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 )
# 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')(lay_in)
a = Dropout(p_drop=0.2)(a)
a = Dense(n_out=n_hid, act='relu')(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()
# Callbacks
dump_fd = 'train_on_batch_models'
if not os.path.exists(dump_fd): os.makedirs(dump_fd)
save_model = SaveModel(dump_fd=dump_fd, call_freq=200, type='iter')
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,
metrics=['categorical_error'],
call_freq=200,
type='iter')
callbacks = [save_model, validation]
def train():
# load data
batch_size = 128
tr_X, tr_y, va_X, va_y, te_X, te_y = pp_data.load_data()
n_batches = int(tr_X.shape[0] / batch_size)
# normalize data between [-1,1]
tr_X = (tr_X - 0.5) * 2
tr_X = tr_X.reshape((50000, 1, 28, 28))
print tr_X.shape
# generator
a0 = InputLayer(100)
a1 = Dense(128 * 7 * 7, act='linear')(a0)
a1 = BN(axis=0)(a1)
a1 = Reshape(out_shape=(128, 7, 7))(a1)
a1 = Convolution2D(64, 5, 5, act='linear', border_mode=(2, 2))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('leaky_relu')(a1)
a1 = UpSampling2D(size=(2, 2))(a1)
a1 = Convolution2D(32, 5, 5, act='linear', border_mode=(2, 2))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('leaky_relu')(a1)
a1 = UpSampling2D(size=(2, 2))(a1)
a8 = Convolution2D(1, 5, 5, act='tanh', border_mode=(2, 2), name='a8')(a1)
g = Model([a0], [a8])
g.compile()
g.summary()
# discriminator
b0 = InputLayer((1, 28, 28), name='b0')
b1 = Convolution2D(64, 5, 5, act='relu', border_mode=(0, 0), name='b1')(b0)
b1 = MaxPooling2D(pool_size=(2, 2))(b1)
b1 = Convolution2D(128, 5, 5, act='relu', border_mode=(0, 0))(b1)
b1 = MaxPooling2D(pool_size=(2, 2))(b1)
b1 = Flatten()(b1)
b8 = Dense(1, act='sigmoid')(b1)
d = Model([b0], [b8])
d.compile()
d.summary()
# discriminator on generator
d_on_g = Model()
d.set_trainability(False)
d_on_g.add_models([g, d])
d.set_trainability(True)
d_on_g.joint_models('a8', 'b0')
d_on_g.compile()
d_on_g.summary()
# optimizer
opt_d = Adam(1e-4)
opt_g = Adam(1e-4)
# optimization function
f_train_d = d.get_optimization_func(target_dims=[2],
loss_func='binary_crossentropy',
optimizer=opt_d,
clip=None)
f_train_g = d_on_g.get_optimization_func(target_dims=[2],
loss_func='binary_crossentropy',
optimizer=opt_g,
clip=None)
noise = np.zeros((batch_size, 100))
for epoch in range(100):
print epoch
for index in range(n_batches):
# concatenate generated img and real image to train discriminator.
noise = np.random.uniform(-1, 1, (batch_size, 100))
batch_x = tr_X[index * batch_size:(index + 1) * batch_size]
batch_gx = g.predict(noise)
batch_x_all = np.concatenate((batch_x, batch_gx))
# assign real img label as 1, generated img label as 0
batch_y_all = np.array([1] * batch_size + [0] * batch_size)
batch_y_all = batch_y_all.reshape((batch_y_all.shape[0], 1))
# save out generated img
if index % 50 == 0:
image = pp_data.combine_images(batch_gx)
image = image * 127.5 + 127.5
if not os.path.exists("img_dcgan"): os.makedirs("img_dcgan")
Image.fromarray(image.astype(
np.uint8)).save("img_dcgan/" + str(epoch) + "_" +
str(index) + ".png")
# train discriminator
d_loss = d.train_on_batch(f_train_d, batch_x_all, batch_y_all)
# assign generate img label as 1, so as to deceive discriminator
noise = np.random.uniform(-1, 1, (batch_size, 100))
batch_y_all = np.array([1] * batch_size)
batch_y_all = batch_y_all.reshape((batch_y_all.shape[0], 1))
# train generator
g_loss = d_on_g.train_on_batch(f_train_g, noise, batch_y_all)
print index, "d_loss:", d_loss, "\tg_loss:", g_loss
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.)
