我正在尝试实现本文 (https://arxiv.org/pdf/1511.00561.pdf) 中描述的 SegNet 的较小规模版本，但我正在尝试对其进行调整以检测边缘
数据集: 我使用的是 BSDS500 边界数据集，我裁剪并旋转了图像，使它们的大小为 320x480x3 而不是 321x481x3
输入形状、200 张训练图像和 100 张验证图像:
x_train: (200, 320, 480, 3)
x_val: (100, 320, 480, 3)
y_train: (200, 153600)
y_val: (100, 153600)
框架: 我正在使用带有 tensorflow 后端的 keras
这些是我用于自定义池化和反池化层的函数:
def pool_argmax2D(x, pool_size=(2,2), strides=(2,2)):
padding = 'SAME'
pool_size = [1, pool_size[0], pool_size[1], 1]
strides = [1, strides[0], strides[1], 1]
ksize = [1, pool_size[0], pool_size[1], 1]
output, argmax = tf.nn.max_pool_with_argmax(
x,
ksize = ksize,
strides = strides,
padding = padding
)
return [output, argmax]
def unpool2D(pool, argmax, ksize=(2,2)):
with tf.variable_scope("unpool"):
input_shape = tf.shape(pool)
output_shape = [input_shape[0],
input_shape[1] * ksize[0],
input_shape[2] * ksize[1],
input_shape[3]]
flat_input_size = tf.cumprod(input_shape)[-1]
flat_output_shape = tf.cast([output_shape[0],
output_shape[1] * output_shape[2] * output_shape[3]], tf.int64)
pool_ = tf.reshape(pool, [flat_input_size])
batch_range = tf.reshape(tf.range(tf.cast(output_shape[0], tf.int64), dtype=tf.int64),
shape=[input_shape[0], 1, 1, 1])
b = tf.ones_like(argmax) * batch_range
b = tf.reshape(b, [flat_input_size, 1])
ind_ = tf.reshape(argmax, [flat_input_size, 1]) % flat_output_shape[1]
ind_ = tf.concat([b, ind_], 1)
ret = tf.scatter_nd(ind_, pool_, shape=flat_output_shape)
ret = tf.reshape(ret, output_shape)
return ret
这是模型的代码:
batch_size = 4
kernel = 3
pool_size=(2,2)
img_shape = (320,480,3)
inputs = Input(shape=img_shape, name='main_input')
conv_1 = Conv2D(32, (kernel, kernel), padding="same")(inputs)
conv_1 = BatchNormalization()(conv_1)
conv_1 = Activation("relu")(conv_1)
conv_2 = Conv2D(32, (kernel, kernel), padding="same")(conv_1)
conv_2 = BatchNormalization()(conv_2)
conv_2 = Activation("relu")(conv_2)
pool_1, mask_1 = Lambda(pool_argmax2D, arguments={'pool_size': pool_size, 'strides': pool_size})(conv_2)
conv_3 = Conv2D(64, (kernel, kernel), padding="same")(pool_1)
conv_3 = BatchNormalization()(conv_3)
conv_3 = Activation("relu")(conv_3)
conv_4 = Conv2D(64, (kernel, kernel), padding="same")(conv_3)
conv_4 = BatchNormalization()(conv_4)
conv_4 = Activation("relu")(conv_4)
pool_2, mask_2 = Lambda(pool_argmax2D, arguments={'pool_size': pool_size, 'strides': pool_size})(conv_4)
conv_5 = Conv2D(64, (kernel, kernel), padding="same")(pool_2)
conv_5 = BatchNormalization()(conv_5)
conv_5 = Activation("relu")(conv_5)
unpool_1 = Lambda(unpool2D, output_shape = (160,240,64), arguments={'ksize':pool_size, 'argmax': mask_2})(conv_5)
conv_6 = Conv2D(64, (kernel, kernel), padding="same")(unpool_1)
conv_6 = BatchNormalization()(conv_6)
conv_6 = Activation("relu")(conv_6)
conv_7 = Conv2D(64, (kernel, kernel), padding="same")(conv_6)
conv_7 = BatchNormalization()(conv_7)
conv_7 = Activation("relu")(conv_7)
unpool_2 = Lambda(unpool2D, output_shape = (320,480,64), arguments={'ksize':pool_size, 'argmax': mask_1})(conv_7)
conv_8 = Conv2D(32, (kernel, kernel), padding="same")(unpool_2)
conv_8 = BatchNormalization()(conv_8)
conv_8 = Activation("relu")(conv_8)
conv_9 = Conv2D(32, (kernel, kernel), padding="same")(conv_8)
conv_9 = BatchNormalization()(conv_9)
conv_9 = Activation("relu")(conv_9)
conv_10 = Conv2D(1, (1, 1), padding="same")(conv_9)
conv_10 = BatchNormalization()(conv_10)
flatten_1 = Flatten()(conv_10)
outputs = Activation('softmax')(flatten_1)
model = Model(inputs=inputs, outputs=outputs)
模型在我运行时正确编译:
model.compile(optimizer='adam', loss='mean_absolute_error', metrics=['accuracy'])
model.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
main_input (InputLayer) (None, 320, 480, 3) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 320, 480, 32) 896
_________________________________________________________________
batch_normalization_1 (Batch (None, 320, 480, 32) 128
_________________________________________________________________
activation_1 (Activation) (None, 320, 480, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 320, 480, 32) 9248
_________________________________________________________________
batch_normalization_2 (Batch (None, 320, 480, 32) 128
_________________________________________________________________
activation_2 (Activation) (None, 320, 480, 32) 0
