yushuai / ftanet-melodic Goto Github PK

`## Load f0 frequency
ref_arr = np.loadtxt(data_folder + 'f0ref/' + fname + '.txt')   # (T, 2)` 

' def create_model(input_shape=(320, 430, 3)):
visible = Input(shape=input_shape)
x = BatchNormalization()(visible)

## Bottom
# bm = BatchNormalization()(x)
bm = x
bm = Conv2D(16, (4, 1), padding='valid', strides=(4, 1), activation='selu')(bm) # 80
bm = Conv2D(16, (4, 1), padding='valid', strides=(4, 1), activation='selu')(bm) # 20
bm = Conv2D(16, (4, 1), padding='valid', strides=(4, 1), activation='selu')(bm) # 5
bm = Conv2D(1,  (5, 1), padding='valid', strides=(5, 1), activation='selu')(bm) # 1

# 保持高分辨率，关注细节
shape=input_shape
x_r, x_t, x_f = FTA_Module(x, (shape[0], shape[1], 32), 3, 3)
x = SF_Module([x_r, x_t, x_f], 32, 4, 4)
x = MaxPooling2D((2, 2))(x)

x_r, x_t, x_f = FTA_Module(x, (shape[0]//2, shape[1]//2, 64), 3, 3)
x = SF_Module([x_r, x_t, x_f], 64, 4, 4)
x = MaxPooling2D((2, 2))(x)

x_r, x_t, x_f = FTA_Module(x, (shape[0]//4, shape[1]//4, 128), 3, 3)
x = SF_Module([x_r, x_t, x_f], 128, 4, 4)

x_r, x_t, x_f = FTA_Module(x, (shape[0]//4, shape[1]//4, 128), 3, 3)
x = SF_Module([x_r, x_t, x_f], 128, 4, 4)

x = UpSampling2D((2, 2))(x)
x_r, x_t, x_f = FTA_Module(x, (shape[0]//2, shape[1]//2, 64), 3, 3)
x = SF_Module([x_r, x_t, x_f], 64, 4, 4)

x = UpSampling2D((2, 2))(x)
x_r, x_t, x_f = FTA_Module(x, (shape[0], shape[1], 32), 3, 3)
x = SF_Module([x_r, x_t, x_f], 32, 4, 4)

x_r, x_t, x_f = FTA_Module(x, (shape[0], shape[1], 1), 3, 3)
x = SF_Module([x_r, x_t, x_f], 1, 4, 4)
x = Concatenate(axis=1)([bm, x])

# Softmax
x = Lambda(K.squeeze, arguments={'axis': -1})(x) # (321, 430)
x = Softmax(axis=-2)(x)

return Model(inputs=visible, outputs=x)`

x_r, x_t, x_f = FTA_Module(x, (shape[0]//4, shape[1]//4, 128), 3, 3)
x = SF_Module([x_r, x_t, x_f], 128, 4, 4)      --------------------------> shape : (None, 80, 107, 128)

x = UpSampling2D((2, 2))(x)-----------------------------------------> shape: (None, 160, 214, 128)
x_r, x_t, x_f = FTA_Module(x, (shape[0]//2, shape[1]//2, 64), 3, 3)----> input shape : (None, 160, 215, 128)