def train():
# prepare data
tr_X, tr_y, _, te_X, te_y, te_na_list = pp_dev_data.GetAllData(
fe_fd, agg_num, hop, fold)
[batch_num, n_time, n_freq] = tr_X.shape
print tr_X.shape, tr_y.shape
print te_X.shape, te_y.shape
# build model
seq = Sequential()
seq.add(InputLayer((n_time, n_freq)))
seq.add(Flatten()) # flatten to 2d: (n_time, n_freq) to 1d:(n_time*n_freq)
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_out, act='sigmoid'))
md = seq.compile()
md.summary()
# optimizer
optimizer = Adam(1e-4)
# callbacks
# tr_err, te_err are frame based. To get event based err, run recognize.py
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=None,
va_y=None,
te_x=te_X,
te_y=te_y,
batch_size=2000,
metrics=['binary_crossentropy'],
call_freq=1,
dump_path=None)
# save model
pp_dev_data.CreateFolder(cfg.dev_md_fd)
save_model = SaveModel(dump_fd=cfg.dev_md_fd, call_freq=10)
# callbacks
callbacks = [validation, save_model]
# fit model
md.fit(x=tr_X,
y=tr_y,
batch_size=2000,
n_epochs=100,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks,
verbose=1)
def train():
# load data
tr_X2d_mix, tr_X3d_mix, tr_y2d_chn0, tr_y2d_chn1, tr_y3d_chn0, tr_y3d_chn1 = pp_data.LoadData(
cfg.fe_fft_fd, n_time, hop, na_list=cfg.tr_list)
# build model
lay_in0 = InputLayer(in_shape=((n_time, n_freq)), name='in1')
lay_a0 = Flatten()(lay_in0)
lay_a1 = Dense(n_hid, act='relu', name='a1')(lay_a0)
lay_a2 = Dropout(0.2, name='a2')(lay_a1)
lay_a3 = Dense(n_hid, act='relu', name='a3')(lay_a2)
lay_a4 = Dropout(0.2, name='a4')(lay_a3)
lay_a5 = Dense(n_hid, act='relu', name='a5')(lay_a4)
lay_a6 = Dropout(0.2, name='a6')(lay_a5)
lay_b1 = Dense(n_freq, act='sigmoid',
name='a7')(lay_a6) # mask_left, shape: (N, n_freq)
lay_c1 = Dense(n_freq, act='sigmoid',
name='a8')(lay_a6) # mask_right, shape: (N, n_freq)
lay_out_b = Lambda(mul,
name='out_b')([lay_b1,
lay_in0]) # out_left, shape: (N, n_freq)
lay_out_c = Lambda(mul, name='out_c')([lay_c1, lay_in0
]) # out_right, shape: (N, n_freq)
md = Model(in_layers=[lay_in0],
out_layers=[lay_out_b, lay_out_c],
any_layers=[lay_in0, lay_b1, lay_c1])
md.summary()
# validation
validation = Validation(tr_x=[np.abs(tr_y3d_chn0) + np.abs(tr_y3d_chn1)],
tr_y=[np.abs(tr_y2d_chn0),
np.abs(tr_y2d_chn1)],
batch_size=100,
metrics=[loss_func],
call_freq=1,
dump_path=None)
# save model
if not os.path.exists(cfg.md_fd): os.makedirs(cfg.md_fd)
save_model = SaveModel(dump_fd=cfg.md_fd, call_freq=2)
# callbacks
callbacks = [validation, save_model]
# optimizer
optimizer = Adam(1e-3)
# fit model
md.fit( [np.abs(tr_y3d_chn0)+np.abs(tr_y3d_chn1)], [np.abs(tr_y2d_chn0), np.abs(tr_y2d_chn1)], \
batch_size=100, n_epochs=100, loss_func='mse', optimizer=optimizer, callbacks=callbacks, verbose=1 )
n_out = 10 # sparse label to 1 of K categorical label tr_y = sparse_to_categorical(tr_y, n_out) va_y = sparse_to_categorical(va_y, n_out) te_y = sparse_to_categorical(te_y, n_out) ### Build model act = 'relu' seq = Sequential() seq.add(InputLayer(in_shape=(1, 28, 28))) seq.add(Convolution2D(n_outfmaps=32, n_row=3, n_col=3, act='relu')) seq.add(MaxPool2D(pool_size=(2, 2))) seq.add(Convolution2D(n_outfmaps=32, n_row=3, n_col=3, act='relu')) seq.add(MaxPool2D(pool_size=(2, 2))) seq.add(Dropout(0.2)) seq.add(Flatten()) seq.add(Dense(n_hid, act='relu')) seq.add(Dropout(0.5)) seq.add(Dense(n_hid, act='relu')) seq.add(Dense(n_out, act='softmax')) md = seq.combine() # print summary info of model md.summary() # optimization method optimizer = Adam(lr=0.001) ### callbacks (optional) # save model every n epoch (optional)
# init params
n_in = 784
n_hid = 500
n_out = 10
# sparse label to 1-of-K categorical label
tr_y = sparse_to_categorical(tr_y, n_out)
va_y = sparse_to_categorical(va_y, n_out)
te_y = sparse_to_categorical(te_y, n_out)
### Build model
lay_in = InputLayer(in_shape=(n_in, ))
a = Dense(n_out=n_hid, act='relu', name='dense1')(lay_in)
a = Dropout(p_drop=0.2)(a)
a = Dense(n_out=n_hid, act='relu', name='dense2')(a)
a = Dropout(p_drop=0.2)(a)
lay_out = Dense(n_out=n_out, act='softmax')(a)
md = Model(in_layers=[lay_in], out_layers=[lay_out])
md.compile()
md.summary()
# observe forward
observe_nodes = [
md.find_layer('dense1').output_,
md.find_layer('dense2').output_
]
f_forward = md.get_observe_forward_func(observe_nodes)
print md.run_function(func=f_forward, z=[te_x], batch_size=500, tr_phase=0.)