_________________________________________________________________
lambda_1 (Lambda) [(None, 160, 240, 32), (N 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 160, 240, 64) 18496
_________________________________________________________________
batch_normalization_3 (Batch (None, 160, 240, 64) 256
_________________________________________________________________
activation_3 (Activation) (None, 160, 240, 64) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 160, 240, 64) 36928
_________________________________________________________________
batch_normalization_4 (Batch (None, 160, 240, 64) 256
_________________________________________________________________
activation_4 (Activation) (None, 160, 240, 64) 0
_________________________________________________________________
lambda_2 (Lambda) [(None, 80, 120, 64), (No 0
_________________________________________________________________
conv2d_5 (Conv2D) (None, 80, 120, 64) 36928
_________________________________________________________________
batch_normalization_5 (Batch (None, 80, 120, 64) 256
_________________________________________________________________
activation_5 (Activation) (None, 80, 120, 64) 0
_________________________________________________________________
lambda_3 (Lambda) (None, 160, 240, 64) 0
_________________________________________________________________
conv2d_6 (Conv2D) (None, 160, 240, 64) 36928
_________________________________________________________________
batch_normalization_6 (Batch (None, 160, 240, 64) 256
_________________________________________________________________
activation_6 (Activation) (None, 160, 240, 64) 0
_________________________________________________________________
conv2d_7 (Conv2D) (None, 160, 240, 64) 36928
_________________________________________________________________
batch_normalization_7 (Batch (None, 160, 240, 64) 256
_________________________________________________________________
activation_7 (Activation) (None, 160, 240, 64) 0
_________________________________________________________________
lambda_4 (Lambda) (None, 320, 480, 64) 0
_________________________________________________________________
conv2d_8 (Conv2D) (None, 320, 480, 32) 18464
_________________________________________________________________
batch_normalization_8 (Batch (None, 320, 480, 32) 128
_________________________________________________________________
activation_8 (Activation) (None, 320, 480, 32) 0
_________________________________________________________________
conv2d_9 (Conv2D) (None, 320, 480, 32) 9248
_________________________________________________________________
batch_normalization_9 (Batch (None, 320, 480, 32) 128
_________________________________________________________________
activation_9 (Activation) (None, 320, 480, 32) 0
_________________________________________________________________
conv2d_10 (Conv2D) (None, 320, 480, 1) 33
_________________________________________________________________
batch_normalization_10 (Batc (None, 320, 480, 1) 4
_________________________________________________________________
flatten_1 (Flatten) (None, 153600) 0
_________________________________________________________________
activation_10 (Activation) (None, 153600) 0
=================================================================
Total params: 205,893
Trainable params: 204,995
Non-trainable params: 898
_________________________________________________________________
但是当试图拟合模型时
history = model.fit(x=x_train, y=y_train, batch_size=batch_size, epochs=3, verbose=2, validation_data=(x_val,y_val))
我遇到这个错误:
InvalidArgumentError: Input to reshape is a tensor with 4915200 values, but the requested shape has 9830400
[[{{node lambda_4/unpool/Reshape_3}} = Reshape[T=DT_INT64, Tshape=DT_INT32, _device="/job:localhost/replica:0/task:0/device:GPU:0"](lambda_1/MaxPoolWithArgmax:1, lambda_4/unpool/Reshape_2/shape)]]
[[{{node lambda_4/unpool/strided_slice_6/_515}} = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_1479_lambda_4/unpool/strided_slice_6", tensor_type=DT_INT64, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]
我查看了每一层之后的所有形状，它们是我所期望的。我还测试了样本张量上的池化/解池化函数，它们产生了预期的输出。我在这里做错了什么？
我一直在努力解决这个问题，非常感谢任何帮助!
请您参考如下方法:
发现问题，mask_1 有 32 个 channel ，而 unpool_2 试图用 64 个 channel reshape 输出。我只是重新安排了一些东西，使深度对齐。