# observe backward
md.set_gt_nodes(target_dim_list=[2])
loss_node = obj.categorical_crossentropy(md.out_nodes_[0], md.gt_nodes_[0])
gparams = K.grad(loss_node + md.reg_value_, md.params_)
observe_nodes = gparams
f_backward = md.get_observe_backward_func(observe_nodes)
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(args):
workspace = cfg.workspace
te_fold = cfg.te_fold
n_events = args.n_events
snr = args.snr
feature_dir = os.path.join(workspace, "features", "logmel",
"n_events=%d" % n_events)
yaml_dir = os.path.join(workspace, "mixed_audio", "n_events=%d" % n_events)
(tr_x, tr_at_y, tr_sed_y, tr_na_list, te_x, te_at_y, te_sed_y,
te_na_list) = pp_data.load_data(feature_dir=feature_dir,
yaml_dir=yaml_dir,
te_fold=te_fold,
snr=snr,
is_scale=is_scale)
print(tr_x.shape, tr_at_y.shape)
print(te_x.shape, te_at_y.shape)
(_, n_time, n_freq) = tr_x.shape
n_out = len(cfg.events)
if False:
for e in tr_x:
plt.matshow(e.T, origin='lower', aspect='auto')
plt.show()
# Build model.
lay_in = InputLayer(in_shape=(n_time, n_freq))
a = Reshape((1, n_time, n_freq))(lay_in)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Dropout(p_drop=0.2)(a)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Dropout(p_drop=0.2)(a)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Conv2D(n_outfmaps=64,
n_row=3,
n_col=5,
act='linear',
strides=(1, 1),
border_mode=(1, 2))(a)
a = BN(axis=(0, 2, 3))(a)
a = Activation('relu')(a)
a = Dropout(p_drop=0.2)(a)
a = Conv2D(n_outfmaps=n_out,
n_row=1,
n_col=1,
act='sigmoid',
border_mode=(0, 0),
name='seg_masks')(a)
a8 = Lambda(_global_avg_pooling, name='a8')(a)
md = Model([lay_in], [a8])
md.compile()
md.summary(is_logging=True)
# Callbacks.
md_dir = os.path.join(workspace, "models", pp_data.get_filename(__file__),
"n_events=%d" % n_events, "fold=%d" % te_fold,
"snr=%d" % snr)
pp_data.create_folder(md_dir)
save_model = SaveModel(md_dir, call_freq=50, type='iter', is_logging=True)
validation = Validation(te_x=te_x,
te_y=te_at_y,
batch_size=50,
call_freq=50,
metrics=['binary_crossentropy'],
dump_path=None,
is_logging=True)
callbacks = [save_model, validation]
observe_nodes = [md.find_layer('seg_masks').output_]
f_forward = md.get_observe_forward_func(observe_nodes)
# Generator.
tr_gen = DataGenerator(batch_size=32, type='train')
eva_gen = DataGenerator2(batch_size=32, type='test')
