keras 输出网络结构_Keras中无损实现复杂（多入参）的损失函数

model.compile(loss=YOUR_CUSTOM_LOSS_FUNCTION)

def simple_loss(y_true, y_pred):
    pass

from keras import layers as KL
from keras import models as KM
 
def create_simple_model():
    input_img = KL.Input([64, 64, 3])
    branch1 = KL.Conv2D(64, (3, 3), strides=(4, 4), activation="relu")(input_img)
    branch2 = KL.Conv2D(64, (3, 3), strides=(4, 4), activation="relu")(input_img)
    concat1 = KL.Concatenate()([branch1, branch2])
    deconv1 = KL.Deconv2D(1, (3, 3), strides=(4, 4), activation="relu")(concat1)
    output = KL.Conv2D(1, (1, 1), strides=(1, 1), activation="sigmoid")(deconv1)
    return KM.Model(inputs=input_img, outputs=output)
 
def my_simple_loss(y_true, y_pred):
    # do what you want here
    return binary_crossentropy(y_true, y_pred)

import numpy as np
 
def fake_data_generator(num_samples):
    while (1):
        imgs = np.random.random((num_samples, 64, 64, 3)).astype("float32")
        masks = np.random.random((num_samples, 64, 64, 1)).astype("float32")
        yield imgs, masks

train_gen = fake_data_generator(10)
val_gen = fake_data_generator(5)
 
model = create_simple_model()
model.summary()
model.compile(optimizer="adam", loss=my_simple_loss)
model.fit_generator(
    train_gen,
    epochs=10,
    steps_per_epoch=50,
    validation_data=val_gen,
    validation_steps=5
)

from keras.losses import mean_squared_error, binary_crossentropy
 
def my_complex_loss_graph(target, deconv1, output):
    mse_deconv1 = mean_squared_error(target, deconv1)
    bce_output = binary_crossentropy(target, output)
    final_loss = mse_deconv1 + bce_output
    return K.mean(final_loss)

import tensorflow as tf 
 
def create_complex_model(mode="train"):
    assert mode in ("train", "predict"), "only 'train' and 'predict' mode supported"
 
    input_img = KL.Input([64, 64, 3])
    branch1 = KL.Conv2D(64, (3, 3), strides=(4, 4), activation="relu")(input_img)
    branch2 = KL.Conv2D(64, (3, 3), strides=(4, 4), activation="relu")(input_img)
    concat1 = KL.Concatenate(name="concat1")([branch1, branch2])
    deconv1 = KL.Deconv2D(1, (3, 3), strides=(4, 4), activation="relu")(concat1)
    output = KL.Conv2D(1, (1, 1), strides=(1, 1), activation="sigmoid")(deconv1)
 
    if mode == "train":
        #本文最开始提到过，keras generator中yield input, target的target是无法获取
        #参考github issues:https://github.com/keras-team/keras/issues/11812
        #所以为了取到target,我们必须须把target也当作inputs的一部分传进来即 
        #yield  [input,target], [], 然后再通过KL.Input按顺序获取
        target = KL.Input([64, 64, 1], name="target")
        my_complex_loss = KL.Lambda(
            lambda x: my_complex_loss_graph(*x), name="complex_loss"
        )([target, deconv1, output])
        inputs = [input_img, target]
        outputs = [output, my_complex_loss]
    else:
        #predict阶段，就不用计算loss了所以这里不加入loss层和metric层
        inputs = input_img
        outputs = output
 
    model = KM.Model(inputs=inputs, outputs=outputs)
 
    #重点
    model._losses = []
    model._per_input_losses = {}
    #通过add_loss来把之前通过KL.Lambda定义的层加入loss，当添加了多个loss层时，optimizer实际优  
    #化的是多个loss的和
    for loss_name in ["complex_loss"]:
        layer = model.get_layer(loss_name)
        if layer.output in model.losses:
            continue
        loss = tf.reduce_mean(layer.output, keepdims=True)
        model.add_loss(loss)
    #其实这里可以添加的不只loss， 有助于监视模型情况的metrics比如f1 score, iou等等也可以通过   
    #model.metrics_tensors.append()来添加
 
    return model

def fake_data_generator_2(num_samples):
    while (1):
        imgs = np.random.random((num_samples, 64, 64, 3)).astype("float32")
        masks = np.random.random((num_samples, 64, 64, 1)).astype("float32")
        inputs = [imgs, masks]
        targets = []
        yield inputs, targets

train_gen = fake_data_generator_2(10)
val_gen = fake_data_generator_2(5)
model = create_complex_model("train")
model.summary()
model.compile(
    optimizer="adam",
    loss=[None] * len(model.outputs)
)
model.fit_generator(
    train_gen,
    epochs=10,
    steps_per_epoch=50,
    validation_data=val_gen,
    validation_steps=5
)