate:
[step],
self.conservativity:
[conservativity]
})
newsong = np.concatenate(
(np.expand_dims(xOpt[0], 0), new_state_matrix))
songname = str(time.time()) + '.mid'
if not os.path.exists(os.path.join(saveto,
cur_model_name)):
os.makedirs(os.path.join(saveto, cur_model_name))
noteStateMatrixToMidi(newsong,
name=os.path.join(
saveto, cur_model_name,
songname))
print('New Songs {} saved to \'{}\''.format(
songname, os.path.join(saveto, cur_model_name)))
# save the models for restoring training
if (i + 1) % save_freq == 0:
if not os.path.exists('model/'):
os.makedirs('model/')
saver.save(sess, 'model/{}'.format(cur_model_name))
print('{} Saved'.format(cur_model_name))
# Get validation data and validate
xIpt_val, xOpt_val = next(val_barch_generator)
val_loss = sess.run((self.loss),
feed_dict={
self.input_mat: xIpt_val,
self.output_mat: xOpt_val
})
print(
"Validation of loss of {} achieved on step {}".format(
val_loss, i))
if val_loss < min_loss:
min_loss = val_loss
# Save the best model
saver.save(
sess, 'model/{}_{}'.format('best', cur_model_name))
print('{}_{} Saved'.format('best', cur_model_name))
transpose_b=True)
#tensorflow变量初始化
init = tf.global_variables_initializer()
#训练
num_steps = 1001
with tf.Session(graph=graph) as session:
#初始化所有变量
init.run()
print("Initialized")
average_loss = 0
for step in range(num_steps):
batch_inputs, batch_labels = data.generate_batch(
batch_size, num_skips, skip_window)
feed_dict = {train_inputs: batch_inputs, train_labels: batch_labels}
#更新参数
_, loss_val = session.run([optimizer, loss], feed_dict=feed_dict)
average_loss += loss_val
if step % 200 == 0:
if step > 0:
average_loss /= 2000
print("Average loss at step", step, ":", average_loss)
average_loss = 0
#计算验证样本的相似度
if step % 1000 == 0:
sim = similarity.eval()
def main():
if X_SIGMA is not None:
vae = models.VAE(x_logvar=2 * math.log(X_SIGMA)).to(device=DEVICE)
else:
vae = models.VAE().to(device=DEVICE)
optimizer = torch.optim.Adamax(vae.parameters(), lr=1e-4, betas=(0.5, 0.9))
scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer,
lambda e: max(0.5**e, 0.01))
# run loop
with open(ROOT + 'out.csv', 'w') as logfile:
print('batch;wi losses;z loss;x loss;ELBO;', file=logfile)
for b in range(ITERNUM):
if b % 20000 == 0:
scheduler.step()
vae.zero_grad()
min_curves = 1
max_curves = max(2, min(16, 1 + int(b / 2500)))
# average the loss over different batch sizes
loss_w = torch.zeros((), device=DEVICE)
loss_z = torch.zeros((), device=DEVICE)
loss_x = torch.zeros((), device=DEVICE)
for _ in range(BATCH_SPLIT):
(fs, curves) = data.generate_batch(BATCHLEN // BATCH_SPLIT,
freqrg=(1, 10),
nfreqrg=(min_curves,
max_curves),
device=DEVICE)
(l_w, l_z, l_x, K) = vae.rec_losses(curves, fs)
loss_w += l_w / BATCH_SPLIT
loss_z += l_z / BATCH_SPLIT
loss_x += l_x / BATCH_SPLIT
elbo = loss_w + loss_z + loss_x
print("{};{:.2e};{:.2e};{:.2e};{:.2e}".format(
b, float(loss_w), float(loss_z), float(loss_x), float(elbo)),
file=logfile)
logfile.flush()
elbo.backward()
# update gradients
optimizer.step()
if b % DUMP_PERIOD == 0:
(mus, sigmas) = vae.rec_output(curves, fs)
with open(ROOT + 'curve-{:06d}.csv'.format(b), 'w') as outfile:
print('x;mu;sigma', file=outfile)
for i in range(curves.size(1)):
