def validate(self, type):
'''Handles validation set data'''
input_eval_path = os.path.join(self.dirs['eval_path'], "inputs")
vel_eval_path = os.path.join(self.dirs['eval_path'], "velocities")
c_input_eval_path = os.path.join(input_eval_path, "classical")
c_vel_eval_path = os.path.join(vel_eval_path, "classical")
j_input_eval_path = os.path.join(input_eval_path, "jazz")
j_vel_eval_path = os.path.join(vel_eval_path, "jazz")
if type == "classical":
input_folder = os.listdir(c_input_eval_path)
file_count = len(input_folder)
vel_eval_path = c_vel_eval_path
input_eval_path = c_input_eval_path
else:
input_folder = os.listdir(j_input_eval_path)
file_count = len(input_folder)
vel_eval_path = j_vel_eval_path
input_eval_path = j_input_eval_path
#CLASSICS
in_list = []
out_list = []
filenames = []
for i, filename in enumerate(input_folder):
if filename.split('.')[-1] == 'npy':
vel_path = os.path.join(vel_eval_path, filename)
input_path = os.path.join(input_eval_path, filename)
true_vel = np.load(vel_path) / 127
loaded = np.load(input_path)
if not self.one_hot:
loaded = loaded / 2
in_list.append(loaded)
out_list.append(true_vel)
filenames.append(filename)
valid_generator = BatchGenerator(in_list,
out_list,
self.batch_count,
self.input_size,
self.output_size,
mini=False)
return valid_generator
def train(self, data, model=None, starting_epoch=0, clip_grad=True, epochs=1001, input_keep_prob=0.5, output_keep_prob=0.5,
learning_rate=0.001 , eval_epoch=20,val_epoch=10, save_epoch=1):
self.data = data
if clip_grad:
jazz_optimizer = self.clip_optimizer(learning_rate,self.jazz_loss)
classical_optimizer = self.clip_optimizer(learning_rate,self.classical_loss)
else:
jazz_optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.jazz_loss)
classical_optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.classical_loss)
self.sess = tf.Session()
self.c_in_list, self.c_out_list,self.c_input_lens, self.c_files = self.eval_set('classical')
self.j_in_list, self.j_out_list,self.j_input_lens, self.j_files = self.eval_set('jazz')
if model:
self.load(model)
else:
self.sess.run(tf.global_variables_initializer())
self.summary_op = tf.summary.merge_all()
self.train_writer = tf.summary.FileWriter(os.path.join(self.dirs['logs_path'], 'train'), graph=self.sess.graph_def)
self.test_writer = tf.summary.FileWriter(os.path.join(self.dirs['logs_path'], 'test'), graph=self.sess.graph_def)
classical_batcher = BatchGenerator(self.data["classical"]["X"], self.data["classical"]["Y"], self.batch_count, self.input_size, self.output_size, mini=self.mini)
jazz_batcher = BatchGenerator(self.data["jazz"]["X"], self.data["jazz"]["Y"], self.batch_count, self.input_size, self.output_size, mini=self.mini)
self.v_classical_batcher = self.validate("classical")
self.v_classical_batcher = self.v_classical_batcher.batch()
self.v_jazz_batcher = self.validate("jazz")
self.v_jazz_batcher = self.v_jazz_batcher.batch()
classical_generator = classical_batcher.batch()
jazz_generator = jazz_batcher.batch()
print("[*] Initiating training...")
