python - 使用 FaceNet 创建三元组损失模型时遇到问题

Question

我正在尝试使用 FacenetModel 实现三元组损失模型。我使用了 coursera 作业中提供的 Facenet 实现。

每当我编译模型时，我都会收到此错误:

ValueError: No data provided for "FaceRecoModel". Need data for each key in: ['FaceRecoModel', 'FaceRecoModel', 'FaceRecoModel']

我的代码:

def batch_generator(batch_size = 64):
    while True:
        pos = positiveImg[np.random.choice(len(positiveImg), batch_size)]
        neg = negativeImg[np.random.choice(len(negativeImg), batch_size)]
        anc = anchorsImg[np.random.choice(len(anchorsImg), batch_size)]
        
        x_data = {'inp1': anc,
                  'inp2': pos,
                  'inp3': neg
                  }
        y_data = {'y1': np.zeros((64,0)),
                  'y2': np.zeros((64,0)),
                  'y3': np.zeros((64,0))}
        yield (x_data, y_data)

def triplet_loss(y_true, y_pred):    
    anchor, positive, negative = y_pred[0], y_pred[1], y_pred[2]
    pos_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, positive)), axis=-1)
    neg_dist = tf.reduce_sum(tf.square(tf.subtract(anchor, negative)), axis=-1)
    basic_loss = tf.add(tf.subtract(pos_dist, neg_dist), 0.2)
    loss = tf.reduce_sum(tf.maximum(basic_loss, 0.0))

    return loss

def getModels():
    FRmodel = keras.models.load_model('FR.h5', custom_objects={'triplet_loss': triplet_loss})
    
    inp1 = Input((3, 96, 96), name= 'inp1')
    inp2 = Input((3, 96, 96), name= 'inp2')
    inp3 = Input((3, 96, 96), name= 'inp3')
    
    pred1 = FRmodel(inp1)
    pred2 = FRmodel(inp2)
    pred3 = FRmodel(inp3)
    
    inputs = [inp1, inp2, inp3]
    outputs = [pred1, pred2, pred3]
    
    model = keras.models.Model(inputs=[inp1, inp2, inp3], outputs= [pred1, pred2, pred3])
   
    return FRmodel, model

generator = batch_generator(64)

FRmodel, my_model = getModels()
my_model.compile(optimizer = 'adam', loss = triplet_loss, metrics = ['accuracy'])   
my_model.fit_generator(generator, epochs=5,steps_per_epoch=30)

预训练 Facenet 模型摘要:

FRmodel.summary() : https://codeshare.io/arxmev

my_model.summary() : https://codeshare.io/arx3N6

Answer 1

在 coursera 的论坛上找到了解决方案。这有点棘手。我必须使用 Lambda 在 keras 层包装器中添加三元组损失的欧几里德距离。根据文档:

Wraps arbitrary expression as a Layer object.

新的实现:

`

    def triplet_loss_v2(y_true, y_pred):
        positive, negative = y_pred[:,0,0], y_pred[:,1,0]
        margin = K.constant(0.2)
        loss = K.mean(K.maximum(K.constant(0), positive - negative + margin))
        return loss # shape = [1]

    def euclidean_distance(vects):
        x, y = vects   # shape = [batch_size, 2, 1]
        dist = K.sqrt(K.maximum(K.sum(K.square(x - y), axis=1, keepdims=True), K.epsilon()))
        return dist   # shape = [batch_size, 1]

    FRmodel = faceRecoModel(input_shape=(3, 96, 96))
    load_weights_from_FaceNet(FRmodel)

    for layer in FRmodel.layers[0: 80]:
        layer.trainable  =  False

    input_shape=(3, 96, 96)
    anchor = Input(shape=input_shape, name = 'anchor')
    anchorPositive = Input(shape=input_shape, name = 'anchorPositive')
    anchorNegative = Input(shape=input_shape, name = 'anchorNegative')

    anchorCode = FRmodel(anchor)
    anchorPosCode = FRmodel(anchorPositive)
    anchorNegCode = FRmodel(anchorNegative)

    positive_dist = Lambda(euclidean_distance, name='pos_dist')([anchorCode, anchorPosCode])
    negative_dist = Lambda(euclidean_distance, name='neg_dist')([anchorCode, anchorNegCode])
    stacked_dists = Lambda(lambda vects: K.stack(vects, axis=1), name='stacked_dists')([positive_dist, negative_dist]) # shape = [batch_size, 2, 1]

    tripletModel = Model([anchor, anchorPositive, anchorNegative], stacked_dists, name='triple_siamese')
    tripletModel.compile(optimizer = 'adadelta', loss = triplet_loss_v2, metrics = None)

    gen = batch_generator(64)
    tripletModel.fit_generator(gen, epochs=1,steps_per_epoch=5)`