# Train.
loss_ary = []
t1 = time.time()
optimizer = Adam(1e-3)
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_at_y]):
if md.iter_ % 50 == 0:
logging.info("iter: %d tr_loss: %f time: %s" % (
md.iter_,
np.mean(loss_ary),
time.time() - t1,
))
t1 = time.time()
loss_ary = []
# if md.iter_ % 200 == 0:
# write_out_at_sed(md, eva_gen, f_forward, te_x, te_at_y, te_sed_y, n_events, snr, te_fold)
if md.iter_ == 5001:
break
loss = md.train_on_batch(batch_x,
batch_y,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks)
loss_ary.append(loss)
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(args):
cpickle_dir = args.cpickle_dir
workspace = args.workspace
# Path of hdf5 data
bal_train_hdf5_path = os.path.join(cpickle_dir, "bal_train.h5")
unbal_train_hdf5_path = os.path.join(cpickle_dir, "unbal_train.h5")
eval_hdf5_path = os.path.join(cpickle_dir, "eval.h5")
# Load data
t1 = time.time()
(tr_x1, tr_y1, tr_id_list1) = pp_data.load_data(bal_train_hdf5_path)
(tr_x2, tr_y2, tr_id_list2) = pp_data.load_data(unbal_train_hdf5_path)
tr_x = np.concatenate((tr_x1, tr_x2))
tr_y = np.concatenate((tr_y1, tr_y2))
tr_id_list = tr_id_list1 + tr_id_list2
(te_x, te_y, te_id_list) = pp_data.load_data(eval_hdf5_path)
logging.info("Loading data time: %s s" % (time.time() - t1))
logging.info(tr_x1.shape, tr_x2.shape)
logging.info("tr_x.shape: %s" % (tr_x.shape, ))
(_, n_time, n_freq) = tr_x.shape
# Build model
n_hid = 500
n_out = tr_y.shape[1]
lay_in = InputLayer(in_shape=(n_time, n_freq))
a = Dense(n_out=n_hid, act='relu')(lay_in)
a = Dropout(p_drop=0.2)(a)
a = Dense(n_out=n_hid, act='relu')(a)
a = Dropout(p_drop=0.2)(a)
a = Dense(n_out=n_hid, act='relu')(a)
a = Dropout(p_drop=0.2)(a)
cla = Dense(n_out=n_out, act='sigmoid', name='cla')(a)
att = Dense(n_out=n_out, act='softmax', name='att')(a)
# Attention
lay_out = Lambda(_attention)([cla, att])
# Compile model
md = Model(in_layers=[lay_in], out_layers=[lay_out])
md.compile()
md.summary(is_logging=True)
# Save model every several iterations
call_freq = 1000
dump_fd = os.path.join(workspace, "models", pp_data.get_filename(__file__))
pp_data.create_folder(dump_fd)
save_model = SaveModel(dump_fd=dump_fd,
call_freq=call_freq,
type='iter',
is_logging=True)
# Callbacks function
callbacks = [save_model]
batch_size = 500
tr_gen = RatioDataGenerator(batch_size=batch_size, type='train')
# Optimization method
optimizer = Adam(lr=args.lr)
# Train
stat_dir = os.path.join(workspace, "stats", pp_data.get_filename(__file__))
pp_data.create_folder(stat_dir)
prob_dir = os.path.join(workspace, "probs", pp_data.get_filename(__file__))
pp_data.create_folder(prob_dir)
tr_time = time.time()
for (tr_batch_x, tr_batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
# Compute stats every several interations
if md.iter_ % call_freq == 0:
# Stats of evaluation dataset
t1 = time.time()
te_err = eval(md=md,
x=te_x,
y=te_y,
out_dir=os.path.join(stat_dir, "test"),
out_probs_dir=os.path.join(prob_dir, "test"))
logging.info("Evaluate test time: %s" % (time.time() - t1, ))
# Stats of training dataset
t1 = time.time()
tr_bal_err = eval(md=md,
x=tr_x1,
y=tr_y1,
out_dir=os.path.join(stat_dir, "train_bal"),
out_probs_dir=None)
logging.info("Evaluate tr_bal time: %s" % (time.time() - t1, ))
# Update params
(tr_batch_x,
tr_batch_y) = pp_data.transform_data(tr_batch_x, tr_batch_y)
md.train_on_batch(batch_x=tr_batch_x,
batch_y=tr_batch_y,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks)
# Stop training when maximum iteration achieves
if md.iter_ == call_freq * 31:
break
def train(args):
