def experiment_1(encoder):
print("...loading&preprocessing dataset")
#train_x, train_y = load_data("data/enRep_USE_None.csv", 40000)
train_x, train_y, test_x, test_y, _, _ = utils.load_dataset(
'data/train.csv', 'data/trial.csv', 'data/trial.labels',
'data/test.csv')
train_x, test_x, _, max_string_length = preprocessing.preprocessing_pipeline(
train_x, test_x, _)
print("...creating encoder instance")
encoder_instance = encoder(train_x)
print("...encoding sentences")
embeddings = numpy.zeros((train_x.shape[0], encoder_instance.getSize()))
for i in range(0, train_x.shape[0], 1000):
print('\r......{0}/{1}'.format(i, train_x.shape[0]), end="")
embeddings[i:i + 1000] = encoder_instance.use(train_x[i:i + 1000])
for i in embeddings[0:10]:
print(i)
print("...preparing labels")
train_y_l = numpy.zeros((train_y.shape[0], 6))
for i in range(train_y.shape[0]):
if train_y[i] == 'anger':
train_y_l[i] = numpy.asarray([1, 0, 0, 0, 0, 0])
elif train_y[i] == 'sad':
train_y_l[i] = numpy.asarray([0, 1, 0, 0, 0, 0])
elif train_y[i] == 'joy':
train_y_l[i] = numpy.asarray([0, 0, 1, 0, 0, 0])
elif train_y[i] == 'disgust':
train_y_l[i] = numpy.asarray([0, 0, 0, 1, 0, 0])
elif train_y[i] == 'fear':
train_y_l[i] = numpy.asarray([0, 0, 0, 0, 1, 0])
elif train_y[i] == 'surprise':
train_y_l[i] = numpy.asarray([0, 0, 0, 0, 0, 1])
model = model_utils.get_SM_model_2(embeddings[0].shape, 6)
print("...Preparations complete:")
print("......Embeddings shape (data size, embedding size): {}".format(
embeddings.shape))
print("......Labels shape (data size, classes): {}".format(train_y.shape))
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=[
'accuracy', evaluation.f1, evaluation.precision,
evaluation.recall
])
callbacks = model_utils.get_callbacks(
early_stop_monitor=config.early_stop_monitor,
early_stop_patience=config.early_stop_patience,
early_stop_mode=config.early_stop_mode)
model_info = model.fit(embeddings,
train_y_l,
epochs=200,
batch_size=config.batch_size,
validation_split=0.1,
shuffle=True,
verbose=2)
def main():
midi_dir = '../../../data/query_symlinks'
model_num = 12
model_dir = '../../../models/keras/01_vanilla_rnn/{}'.format(model_num)
window_size = 20
# only creates if doesn't already exist
model_utils.create_model_dir(model_dir)
files = [os.path.join(midi_dir, path) for path in os.listdir(midi_dir)]
train_generator = data_utils.get_data_generator(files[0:1000],
form=FORMAT,
window_size=20)
val_generator = data_utils.get_data_generator(files[1000:1100],
form=FORMAT,
window_size=20)
model, epoch = get_model(window_size=window_size)
print(model.summary())
model_utils.save_model(model, model_dir)
callbacks = model_utils.get_callbacks(model_dir)
print('fitting model...')