print("{};{:.3f};{:.3f};{:.3f}".format(
i, curves[0, i], mus[0, i], sigmas[0, i]),
file=outfile)
xs = vae.generate_same_latent(fs[0:1, :], 4)
with open(ROOT + 'gen-{:06d}.csv'.format(b), 'w') as outfile:
for i in range(xs.size(1)):
print("{}".format(i), file=outfile, end='')
for j in range(xs.size(0)):
print(";{:.3f}".format(xs[j, i]),
file=outfile,
end='')
print("", file=outfile)
if b % MODEL_DUMP_PERIOD == 0:
torch.save(vae.state_dict(), ROOT + 'dump-{:06d}.pt'.format(b))
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import numpy as np
from data import produce_signals
from data import generate_batch
signals = produce_signals()
# for a in signals:
# print(a)
x_batches, y_batches = generate_batch(10, signals)
for i in range(10):
print("batch %d", i)
print(x_batches[i])
print(y_batches[i])
# tmp = [1, 2, 3, 4, 5, 6]
# xdata = np.zeros((1, 5))
# ydata = np.zeros((1, 5))
# print(tmp)
# xdata = tmp[: -1]
# ydata = tmp[: 1]
# print(xdata)
# print(ydata)
def run_episode(mode, sumvar, model, update_importance_weights=False):
'''run over batches
return different losses
type: 'train', 'val' or 'test'
'''
depth_predictions = []
start_time = time.time()
data_loading_time = 0
calculation_time = 0
start_data_time = time.time()
# results = {}
# results['control'] = []
results = {'total': []}
results['ce'] = []
for k in model.lll_losses.keys():
results['lll_' + k] = []
# results['accuracy'] = []
# if FLAGS.auxiliary_depth: results['depth'] = []
all_inputs = []
for index, ok, batch in data.generate_batch(mode):
data_loading_time += (time.time() - start_data_time)
start_calc_time = time.time()
if ok:
inputs = np.array([_['img'] for _ in batch])
targets = np.array([[_['ctr']] for _ in batch])
if update_importance_weights and len(all_inputs) < 2000:
try:
all_inputs = np.concatenate([all_inputs, inputs], axis=0)
except:
all_inputs = inputs[:]
# print("targets: {}".format(targets))
# try:
target_depth = np.array([_['depth'] for _ in batch]).reshape(
(-1, 55, 74)) if FLAGS.auxiliary_depth else []
if len(target_depth) == 0 and FLAGS.auxiliary_depth:
raise ValueError('No depth in batch.')
if mode == 'train':
# model.backward(inputs, targets=targets, depth_targets=target_depth, sumvar=sumvar)
losses = model.backward(inputs,
targets=targets,
depth_targets=target_depth,
sumvar=sumvar)
for k in losses.keys():
results[k].append(losses[k])
elif mode == 'val' or mode == 'test':
_, aux_results = model.forward(inputs,
auxdepth=False,
targets=targets,
depth_targets=target_depth)
if index == 1 and mode == 'val' and FLAGS.plot_depth:
depth_predictions = tools.plot_depth(inputs, target_depth,
model)
else:
print('Failed to run {}.'.format(mode))
calculation_time += (time.time() - start_calc_time)
start_data_time = time.time()
if update_importance_weights:
model.update_importance_weights(all_inputs)
for k in results.keys():
if len(results[k]) != 0:
sumvar['Loss_' + mode + '_' + k] = np.mean(results[k])
if len(depth_predictions) != 0:
sumvar['depth_predictions'] = depth_predictions
print('>>{0} [{1[2]}/{1[1]}_{1[3]:02d}:{1[4]:02d}]: data {2}; calc {3}'.
format(mode.upper(), tuple(time.localtime()[0:5]),
tools.print_dur(data_loading_time),
tools.print_dur(calculation_time)))
return sumvar
