def train(num_layers, length, rate, batch, epochs, load_path, save_path,
preproc_dir):
if rate is None:
rate = callback.CustomSchedule(par.embedding_dim)
preproc_dir = Path(preproc_dir)
model = MusicTransformerDecoder(
embedding_dim=256,
vocab_size=par.vocab_size,
num_layer=num_layers,
max_seq=length,
dropout=0.2,
debug=False,
loader_path=load_path,
)
model.compile(
optimizer=Adam(rate, beta_1=0.9, beta_2=0.98, epsilon=1e-9),
loss=callback.transformer_dist_train_loss,
)
time = datetime.datetime.now().strftime("%Y-%m-%d-%H:%M:%S")
train_summary_writer = tf.summary.create_file_writer(
f"logs/mt_decoder/{time}/train")
eval_summary_writer = tf.summary.create_file_writer(
f"logs/mt_decoder/{time}/eval")
dataset = Data(preproc_dir)
idx = 0
with click.progressbar(length=epochs) as prog:
for e in prog:
model.reset_metrics()
with click.progressbar(length=len(dataset.files) //
batch) as prog2:
for b in prog2:
batch_x, batch_y = dataset.slide_seq2seq_batch(
batch, length)
loss, acc = model.train_on_batch(batch_x, batch_y)
if b % 100 == 0:
eval_x, eval_y = dataset.slide_seq2seq_batch(
batch, length, "eval")
(eloss, eacc), weights = model.evaluate(eval_x, eval_y)
if save_path is not None:
model.save(save_path)
with train_summary_writer.as_default():
if b == 0:
tf.summary.histogram("target_analysis",
batch_y,
step=e)
tf.summary.histogram("source_analysis",
batch_x,
step=e)
tf.summary.scalar("loss", loss, step=idx)
tf.summary.scalar("accuracy", acc, step=idx)
with eval_summary_writer.as_default():
if b == 0:
model.sanity_check(eval_x, eval_y, step=e)
tf.summary.scalar("loss", eloss, step=idx)
tf.summary.scalar("accuracy", eacc, step=idx)
for i, weight in enumerate(weights):
with tf.name_scope("layer_%d" % i):
with tf.name_scope("w"):
utils.attention_image_summary(weight,
step=idx)
print(
f"Loss: {loss:6.6} (e: {eloss:6.6}), Accuracy: {acc} (e: {eacc})"
)
idx += 1
result_metrics = mt.train_on_batch(batch_x, batch_y)
if b % 100 == 0:
eval_x, eval_y = dataset.slide_seq2seq_batch(batch_size, max_seq, 'eval')
eval_result_metrics, weights = mt.evaluate(eval_x, eval_y)
mt.save(save_path)
with train_summary_writer.as_default():
if b == 0:
tf.summary.histogram("target_analysis", batch_y, step=e)
tf.summary.histogram("source_analysis", batch_x, step=e)
tf.summary.scalar('loss', result_metrics[0], step=idx)
tf.summary.scalar('accuracy', result_metrics[1], step=idx)
with eval_summary_writer.as_default():
if b == 0:
mt.sanity_check(eval_x, eval_y, step=e)
tf.summary.scalar('loss', eval_result_metrics[0], step=idx)
tf.summary.scalar('accuracy', eval_result_metrics[1], step=idx)
for i, weight in enumerate(weights):
with tf.name_scope("layer_%d" % i):
with tf.name_scope("w"):
utils.attention_image_summary(weight, step=idx)
# for i, weight in enumerate(weights):
# with tf.name_scope("layer_%d" % i):
# with tf.name_scope("_w0"):
# utils.attention_image_summary(weight[0])
# with tf.name_scope("_w1"):
# utils.attention_image_summary(weight[1])
idx += 1
print('\n====================================================')
def train(input_path, save_path, l_r=None, batch_size=2,
max_seq=1024, epochs=100,
load_path=None, num_layer=6, log_dir='/pfs/out/logs'):
# load data
dataset = Data(input_path)
print('dataset', dataset)
# load model
learning_rate = callback.CustomSchedule(par.embedding_dim) if l_r is None else l_r
opt = Adam(learning_rate, beta_1=0.9, beta_2=0.98, epsilon=1e-9)
# define model
mt = MusicTransformerDecoder(
embedding_dim=256,
vocab_size=par.vocab_size,
num_layer=num_layer,
max_seq=max_seq,
dropout=0.2,
debug=False, loader_path=load_path)
mt.compile(optimizer=opt, loss=callback.transformer_dist_train_loss)
# define tensorboard writer
current_time = datetime.datetime.now().strftime('%Y%m%d-%H%M%S')
train_log_dir = '{log_dir}/{time}/train'.format(log_dir=log_dir, time=current_time)
eval_log_dir = '{log_dir}/{time}/eval'.format(log_dir=log_dir, time=current_time)
train_summary_writer = tf.summary.create_file_writer(train_log_dir)
eval_summary_writer = tf.summary.create_file_writer(eval_log_dir)
# Train Start
idx = 0
batchings = len(dataset.files) // batch_size
how_often_to_print = 50
for e in tqdm(range(epochs), desc='epochs'):
mt.reset_metrics()
for b in tqdm(range(batchings), desc='batches'):
try:
batch_x, batch_y = dataset.slide_seq2seq_batch(batch_size, max_seq)
except:
continue
result_metrics = mt.train_on_batch(batch_x, batch_y)
if b % how_often_to_print == 0:
eval_x, eval_y = dataset.slide_seq2seq_batch(batch_size, max_seq, 'eval')
eval_result_metrics, weights = mt.evaluate(eval_x, eval_y)
mt.save(save_path)
with train_summary_writer.as_default():
if b == 0:
tf.summary.histogram("target_analysis", batch_y, step=e)
tf.summary.histogram("source_analysis", batch_x, step=e)
tf.summary.scalar('loss', result_metrics[0], step=idx)
tf.summary.scalar('accuracy', result_metrics[1], step=idx)
with eval_summary_writer.as_default():
if b == 0:
mt.sanity_check(eval_x, eval_y, step=e)
tf.summary.scalar('loss', eval_result_metrics[0], step=idx)
tf.summary.scalar('accuracy', eval_result_metrics[1], step=idx)
for i, weight in enumerate(weights):
with tf.name_scope("layer_%d" % i):
with tf.name_scope("w"):
utils.attention_image_summary(weight, step=idx)
# for i, weight in enumerate(weights):
# with tf.name_scope("layer_%d" % i):
# with tf.name_scope("_w0"):
# utils.attention_image_summary(weight[0])
# with tf.name_scope("_w1"):
# utils.attention_image_summary(weight[1])
idx += 1
print('\n====================================================')
print('Epoch/Batch: {}/{}'.format(e, b))
print('Train >>>> Loss: {:6.6}, Accuracy: {}'.format(result_metrics[0], result_metrics[1]))
print('Eval >>>> Loss: {:6.6}, Accuracy: {}'.format(eval_result_metrics[0], eval_result_metrics[1]))