fold = 1
n_out = len(cfg.labels)
# prepare data
tr_X, tr_y, _ = pp_dev_data.GetAllData(cfg.dev_fe_mel_fd,
agg_num,
hop,
fold=None)
[batch_num, n_time, n_freq] = tr_X.shape
print tr_X.shape, tr_y.shape
# build model
seq = Sequential()
seq.add(InputLayer((n_time, n_freq)))
seq.add(Flatten()) # flatten to 2d: (n_time, n_freq) to 1d:(n_time*n_freq)
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_out, act='sigmoid'))
md = seq.compile()
md.summary()
# validation
# tr_err, te_err are frame based. To get event based err, run recognize.py
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=None,
def train():
# create empty folders in workspace
create_folders()
# get dev & eva data
tr_X, tr_y, _, _, _, _ = pp_dev_data.GetSegData(dev_fe_fd,
agg_num,
hop,
fold=None)
te_X, te_na_list = pp_eva_data.GetEvaSegData(eva_fe_fd, agg_num, hop)
[n_songs, n_chunks, _, n_in] = tr_X.shape
print tr_X.shape, tr_y.shape
print te_X.shape
# model
lay_in0 = InputLayer(
(n_chunks, agg_num, n_in),
name='in0') # shape: (n_songs, n_chunk, agg_num, n_in)
lay_a1 = Flatten(3, name='a1')(
lay_in0) # shape: (n_songs, n_chunk, agg_num*n_in)
lay_a2 = Dense(n_hid,
act='relu')(lay_a1) # shape: (n_songs, n_chunk, n_hid)
lay_a3 = Dropout(0.2)(lay_a2)
lay_a4 = Dense(n_hid, act='relu')(lay_a3)
lay_a5 = Dropout(0.2)(lay_a4)
lay_a6 = Dense(n_hid, act='relu')(lay_a5)
lay_a7 = Dropout(0.2)(lay_a6)
lay_a8 = Dense(n_out, act='sigmoid', b_init=-1,
name='a8')(lay_a7) # shape: (n_songs, n_chunk, n_out)
md = Model(in_layers=[lay_in0], out_layers=[lay_a8], any_layers=[])
md.compile()
md.summary()
# callback, write out dection scores to .txt each epoch
dump_fd = cfg.scrap_fd + '/Results_eva/bob_eer'
print_scores = cb_eer.PrintScoresBagOfBlocks(te_X,
te_na_list,
dump_fd,
call_freq=1)
# callback, print loss each epoch
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=None,
va_y=None,
te_x=None,
te_y=None,
metrics=[loss_func],
call_freq=1,
dump_path=None)
# callback, save model every N epochs
save_model = SaveModel(dump_fd=cfg.scrap_fd + '/Md_eva_bob', call_freq=10)
# combine all callbacks
callbacks = [validation, save_model, print_scores]
# optimizer
optimizer = Adam(2e-4)
# fit model
md.fit(x=tr_X,
y=tr_y,
batch_size=10,
n_epochs=301,
loss_func=loss_func,
optimizer=optimizer,
callbacks=callbacks)
def train_cv_model():
# init path
if type == 'home':
fe_fd = cfg.dev_fe_mel_home_fd
labels = cfg.labels_home
lb_to_id = cfg.lb_to_id_home
tr_txt = cfg.dev_evaluation_fd + '/home_fold' + str(
fold) + '_train.txt'
te_txt = cfg.dev_evaluation_fd + '/home_fold' + str(
fold) + '_evaluate.txt'
if type == 'resi':
fe_fd = cfg.dev_fe_mel_resi_fd
labels = cfg.labels_resi
lb_to_id = cfg.lb_to_id_resi
tr_txt = cfg.dev_evaluation_fd + '/residential_area_fold' + str(
fold) + '_train.txt'
te_txt = cfg.dev_evaluation_fd + '/residential_area_fold' + str(
fold) + '_evaluate.txt'
n_out = len(labels)
# load data to list
tr_X, tr_y = pp_dev_data.LoadAllData(fe_fd, tr_txt, lb_to_id, agg_num, hop)
tr_y = sparse_to_categorical(tr_y, n_out)
print tr_X.shape
print tr_y.shape
n_freq = tr_X.shape[2]
# build model
seq = Sequential()