model.fit_generator(train_generator,
steps_per_epoch=10000,
epochs=10,
validation_data=val_generator,
validation_steps=2000,
verbose=1,
callbacks=callbacks,
initial_epoch=epoch)
# generate 10 tracks using random seeds
X, y = val_generator.next()
generated = model_utils.generate(model, X)
for i, midi in enumerate(generated):
file = os.path.join(model_dir, 'generated', '{}.mid'.format(i + 1))
midi.write(file.format(i + 1))
print('wrote midi file to {}'.format(file))
def experiment_2(): # w/o candidates
test_x_s, test_y, test_label = prepare_testdata()
print("...initializing model")
model = model_utils.get_SM_model_2((4096, ),
embedding_layer=None,
units=config.units)
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
callbacks = model_utils.get_callbacks(
early_stop_monitor=config.early_stop_monitor,
early_stop_patience=config.early_stop_patience,
early_stop_mode=config.early_stop_mode)
model_info = model.fit_generator(
generators.MySMGenerator("data/enRep_InferSent_None_original.csv",
batch=32,
embedding_size=4096),
validation_data=generators.MySMValidationGenerator(
"data/enRep_InferSent_None_original.csv",
batch=32,
embedding_size=4096),
epochs=40,
callbacks=callbacks,
verbose=2)
print('...Evaluation')
loss, acc = model.evaluate(test_x_s, test_y, verbose=2)
print("...Loss = ", loss)
print("...Test accuracy = ", acc)
probabilities = model.predict(test_x_s)
predictions = utils.indices_to_labels(probabilities.argmax(axis=-1),
config.index_to_label)
evaluation.calculate_prf(test_label.tolist(), predictions)
embedding_layer=embeddings_layer,
units=config.units,
dtype=tf.string)
else:
model = model_utils.get_model(shape,
embeddings_layer,
config.classes,
units=config.units,
dtype=inp_type)
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
callbacks = model_utils.get_callbacks(
early_stop_monitor=config.early_stop_monitor,
early_stop_patience=config.early_stop_patience,
early_stop_mode=config.early_stop_mode)
weights = model_utils.get_sample_weights_prim(train_y, config.class_weight)
# model_info = model.fit(train_x_indices, train_y_oh, epochs=config.epochs, batch_size=config.batch_size, validation_split=0.05, callbacks=callbacks, shuffle=True, verbose=config.verbose)
model_info = model.fit(train_x_indices,
train_y_oh,
epochs=config.epochs,
batch_size=config.batch_size,
validation_data=(trial_x_indices, trial_y_oh),
callbacks=callbacks,
shuffle=True,
verbose=config.verbose,
sample_weight=weights)
def main():
midi_dir = '../../../data/query_symlinks'
model_num = 3
model_dir = '../../../models/keras/02_event_rnn/{}'.format(model_num)
rythm_test = False
window_size = 20
# only creates if doesn't already exist
model_utils.create_model_dir(model_dir)
files = [os.path.join(midi_dir, path) for path in os.listdir(midi_dir)]
train_files = files[0:1000]
val_files = files[1000:1250]
# get the train/val d
train_generator = data_utils.get_data_generator(train_files,
form=data_utils.F_EVENT_WINDOW_TIMES,
window_size=20)
val_generator = data_utils.get_data_generator(val_files,
form=data_utils.F_EVENT_WINDOW_TIMES,
window_size=20)
# train_generator = scale_data(train_generator)
# val_generator = scale_data(val_generator)
X_timing_seed, _ = val_generator.next()
# pdb.set_trace()
timing_model, epoch = get_timing_model(model_dir,
window_size=window_size,
model_index=1)
note_model, _ = get_note_model(model_dir,
window_size=window_size,
model_index=0)
model_utils.save_model(note_model, model_dir, model_index=0)
model_utils.save_model(timing_model, model_dir, model_index=1)
callbacks = model_utils.get_callbacks(model_dir,
checkpoint_monitor='val_mae',
model_index=1)
# print('fitting timing model...')
# timing_model.fit_generator(train_generator,
# steps_per_epoch=data_utils.WINDOWS_PER_FILE * len(train_files),
# epochs=30,
# validation_data=val_generator,
# validation_steps=data_utils.WINDOWS_PER_FILE * len(val_files),
# verbose=1,
# callbacks=callbacks,
# initial_epoch=epoch)
train_generator = data_utils.get_data_generator(train_files,
form=data_utils.F_EVENT_WINDOW_NOTES,
window_size=20)
callbacks = model_utils.get_callbacks(model_dir,
checkpoint_monitor='val_mae',
model_index=0)
val_generator = data_utils.get_data_generator(val_files,
form=data_utils.F_EVENT_WINDOW_NOTES,
window_size=20)
# print('fitting note model...')
# note_model.fit_generator(train_generator,
# steps_per_epoch=data_utils.WINDOWS_PER_FILE * 10,
# epochs=10,
# validation_data=val_generator,
# validation_steps=data_utils.WINDOWS_PER_FILE * 2,
# verbose=1,
# callbacks=callbacks,
# initial_epoch=epoch)
# generate 10 tracks using random seeds
X_note_seed, _ = val_generator.next()
print('generating notes...')
generated_notes = model_utils.generate_notes(note_model, X_note_seed)
print('generating timings...')
