def recognize(args, at_bool, sed_bool):
(te_x, te_y, te_na_list) = load_hdf5_data(args.te_hdf5_path, verbose=1)
x = te_x
y = te_y
na_list = te_na_list
x = do_scale(x, args.scaler_path, verbose=1)
fusion_at_list = []
fusion_sed_list = []
for epoch in range(20, 30, 1):
t1 = time.time()
print(epoch)
model_p = "gatedAct_rationBal44_lr0.001_normalization_at_cnnRNN_64newMel_240fr.06-0.8333.hdf5"
#[model_path] = glob.glob(os.path.join(args.model_dir,
# "*.%02d-0.*.hdf5" % epoch))
[model_path] = glob.glob(os.path.join(args.model_dir, model_p))
model = load_model(model_path)
# Audio tagging
if at_bool:
pred = model.predict(x)
fusion_at_list.append(pred)
# Sound event detection
if sed_bool:
in_layer = model.get_layer('in_layer')
loc_layer = model.get_layer('localization_layer')
func = K.function(
[in_layer.input, K.learning_phase()], [loc_layer.output])
pred3d = run_func(func, x, batch_size=20)
fusion_sed_list.append(pred3d)
print("Prediction time: %s" % (time.time() - t1, ))
# Write out AT probabilities
if at_bool:
fusion_at = np.mean(np.array(fusion_at_list), axis=0)
print("AT shape: %s" % (fusion_at.shape, ))
io_task4.at_write_prob_mat_to_csv(na_list=na_list,
prob_mat=fusion_at,
out_path=os.path.join(
args.out_dir,
"at_prob_mat.csv.gz"))
# Write out SED probabilites
if sed_bool:
fusion_sed = np.mean(np.array(fusion_sed_list), axis=0)
print("SED shape:%s" % (fusion_sed.shape, ))
io_task4.sed_write_prob_mat_list_to_csv(
na_list=na_list,
prob_mat_list=fusion_sed,
out_path=os.path.join(args.out_dir, "sed_prob_mat_list.csv.gz"))
print("Prediction finished!")
def recognize(args, at_bool, sed_bool):
(te_x, _, te_na_list) = load_hdf5_data(args.hdf5_path, verbose=1)
x = te_x
na_list = te_na_list
# x[:, :64] = do_scale(x[:, :64], args.scaler_path, verbose=1)
fusion_at_list = []
fusion_sed_list = []
for epoch in range(25, 30, 1):
t1 = time.time()
file_name = os.path.join(args.model_dir, "*.%02d-0.*.hdf5" % epoch)
model_path = glob.glob(file_name)[0] # returns more than one item so can't unpack
model = load_model(model_path)
# Audio tagging
if at_bool:
pred = model.predict(x, batch_size=5)
fusion_at_list.append(pred)
# Sound event detection
if sed_bool:
in_layer = model.get_layer('in_layer')
loc_layer = model.get_layer('localization_layer')
func = K.function([in_layer.input, K.learning_phase()],
[loc_layer.output])
pred3d = run_func(func, x, batch_size=20)
fusion_sed_list.append(pred3d)
print("Prediction time: %s" % (time.time() - t1,))
# Write out AT probabilities
if at_bool:
fusion_at = np.mean(np.array(fusion_at_list), axis=0)
print("AT shape: %s" % (fusion_at.shape,))
io_task4.at_write_prob_mat_to_csv(
na_list=na_list,
prob_mat=fusion_at,
out_path=args.out_dir) # "at_audio_prob_mat.csv.gz"
# Write out SED probabilites
if sed_bool:
fusion_sed = np.mean(np.array(fusion_sed_list), axis=0)
print("SED shape:%s" % (fusion_sed.shape,))
io_task4.sed_write_prob_mat_list_to_csv(
na_list=na_list,
prob_mat_list=fusion_sed,
out_path=os.path.join(args.out_dir, "sed_prob_mat_list.csv.gz"))
print("Prediction finished!")
