自定义损失缺少梯度操作

Question

我不太确定如何处理这个问题以及为什么会收到此错误。

    raise ValueError('An operation has `None` for gradient. '
    ValueError: An operation has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.

因此，我使用自定义三重损失作为此博客中的损失函数。 https://omoindrot.github.io/triplet-loss 而我在 keras 中运行它应该不是问题。但我无法让它与我的模型一起正常工作。

这就是他们的损失函数。它需要的其他代码是直接复制：

def batch_hard_triplet_loss(embeddings, labels, margin = 0.3, squared=False):

    # Get the pairwise distance matrix
    pairwise_dist = pairwise_distances(embeddings, squared=squared)
    mask_anchor_positive = _get_anchor_positive_triplet_mask(labels)
    mask_anchor_positive = tf.to_float(mask_anchor_positive)
    anchor_positive_dist = tf.multiply(mask_anchor_positive, pairwise_dist)
    hardest_positive_dist = tf.reduce_max(anchor_positive_dist, axis=1, keepdims=True)
    mask_anchor_negative = _get_anchor_negative_triplet_mask(labels)
    mask_anchor_negative = tf.to_float(mask_anchor_negative)
    max_anchor_negative_dist = tf.reduce_max(pairwise_dist, axis=1, keepdims=True)
    anchor_negative_dist = pairwise_dist + max_anchor_negative_dist * (1.0 - mask_anchor_negative)
    hardest_negative_dist = tf.reduce_min(anchor_negative_dist, axis=1, keepdims=True)
    # Combine biggest d(a, p) and smallest d(a, n) into final triplet loss
    triplet_loss = tf.maximum(hardest_positive_dist - hardest_negative_dist + margin, 0.0)
    triplet_loss = tf.reduce_mean(triplet_loss)
#triplet_loss = k.mean(triplet_loss) # use keras mean

    return triplet_loss

现在这是我正在使用的模型。

train_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,
    ....
    validation_split=0.2) # set validation split

train_generator = train_datagen.flow_from_directory(
    IMAGE_DIR,
    target_size=(224, 224),
    batch_size=BATCHSIZE,
    class_mode='categorical',
    subset='training') # set as training data

validation_generator = train_datagen.flow_from_directory(
    IMAGE_DIR, # same directory as training data
    target_size=(224, 224),
    batch_size=BATCHSIZE,
    class_mode='categorical',
    subset='validation') # set as validation data

print("Initializing Model...")
# Get base model
input_layer = preloadmodel.get_layer('model_1').get_layer('input_1').input
layer_output = preloadmodel.get_layer('model_1').get_layer('glb_avg_pool').output
# Make extractor
base_network = Model(inputs=input_layer, outputs=layer_output)

# Define new model
input_images = Input(shape=(224, 224, 3), name='input_image')  # input layer for images
#input_labels = Input(shape=(num_classes,), name='input_label')  # input layer for labels
embeddings = base_network(input_images)  # output of network -> embeddings
output = Dense(1, activation='sigmoid')(embeddings)
model = Model(inputs=input_images,  outputs=output)
# Compile model
model.compile(loss=batch_hard_triplet_loss, optimizer='adam')

Answer 1

好的，我通过大量研究解决了这些问题。现在它没有解决我的问题，因为代码仍然不起作用，但损失函数的问题已解决。关注此博客 https://medium.com/@Bloomore/how-to-write-a-custom-loss-function-with-additional-arguments-in-keras-5f193929f7a0

我把损失函数改成这样：

def batch_hard_triplet_loss(embeddings, labels, margin = 0.3, squared=False):
    # Get the pairwise distance matrix
    pairwise_dist = pairwise_distances(embeddings, squared=squared)
    mask_anchor_positive = _get_anchor_positive_triplet_mask(labels)
    mask_anchor_positive = tf.to_float(mask_anchor_positive)
    anchor_positive_dist = tf.multiply(mask_anchor_positive, pairwise_dist)
    hardest_positive_dist = tf.reduce_max(anchor_positive_dist, axis=1, keepdims=True)
    mask_anchor_negative = _get_anchor_negative_triplet_mask(labels)
    mask_anchor_negative = tf.to_float(mask_anchor_negative)
    max_anchor_negative_dist = tf.reduce_max(pairwise_dist, axis=1, keepdims=True)
    anchor_negative_dist = pairwise_dist + max_anchor_negative_dist * (1.0 - mask_anchor_negative)
    hardest_negative_dist = tf.reduce_min(anchor_negative_dist, axis=1, keepdims=True)
    def loss(y_true, y_pred):

        # Combine biggest d(a, p) and smallest d(a, n) into final triplet loss
        #triplet_loss = tf.maximum(hardest_positive_dist - hardest_negative_dist + margin, 0.0)
        #triplet_loss = tf.reduce_mean(triplet_loss)
        triplet_loss = k.maximum(hardest_positive_dist - hardest_negative_dist + margin, 0.0)
        triplet_loss = k.mean(triplet_loss) # use keras mean
        return triplet_loss

    return loss

然后在模型中这样调用它：

batch_loss = batch_hard_triplet_loss(embeddings, input_labels, 0.4, False)
model = Model(inputs=input_images,  outputs=embeddings)
model.compile(loss=batch_loss, optimizer='adam')

它现在给了我这些问题

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'input_label' with dtype float and shape [?,99]
 [[{{node input_label}}]]

但是，嘿，我们继续前进。问题是 keras 只接受带有 2 个参数的损失，所以你需要像我在这里做的那样从另一个函数调用损失。

自定义损失缺少梯度操作

Custom loss is missing an operation for gradient

python-3.x

keras

tensorflow

loss-function