# replace start/end note events with values generated from timing models
generated_timings = model_utils.generate_timings(timing_model, X_timing_seed)
for i, midi in enumerate(generated_notes):
for instrument in midi.instruments:
wall_time = 0
if rythm_test:
instrument.program = 125
for j, note in enumerate(instrument.notes):
if rythm_test:
note.pitch = 60
offset = generated_timings[i][j][0]
duration = generated_timings[i][j][1]
note.start = wall_time + offset
note.end = wall_time + offset + duration
wall_time = wall_time + offset + duration
for i, midi in enumerate(generated_notes):
file = os.path.join(model_dir, 'generated', '{}.mid'.format(i + 1))
midi.write(file.format(i + 1))
print('wrote midi file to {}'.format(file))
val_ds = DictImageDataGenerator(preprocessing_function=preprocess_func)
val_gen = val_ds.flow_from_dict(val_d,
target_size=(IMG_SIZE, IMG_SIZE),
batch_size=BATCH_SIZE)
val_steps = val_gen.samples // BATCH_SIZE
print(val_steps)
# test
for x, y in train_gen:
print(x.shape, y.shape)
break
print(train_gen.class_indices)
mob_model = get_xception_model(IMG_SIZE, LABEL_CNT, dense_dim=None)
# mob_model.summary()
cb_list = get_callbacks(
'../data/checkpoints/detect_if_new_whale_bbox_xception.h5', mob_model)
adam_opt = keras.optimizers.Adam(lr=0.00001)
mob_model.compile(optimizer=adam_opt,
loss='categorical_crossentropy',
metrics=['acc'])
print('compile done')
# train
mob_model.fit_generator(train_gen,
steps_per_epoch=1000,
epochs=100,
verbose=1,
callbacks=cb_list,
validation_data=val_gen,
validation_steps=val_steps)
def main():
midi_dir = '../../../data/t-sne_selections/1/0/'
model_num = 7 # same as 3
model_dir = '../../../models/keras/' \
'03_vanilla_rnn_intervals/{}'.format(model_num)
window_size = 20
# only creates if doesn't already exist
model_utils.create_model_dir(model_dir)
files = [os.path.join(midi_dir, path) for path in os.listdir(midi_dir)]
val_split = 0.2
print(int(float(len(files)) * val_split))
train_generator = data_utils.get_data_generator(
files[0:int(float(len(files)) * val_split)],
form=FORMAT,
window_size=20)
val_generator = data_utils.get_data_generator(
files[int(float(len(files)) * val_split):],
form=FORMAT,
window_size=20)
# intervals = []
# for i in range(1000):
# data = train_generator.next()
# for window in data[0]:
# for interval in window:
# if interval not in intervals:
# intervals.append(interval)
# pdb.set_trace()
# for i in range(1000):
# data = train_generator.next()
# pdb.set_trace()
model, epoch = get_model(window_size=window_size)
print(model.summary())
model_utils.save_model(model, model_dir)
callbacks = model_utils.get_callbacks(model_dir)
# print('fitting model...')
model.fit_generator(train_generator,
steps_per_epoch=1000,
epochs=100,
validation_data=val_generator,
validation_steps=200,
verbose=1,
callbacks=callbacks,
initial_epoch=epoch)
# generate 10 tracks using random seeds
X, y = val_generator.next()
generated = model_utils.generate_notes_from_intervals(model, X, length=200)
for i, midi in enumerate(generated):
file = os.path.join(model_dir, 'generated', '{}.mid'.format(i + 1))
midi.write(file.format(i + 1))
print('wrote midi file to {}'.format(file))
print(x1.shape, x2.shape, y.shape)
print(y)
break
# In[ ]:
all_model, feat_model, compare_model = create_simaese_model(
img_shape=IMAGE_SHAPE, mid_feat_dim=256, mob_alpha=0.75)
print(all_model.input_shape, all_model.output_shape)
all_model.summary()
# In[ ]:
# get generator and train
cb_list = get_callbacks('../data/checkpoints/mob_299_sim.h5', all_model)
adam_opt = keras.optimizers.Adam(lr=0.00005)
all_model.compile(optimizer=adam_opt,
loss='binary_crossentropy',
metrics=['acc'])
print('compile done')
# In[ ]:
# train
all_model.fit_generator(train_gen,
steps_per_epoch=1000,
epochs=100,
verbose=2,
callbacks=cb_list,
validation_data=val_gen,
# In[ ]:
all_model, feat_model, compare_model = create_simaese_model(
img_shape=IMAGE_SHAPE,
mid_feat_dim=512,
mob_alpha=0.75,
head_model_name='mobilenet_hbp',
l2_flag=True)
print(all_model.input_shape, all_model.output_shape)
all_model.summary()
# In[ ]:
# get generator and train
cb_list = get_callbacks('../data/checkpoints/mob_384_sim_hbp.h5', all_model)
adam_opt = keras.optimizers.Adam(lr=0.0003)
all_model.compile(optimizer=adam_opt,
loss='binary_crossentropy',
metrics=['acc'])
print('compile done')
# In[ ]:
# train
all_model.fit_generator(train_gen,
steps_per_epoch=1000,
epochs=100,
verbose=2,
callbacks=cb_list,
validation_data=val_gen,