def train(args):
num_classes = cfg.num_classes
# Load training & testing data
(tr_x, tr_y, tr_na_list) = load_hdf5_data(args.tr_hdf5_path, verbose=1)
(te_x, te_y, te_na_list) = load_hdf5_data(args.te_hdf5_path, verbose=1)
print("tr_x.shape: %s" % (tr_x.shape,))#removed this dec4 since its not helpful really
# Scale data
tr_x = do_scale(tr_x, args.scaler_path, verbose=1)
te_x = do_scale(te_x, args.scaler_path, verbose=1)
#print("delme dec 1, tr_x.shape", tr_x.shape)#output=51, 240, 64
#print("delme dec 1, te_x.shape", te_x.shape)#:51, 240, 64
# Build model
(_, n_time, n_freq) = tr_x.shape # (N, 240, 64)
input_logmel = Input(shape=(n_time, n_freq), name='in_layer') # (N, 240, 64)
a1 = Reshape((n_time, n_freq, 1))(input_logmel) # (N, 240, 64, 1)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 32, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 16, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 8, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 4, 128)
a1 = Conv2D(256, (3, 3), padding="same", activation="relu", use_bias=True)(a1)
a1 = MaxPooling2D(pool_size=(1, 4))(a1) # (N, 240, 1, 256)
a1 = Reshape((240, 256))(a1) # (N, 240, 256)
# Gated BGRU
rnnout = Bidirectional(GRU(128, activation='linear', return_sequences=True))(a1)
rnnout_gate = Bidirectional(GRU(128, activation='sigmoid', return_sequences=True))(a1)
a2 = Multiply()([rnnout, rnnout_gate])
# Attention
cla = TimeDistributed(Dense(num_classes, activation='sigmoid'), name='localization_layer')(a2)
att = TimeDistributed(Dense(num_classes, activation='softmax'))(a2)
out = Lambda(outfunc, output_shape=(num_classes,))([cla, att])
model = Model(input_logmel, out)
model.summary()
adam_optimizer = keras.optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)
# Compile model
model.compile(loss='binary_crossentropy',
optimizer=adam_optimizer,
metrics=['accuracy'])#finn delme dec1 you can change this to categorical_accuracy to see if you can subvert the keras error. However you dont know if its the right hting to do. Keep a look out at the results to determineif it did what you wanted
# Save model callback
print("working here 1")
filepath = os.path.join(args.out_model_dir, "{0}_{1}.hdf5".format(args.model_name, args.epochs))
create_folder(os.path.dirname(filepath))
save_model = ModelCheckpoint(filepath=filepath,
monitor='val_acc',
verbose=0,
save_best_only=False,
save_weights_only=False,
mode='auto',
period=1)
# Data generator
# Train
t_train = time.time()
print("FINN Training started")#this is really just me seeing if this is where most of the time is spent
use_generator = False
if use_generator:
gen = RatioDataGenerator(batch_size=args.batch_size, type='train')#batch size should be manipulated from 44
model.fit_generator(generator=gen.generate([tr_x], [tr_y]),
steps_per_epoch=args.steps_p_epoch, # 100 iters is called an 'epoch'
epochs=args.epochs, #31 # Maximum 'epoch' to train
verbose=1,
callbacks=[save_model],
validation_data=(te_x, te_y))
else:
model.fit(x=tr_x, y=tr_y, batch_size=20, epochs=args.epochs, verbose=1, callbacks=[save_model], validation_split=0.05, shuffle=True, class_weight=None, sample_weight=None, initial_epoch=args.init_epoch, steps_per_epoch=None, validation_steps=None)
model.save(os.path.join(args.out_model_dir, "final_model_{}_{}epochs.h5".format(args.model_name, args.epochs)))#am not sure if fit will save the final epoch.. pretty sure it does tho
print("FINN Training finished, time taken: ", (time.time()-t_train))#this is really just me seeing if this is where most of the time is spent
def recognize(args, at_bool, sed_bool):
t_rec = time.time()
print("FINN Recognize started")
#print("recognize!")#todo remove
(te_x, te_y, te_na_list) = load_hdf5_data(args.te_hdf5_path, verbose=1)
x = te_x
y = te_y
na_list = te_na_list
#print("delme dec 1, num_classes", cfg.num_classes)#num_classes 3
x = do_scale(x, args.scaler_path, verbose=1)
fusion_at_list = []
fusion_sed_list = []
#for epoch in range(20, 30, 1):#hmm this value might need to be changed depending on nmber of epochs in model.fit_generator(... although this isnt a train session so maybe not!
for epoch in range(1, args.epochs+1):#chane this when you want to increase epochs and reduce the amount of epochs pred'd over eg: range(20, epochs, 1)# range(1, args.epochs+1) is for args.epochs epochs btw
#this allows you to go over all the epochs and average the scroe. but why?