for epoch in range(starting_epoch, epochs):
classical_epoch_avg = 0
jazz_epoch_avg = 0
print("[*] Epoch %d" % epoch)
for batch in range(classical_batcher.batch_count):
batch_X, batch_Y, batch_len = classical_generator.next()
batch_len = [batch_len] * len(batch_X)
epoch_error, classical_summary, _ = self.sess.run([self.classical_loss,
self.summary_op,
classical_optimizer,
], feed_dict={ self.inputs: batch_X,
self.true_classical_outputs: batch_Y,
self.true_jazz_outputs: batch_Y,
self.seq_len: batch_len,
self.input_keep_prob: input_keep_prob,
self.output_keep_prob: output_keep_prob})
classical_epoch_avg += epoch_error
print("\tBatch %d/%d, Training MSE for Classical batch: %.9f" % (batch+1, classical_batcher.batch_count, epoch_error))
self.train_writer.add_summary(classical_summary, epoch*classical_batcher.batch_count + epoch)
for batch in range(jazz_batcher.batch_count):
batch_X, batch_Y, batch_len = jazz_generator.next()
batch_len = [batch_len] * len(batch_X)
epoch_error, jazz_summary, _ = self.sess.run([self.jazz_loss,
self.summary_op,
jazz_optimizer,
], feed_dict={ self.inputs: batch_X,
self.true_jazz_outputs: batch_Y,
self.true_classical_outputs: batch_Y,
self.seq_len: batch_len,
self.input_keep_prob: input_keep_prob,
self.output_keep_prob: output_keep_prob})
jazz_epoch_avg += epoch_error
print("\tBatch %d/%d, Training MSE for Jazz batch: %.9f" % (batch+1, jazz_batcher.batch_count, epoch_error))
self.train_writer.add_summary(jazz_summary, epoch*jazz_batcher.batch_count + epoch)
# self.validation(epoch)
print("[*] Average Training MSE for Classical epoch %d: %.9f" % (epoch, classical_epoch_avg/classical_batcher.batch_count))
print("[*] Average Training MSE for Jazz epoch %d: %.9f" % (epoch, jazz_epoch_avg/jazz_batcher.batch_count))
if epoch % val_epoch == 0 :
print("[*] Validating model...")
self.validation(epoch)
if epoch % save_epoch == 0 :
self.save(epoch)
if epoch % eval_epoch == 0 :
print("[*] Evaluating model...")
self.evaluate(epoch)
print("[*] Training complete.")
def validate(self, type):
'''Handles validation set data'''
# input_eval_path = os.path.join(self.dirs['eval_path'], "inputs")
# vel_eval_path = os.path.join(self.dirs['eval_path'], "velocities")
# c_input_eval_path = os.path.join(input_eval_path, "classical")
# c_vel_eval_path = os.path.join(vel_eval_path, "classical")
# # j_input_eval_path = os.path.join(input_eval_path, "jazz")
# # j_vel_eval_path = os.path.join(vel_eval_path, "jazz")
# if type == "classical":
# input_folder = os.listdir(c_input_eval_path)
# file_count = len(input_folder)
# vel_eval_path = c_vel_eval_path
# input_eval_path = c_input_eval_path
# else:
# input_folder = os.listdir(j_input_eval_path)
# file_count = len(input_folder)
# vel_eval_path = j_vel_eval_path
# input_eval_path = j_input_eval_path
# #CLASSICS
input_folder = self.dirs['eval_path']
in_list = []
out_list = []
filenames = []
input_lens = []
for i, filename in enumerate(os.listdir(input_folder)):
# if filename.split('.')[-1] == 'npy':
if filename.split('.')[-1] == 'mat':
filepath = os.path.join(input_folder, filename)
data = scipy.io.loadmat(filepath)
loaded_x = data['Xin']
loaded_y = data['Yout']
assert (loaded_x.shape[0] == loaded_y.shape[0])
in_list.append(loaded_x)
out_list.append(loaded_y)
filenames.append(filename)
input_len = [len(loaded_x)]
input_lens.append(input_len)
# in_list = []
# out_list = []
# filenames = []
# for i, filename in enumerate(input_folder):
# if filename.split('.')[-1] == 'npy':
# vel_path = os.path.join(vel_eval_path, filename)
# input_path = os.path.join(input_eval_path, filename)
# true_vel = np.load(vel_path)/127
# loaded = np.load(input_path)
# if not self.one_hot:
# loaded = loaded/2
# in_list.append(loaded)
# out_list.append(true_vel)
# filenames.append(filename)
valid_generator = BatchGenerator(in_list,
out_list,
self.batch_count,
self.input_size,
self.output_size,
mini=False)
return valid_generator