workspace = args.workspace
cla_mapping = args.cla_mapping
# Load data.
t1 = time.time()
tr_pack_path = os.path.join(workspace, "packed_features", "logmel",
"training.h5")
te_pack_path = os.path.join(workspace, "packed_features", "logmel",
"testing.h5")
with h5py.File(tr_pack_path, 'r') as hf:
tr_na_list = list(hf.get('na_list'))
tr_x = np.array(hf.get('x'))
tr_y = np.array(hf.get('y'))
with h5py.File(te_pack_path, 'r') as hf:
te_na_list = list(hf.get('na_list'))
te_x = np.array(hf.get('x'))
te_y = np.array(hf.get('y'))
logging.info("Loading data time: %s" % (time.time() - t1, ))
# Scale.
t1 = time.time()
scaler_path = os.path.join(workspace, "scalers", "logmel",
"training.scaler")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.do_scaler_on_x3d(tr_x, scaler)
te_x = pp_data.do_scaler_on_x3d(te_x, scaler)
logging.info("Scale time: %s" % (time.time() - t1, ))
logging.info("tr_x: %s %s" % (tr_x.shape, tr_x.dtype))
logging.info("tr_y: %s %s" % (tr_y.shape, tr_y.dtype))
logging.info("y: 1-of-4 representation: %s" % (cfg.events + ['bg'], ))
# Build model.
(_, n_time, n_freq) = tr_x.shape
n_out = len(cfg.events) + 1
in0 = InputLayer(in_shape=(n_time, n_freq))
a1 = Reshape((1, n_time, n_freq))(in0)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Dropout(0.3)(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Dropout(0.3)(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Dropout(0.3)(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Conv2D(n_outfmaps=64,
n_row=3,
n_col=3,
act='linear',
border_mode=(1, 1))(a1)
a1 = BN(axis=(0, 2, 3))(a1)
a1 = Activation('relu')(a1)
a1 = Dropout(0.3)(a1)
# Segmentation mask for 'babycry', 'glassbreak' and 'gunshot'.
a1 = Conv2D(n_outfmaps=len(cfg.events),
n_row=1,
n_col=1,
act='sigmoid',
border_mode=(0, 0))(a1)
# Extend segmentation mask to 'babycry', 'glassbreak', 'gunshot' and 'background'.
a1 = Lambda(_seg_mask_ext_bg, name='seg_masks')(a1)
# Classification mapping.
cla_mapping = args.cla_mapping
if cla_mapping == 'global_rank_pooling':
weight1d = np.power(r * np.ones(120 * 64), np.arange(120 * 64))
a8 = Lambda(_global_rank_pooling, weight1d=weight1d, name='a5')(a1)
elif cla_mapping == 'global_max_pooling':
a8 = Lambda(_global_max_pooling)(a1)
elif cla_mapping == 'global_avg_pooling':
a8 = Lambda(_global_avg_pooling)(a1)
else:
raise Exception("Incorrect cla_mapping!")
md = Model([in0], [a8])
md.compile()
md.summary(is_logging=True)
# Callbacks.
md_dir = os.path.join(workspace, "models", pp_data.get_filename(__file__))
pp_data.create_folder(md_dir)
save_model = SaveModel(md_dir, call_freq=100, type='iter')
validation = Validation(te_x=te_x,
te_y=te_y,
batch_size=100,
call_freq=50,
metrics=['binary_crossentropy'],
dump_path=None,
is_logging=True)
callbacks = [save_model, validation]
# Train.
generator = DataGenerator(batch_size=20, type='train')
loss_ary = []
t1 = time.time()
optimizer = Adam(1e-4)
for (batch_x, batch_y) in generator.generate(xs=[tr_x], ys=[tr_y]):
np.set_printoptions(threshold=np.nan,
linewidth=1000,
precision=2,
suppress=True)
loss = md.train_on_batch(batch_x,
batch_y,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks)
loss_ary.append(loss)
if md.iter_ % 50 == 0: # Evalute training loss every several iterations.
logging.info("iter: %d, tr loss: %d" %
(md.iter_, np.mean(loss_ary)))
logging.info("time: %s" % (time.time() - t1, ))
t1 = time.time()
loss_ary = []
if md.iter_ == 10001: # Stop after several iterations.
break
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)