seq.add(InputLayer((agg_num, n_freq)))
seq.add(Flatten())
seq.add(Dense(n_hid, act='relu'))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act='relu'))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act='relu'))
seq.add(Dropout(0.1))
seq.add(Dense(n_out, 'sigmoid'))
md = seq.combine()
# print summary info of model
md.summary()
# optimization method
optimizer = Adam(1e-3)
# callbacks (optional)
# save model every n epoch
pp_dev_data.CreateFolder(cfg.dev_md_fd)
save_model = SaveModel(dump_fd=cfg.dev_md_fd, call_freq=5)
# validate model every n epoch
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=None,
va_y=None,
te_x=None,
te_y=None,
metrics=['binary_crossentropy'],
call_freq=1,
dump_path=None)
# callbacks function
callbacks = [validation, save_model]
# train model
md.fit(x=tr_X,
y=tr_y,
batch_size=20,
n_epochs=100,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks)
# init params
n_in = 784
n_hid = 500
n_out = 10
# sparse label to 1-of-K categorical label
tr_y = sparse_to_categorical(tr_y, n_out)
va_y = sparse_to_categorical(va_y, n_out)
te_y = sparse_to_categorical(te_y, n_out)
### Build model
seq = Sequential()
seq.add(InputLayer(n_in))
seq.add(Dense(n_hid, act='relu'))
seq.add(Dropout(p_drop=0.2))
seq.add(Dense(n_hid, act='relu'))
seq.add(Dropout(p_drop=0.2))
seq.add(Dense(n_out, act='softmax'))
md = seq.combine()
# print summary info of model
md.summary()
# md.plot_connection()
### optimization method
optimizer = Adam(lr=0.001) # Try SGD, Adagrad, Rmsprop, etc. instead
### callbacks (optional)
# save model every n epoch (optional)
if not os.path.exists('Md'): os.makedirs('Md')
def train():
_loss_func = _jdc_loss_func0
# load data
t1 = time.time()
dict = cPickle.load( open( cfg.scrap_fd+'/denoise_enhance_pool_fft_all0.p', 'rb' ) )
tr_X, tr_mask, tr_y, tr_na_list, te_X, te_mask, te_y, te_na_list = dict['tr_X'], dict['tr_mask'], dict['tr_y'], dict['tr_na_list'], dict['te_X'], dict['te_mask'], dict['te_y'], dict['te_na_list']
t2 = time.time()
tr_X = pp_data.wipe_click( tr_X, tr_na_list )
te_X = pp_data.wipe_click( te_X, te_na_list )
# balance data
tr_X, tr_mask, tr_y = pp_data.BalanceData2( tr_X, tr_mask, tr_y )
te_X, te_mask, te_y = pp_data.BalanceData2( te_X, te_mask, te_y )
print tr_X.shape, tr_y.shape, te_X.shape, te_y.shape
[n_songs, n_chunks, n_freq] = te_X.shape
tr_y = tr_y.reshape( (len(tr_y), 1) )
te_y = te_y.reshape( (len(te_y), 1) )
# jdc model
# classifier
lay_z0 = InputLayer( (n_chunks,) ) # shape:(n_songs, n_chunks) keep the length of songs
lay_in0 = InputLayer( (n_chunks, n_freq), name='in0' ) # shape: (n_songs, n_chunk, n_freq)
lay_a1 = lay_in0
# lay_a1 = Lambda( _conv2d )( lay_a1 )
lay_a1 = Lambda( _reshape_3d_to_4d )( lay_a1 )
lay_a1 = Convolution2D( 32, 3, 3, act='relu', init_type='glorot_uniform', border_mode=(1,1), strides=(1,1), name='a11' )( lay_a1 )
lay_a1 = Dropout( 0.2 )( lay_a1 )
lay_a1 = MaxPool2D( pool_size=(1,2) )( lay_a1 )
lay_a1 = Convolution2D( 64, 3, 3, act='relu', init_type='glorot_uniform', border_mode=(1,1), strides=(1,1), name='a12' )( lay_a1 )
lay_a1 = Dropout( 0.2 )( lay_a1 )