t1 = time.time()
model_path = os.path.join(args.model_dir, "{0}_{1}.hdf5".format(args.model_name, epoch))
print("model_path", model_path)
model = load_model(model_path)
# Audio tagging
if at_bool:
#dec4 this is where we find that shapes change from y having (#, n_classes) to (#, 2)... Dont know how to fix but i think its in changing the keras model
#x.shape delme (22, 240, 64)
#pred.shape delme (22, 2)
t_pred = time.time()
print("FINN at_pred started")
pred = model.predict(x)
print("FINN at_pred ended", (time.time()-t_pred))
fusion_at_list.append(pred)
# Sound event detection
if sed_bool:
t_pred = time.time()
print("FINN pred_sed started")
in_layer = model.get_layer('in_layer')
loc_layer = model.get_layer('localization_layer')
func = K.function([in_layer.input, K.learning_phase()],
[loc_layer.output])
pred3d = run_func(func, x, batch_size=20)
fusion_sed_list.append(pred3d)
print("FINN pred_sed ended", (time.time()-t_pred))
print("Prediction time: %s" % (time.time() - t1,))
# Write out AT probabilities
if at_bool:
fusion_at = np.mean(np.array(fusion_at_list), axis=0)
print("AT shape: %s" % (fusion_at.shape,))
io_task4.at_write_prob_mat_to_csv(
na_list=na_list,
prob_mat=fusion_at,
out_path=os.path.join(args.out_dir, "at_prob_mat.csv.gz"))
# Write out SED probabilites
if sed_bool:
fusion_sed = np.mean(np.array(fusion_sed_list), axis=0)
print("SED shape:%s" % (fusion_sed.shape,))
io_task4.sed_write_prob_mat_list_to_csv(
na_list=na_list,
prob_mat_list=fusion_sed,
out_path=os.path.join(args.out_dir, "sed_prob_mat_list.csv.gz"))
print("FINN Prediction finished!, time: ", (time.time()-t_rec))#this is really just me seeing if this is where most of the time is spent)
def train(args):
num_classes = cfg.num_classes
# Load training & testing data
(tr_x, tr_y, tr_na_list) = load_hdf5_data(args.tr_hdf5_path, verbose=1)
(te_x, te_y, te_na_list) = load_hdf5_data(args.te_hdf5_path, verbose=1)
print("tr_x.shape: %s" % (tr_x.shape,))
# Scale data
tr_x = do_scale(tr_x, args.scaler_path, verbose=1)
te_x = do_scale(te_x, args.scaler_path, verbose=1)
# Build model
(_, n_time, n_freq) = tr_x.shape # (N, 240, 64)
input_logmel = Input(shape=(n_time, n_freq), name='in_layer') # (N, 240, 64)
a1 = Reshape((n_time, n_freq, 1))(input_logmel) # (N, 240, 64, 1)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 32, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 16, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 8, 128)
a1 = block(a1)
a1 = block(a1)
a1 = MaxPooling2D(pool_size=(1, 2))(a1) # (N, 240, 4, 128)
a1 = Conv2D(256, (3, 3), padding="same", activation="relu", use_bias=True)(a1)
a1 = MaxPooling2D(pool_size=(1, 4))(a1) # (N, 240, 1, 256)
a1 = Reshape((240, 256))(a1) # (N, 240, 256)
# Gated BGRU
rnnout = Bidirectional(GRU(128, activation='linear', return_sequences=True))(a1)
rnnout_gate = Bidirectional(GRU(128, activation='sigmoid', return_sequences=True))(a1)
a2 = Multiply()([rnnout, rnnout_gate])
# Attention
cla = TimeDistributed(Dense(num_classes, activation='sigmoid'), name='localization_layer')(a2)
att = TimeDistributed(Dense(num_classes, activation='softmax'))(a2)
out = Lambda(outfunc, output_shape=(num_classes,))([cla, att])
model = Model(input_logmel, out)
model.summary()
# Compile model
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
# Save model callback
filepath = os.path.join(args.out_model_dir, "gatedAct_rationBal44_lr0.001_normalization_at_cnnRNN_64newMel_240fr.{epoch:02d}-{val_acc:.4f}.hdf5")
create_folder(os.path.dirname(filepath))
save_model = ModelCheckpoint(filepath=filepath,
monitor='val_acc',
verbose=0,
save_best_only=False,
save_weights_only=False,
mode='auto',
period=1)
# Data generator
gen = RatioDataGenerator(batch_size=44, type='train')
# Train
model.fit_generator(generator=gen.generate([tr_x], [tr_y]),
steps_per_epoch=100, # 100 iters is called an 'epoch'
epochs=31, # Maximum 'epoch' to train
verbose=1,
callbacks=[save_model],
validation_data=(te_x, te_y))