如何培养和实施对keras的信心

Question

我试图训练一个CNN，它将输出（x，y）坐标的面部特征，如左眼开始，左眼结束，左眼中心，右眼开始，右眼中心和右眼结束。根据coursera课程深度学习专业化，我需要为每个点获取0和1值（真或假），然后为每个点输出x和y坐标。如果值为0（False），我需要在训练期间忽略x和y的输出。在

我试图寻找一种方法来做这件事，但没找到。最后，我考虑将所有缺失的（我们没有特定特性的数据）x，y坐标设置为0，然后训练下面的网络。在

def get_model():
    inputs = Input(shape=(96, 96, 1))

    # a layer instance is callable on a tensor, and returns a tensor
    x = Conv2D(16, kernel_size=5, padding='same', activation='relu')(inputs)
    x = Conv2D(32, kernel_size=5, padding='valid', activation='relu')(x)
    x = Dropout(0.25)(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)
    x = BatchNormalization()(x)
    x = Conv2D(64, kernel_size=5, padding='valid', activation='relu')(x)
    x = Conv2D(128, kernel_size=5, padding='valid', activation='relu')(x)
    x = Dropout(0.25)(x)
    x = MaxPooling2D(pool_size=(2, 2))(x)
    x = BatchNormalization()(x)
    x = Conv2D(256, kernel_size=3, padding='valid', activation='relu')(x)
    x = Dropout(0.25)(x)
    x = MaxPooling2D(pool_size=(1, 1))(x)
    x = Flatten()(x)
    present = Dense(256)(x)
    present = PReLU()(present)
    present = Dense(15, activation="sigmoid", name="classification")(present)
    position = Dense(256)(x)
    position = PReLU()(position)
    position = BatchNormalization()(position)
    position = Dense(128)(position)
    position = PReLU()(position)
    position = BatchNormalization()(position)
    position = Dense(64)(position)
    position = PReLU()(position)
    position = Dense(30, activation="relu", name="position")(position)

    # This creates a model that includes
    # the Input layer and three Dense layers
    model = Model(inputs=inputs, outputs=[present, position])
    model.compile(optimizer='adam',
                loss={'classification':'binary_crossentropy','position':'mse'},
                metrics=['accuracy'])
    return model

虽然分类精度很好，但定位精度在1%以下。在

Answer 1

可以创建自定义损失，以便在坐标损失计算中忽略置信值低于阈值的点，但不忽略置信度损失计算：

def custom_loss(y_true, y_pred, conf_dim=-1, conf_thresh=0, px_weight=0.5, conf_weight=0.5):
   # compute sum of squared coordinate errors
   px_err = K.sum(K.square(y_pred[:, :conf_dim] - y_true[:, :conf_dim]), axis=-1)

   # compute squared confidence errors
   conf_err = K.square(y_pred[:, conf_dim] - y_true[:, conf_dim])

   # set loss of points whose confidences are below our threshold to zero
   px_err = px_err * tf.cast(y_true[:, conf_dim] > conf_thresh, y_true.dtype)

   # calculate mean over errors
   px_err, conf_err = K.mean(px_err), K.mean(conf_err)

   # return sum of weighted pixel and confidence errors
   return px_weight * px_err + conf_weight * conf_err


# compute a small example with numpy arrays
#                      -
#                   x, y, c
#                      -
y_true = np.array([[1, 2, 1],
                   [3, 4, 0],  # confidence zero
                   [5, 6, 1]])

y_pred = np.array([[1, 2, 1],  # correct
                   [1, 2, 1],  # wrong (not included in pixel loss because of confidence)
                   [2, 3, 1]]) # wrong

# run through loss function
print("Loss:", K.eval(custom_loss(K.variable(y_true), K.variable(y_pred)))) 
>>> Loss: 3.1666667