lay_a1 = MaxPool2D( pool_size=(1,2) )( lay_a1 )
lay_a1 = Lambda( _reshape_4d_to_3d )( lay_a1 )
lay_a1 = Dense( n_hid, act='relu', name='a2' )( lay_a1 ) # shape: (n_songs, n_chunk, n_hid)
lay_a1 = Dropout( 0.2 )( lay_a1 )
lay_a1 = Dense( n_hid, act='relu', name='a4' )( lay_a1 )
lay_a1 = Dropout( 0.2 )( lay_a1 )
lay_a1 = Dense( n_hid, act='relu', name='a6' )( lay_a1 )
lay_a1 = Dropout( 0.2 )( lay_a1 )
lay_a8 = Dense( n_out, act='sigmoid', init_type='zeros', b_init=0, name='a8' )( lay_a1 ) # shape: (n_songs, n_chunk, n_out)
# detector
lay_b1 = lay_in0 # shape: (n_songs, n_chunk, n_freq)
lay_b2 = Lambda( _conv2d )( lay_b1 ) # shape: (n_songs, n_chunk, n_freq)
lay_b2 = Lambda( _reshape_3d_to_4d )( lay_b1 )
lay_b2 = MaxPool2D( pool_size=(1,2) )( lay_b2 )
lay_b2 = Lambda( _reshape_4d_to_3d )( lay_b2 )
lay_b8 = Dense( n_out, act='hard_sigmoid', init_type='zeros', b_init=-2.3, name='b8' )( lay_b2 )
md = Model( in_layers=[lay_in0, lay_z0], out_layers=[lay_a8, lay_b8], any_layers=[] )
# print summary info of model
md.summary()
# callbacks (optional)
# save model every n epoch (optional)
pp_data.CreateFolder( cfg.wbl_dev_md_fd )
pp_data.CreateFolder( cfg.wbl_dev_md_fd+'/cnn_fft' )
save_model = SaveModel( dump_fd=cfg.wbl_dev_md_fd+'/cnn_fft', call_freq=20, type='iter' )
validation = Validation( tr_x=None, tr_y=None, va_x=None, va_y=None, te_x=[te_X, te_mask], te_y=te_y, batch_size=100, metrics=[_loss_func], call_freq=20, dump_path=None, type='iter' )
# callbacks function
callbacks = [save_model, validation]
# EM training
md.set_gt_nodes( tr_y )
md.find_layer('a11').set_trainable_params( ['W','b'] )
md.find_layer('a12').set_trainable_params( ['W','b'] )
md.find_layer('a2').set_trainable_params( ['W','b'] )
md.find_layer('a4').set_trainable_params( ['W','b'] )
md.find_layer('a6').set_trainable_params( ['W','b'] )
md.find_layer('a8').set_trainable_params( ['W','b'] )
md.find_layer('b8').set_trainable_params( [] )
opt_classifier = Adam( 1e-3 )
f_classify = md.get_optimization_func( loss_func=_loss_func, optimizer=opt_classifier, clip=None )
md.find_layer('a11').set_trainable_params( [] )
md.find_layer('a12').set_trainable_params( [] )
md.find_layer('a2').set_trainable_params( [] )
md.find_layer('a4').set_trainable_params( [] )
md.find_layer('a6').set_trainable_params( [] )
md.find_layer('a8').set_trainable_params( [] )
md.find_layer('b8').set_trainable_params( ['W','b'] )
opt_detector = Adam( 1e-3 )
f_detector = md.get_optimization_func( loss_func=_loss_func, optimizer=opt_detector, clip=None )
_x, _y = md.preprocess_data( [tr_X, tr_mask], tr_y, shuffle=True )
for i1 in xrange(500):
print '-----------------------'
opt_classifier.reset()
md.do_optimization_func_iter_wise( f_classify, _x, _y, batch_size=100, n_iters=80, callbacks=callbacks, verbose=1 )
print '-----------------------'
opt_detector.reset()
md.do_optimization_func_iter_wise( f_detector, _x, _y, batch_size=100, n_iters=20, callbacks=callbacks, verbose=1 )
def train():
# create empty folders in workspace
create_folders()
# prepare data
tr_X, tr_y, _, va_X, va_y, va_na_list = pp_dev_data.GetSegData(
fe_fd, agg_num, hop, fold)
[n_songs, n_chunks, _, n_in] = tr_X.shape
print tr_X.shape, tr_y.shape
print va_X.shape, va_y.shape
# model
# classifier
lay_in0 = InputLayer(
(n_chunks, agg_num, n_in),
name='in0') # shape: (n_songs, n_chunk, agg_num, n_in)
lay_a1 = Flatten(3, name='a1')(
lay_in0) # shape: (n_songs, n_chunk, agg_num*n_in)
lay_a2 = Dense(n_hid,
act='relu')(lay_a1) # shape: (n_songs, n_chunk, n_hid)
lay_a3 = Dropout(0.2)(lay_a2)
lay_a4 = Dense(n_hid, act='relu')(lay_a3)
lay_a5 = Dropout(0.2)(lay_a4)
lay_a6 = Dense(n_hid, act='relu')(lay_a5)
lay_a7 = Dropout(0.2)(lay_a6)
lay_a8 = Dense(n_out, act='sigmoid', b_init=-1,
name='a8')(lay_a7) # shape: (n_songs, n_chunk, n_out)
# detector
lay_b1 = Lambda(mean_pool)(lay_in0) # shape: (n_songs, n_chunk, n_out)
lay_b8 = Dense(n_out, act='sigmoid',
name='b4')(lay_b1) # shape: (n_songs, n_chunk, n_out)
md = Model(in_layers=[lay_in0], out_layers=[lay_a8, lay_b8], any_layers=[])
md.compile()
md.summary()
# callback, write out dection scores to .txt each epoch
dump_fd = cfg.scrap_fd + '/Results_dev/jdc_eer/fold' + str(fold)
print_scores = cb_eer.PrintScoresDetectionClassification(va_X,
va_na_list,
dump_fd,
call_freq=1)
# callback, print loss each epoch
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=va_X,
va_y=va_y,
te_x=None,
te_y=None,
metrics=[loss_func],
call_freq=1,
dump_path=None)
# callback, save model every N epochs
save_model = SaveModel(dump_fd=cfg.scrap_fd + '/Md_dev_jdc', call_freq=10)
# combine all callbacks
callbacks = [validation, save_model, print_scores]
# optimizer
# optimizer = SGD( 0.01, 0.95 )
optimizer = Adam(2e-4)
# fit model
md.fit(x=tr_X,
y=tr_y,
batch_size=10,
n_epochs=301,
loss_func=loss_func,
optimizer=optimizer,
callbacks=callbacks)
def train(tr_fe_fd, tr_csv_file, te_fe_fd, te_csv_file, n_concat, hop, scaler,
out_md_fd):
# Prepare data
tr_x, tr_y = pp_data.get_matrix_format_data(fe_fd=tr_fe_fd,
csv_file=tr_csv_file,
n_concat=n_concat,
hop=hop,
scaler=scaler)
te_x, te_y = pp_data.get_matrix_format_data(fe_fd=te_fe_fd,
csv_file=te_csv_file,
n_concat=n_concat,
hop=hop,
scaler=scaler)
n_freq = tr_x.shape[2]
print 'tr_x.shape:', tr_x.shape # (n_samples, n_concat, n_freq)
print 'tr_y.shape:', tr_y.shape # (n_samples, n_labels)
# Build model
n_out = len(cfg.labels)
seq = Sequential()
seq.add(InputLayer((n_concat, n_freq)))
seq.add(Flatten())
seq.add(Dropout(0.2))
seq.add(Dense(200, act='relu'))
seq.add(Dropout(0.2))
seq.add(Dense(200, act='relu'))
seq.add(Dropout(0.2))
seq.add(Dense(n_out, act='softmax'))
md = seq.compile()
md.summary()
# Validation.
# tr_err, te_err are frame based. To get event based err, run recognize.py
validation = Validation(tr_x=tr_x,
tr_y=tr_y,
va_x=None,
va_y=None,
te_x=te_x,
te_y=te_y,
batch_size=500,
call_freq=1,
dump_path=None)
# Save model
pp_data.create_folder(out_md_fd)
save_model = SaveModel(out_md_fd, call_freq=2)
# Callbacks
callbacks = [validation, save_model]
# Optimizer
optimizer = Adam(1e-3)
# fit model
md.fit(x=tr_x,
y=tr_y,
batch_size=100,
n_epochs=101,
loss_func='categorical_crossentropy',
optimizer=optimizer,
callbacks=callbacks)
