tensorflow - TF2.1 : SegNet model architecture problem. 度量计算错误，保持不变并收敛到确定的值-6ren

tensorflow - TF2.1 : SegNet model architecture problem. 度量计算错误，保持不变并收敛到确定的值

转载作者：行者123 更新时间：2023-12-03 14:53:53

我正在 Tensorflow 2.1.0 中构建自定义模型 ( SegNet ) .

我面临的第一个问题是重新利用本文所述所需的最大池操作的索引。
基本上，由于它是一种编码器-解码器架构，因此在解码中需要网络编码部分的池化索引来对特征图进行上采样并保持相应索引的目标值。

现在，在 TF 中，tf.keras.layers.MaxPool2D 层默认不导出这些索引。 (例如在 PyTorch 中)。
要获得最大池操作的索引，需要使用 tf.nn.max_pool_with_argmax .
无论如何，此操作以扁平格式返回索引(argmax)，这需要进一步的操作才能在网络的其他部分有用。

为了实现一个执行 MaxPooling2D 并导出这些索引(扁平化)的层，我在 keras 中定义了一个自定义层。

class MaxPoolingWithArgmax2D(Layer):

def __init__(
        self,
        pool_size=(2, 2),
        strides=2,
        padding='same',
        **kwargs):
    super(MaxPoolingWithArgmax2D, self).__init__(**kwargs)
    self.padding = padding
    self.pool_size = pool_size
    self.strides = strides

def call(self, inputs, **kwargs):
    padding = self.padding
    pool_size = self.pool_size
    strides = self.strides
    output, argmax = tf.nn.max_pool_with_argmax(
        inputs,
        ksize=pool_size,
        strides=strides,
        padding=padding.upper(),
        output_dtype=tf.int64)
    return output, argmax

显然，该层用于网络的编码部分，因此需要解码各个层来执行逆运算(UpSampling2D)，并利用索引(本文中有关此操作的更多详细信息)。

经过一番研究，我找到了遗留代码(TF<2.1.0)并对其进行了修改以执行操作。
无论如何，我不是 100% 相信这段代码运行良好，事实上有些事情我不喜欢。

class MaxUnpooling2D(Layer):
def __init__(self, size=(2, 2), **kwargs):
    super(MaxUnpooling2D, self).__init__(**kwargs)
    self.size = size

def call(self, inputs, output_shape=None):
    updates, mask = inputs[0], inputs[1]
    with tf.name_scope(self.name):
        mask = tf.cast(mask, 'int32')
        #input_shape = tf.shape(updates, out_type='int32')
        input_shape = updates.get_shape()

        # This statement is required if I don't want to specify a batch size
        if input_shape[0] == None:
            batches = 1
        else:
            batches = input_shape[0]

        #  calculation new shape
        if output_shape is None:
            output_shape = (
                    batches,
                    input_shape[1]*self.size[0],
                    input_shape[2]*self.size[1],
                    input_shape[3])

        # calculation indices for batch, height, width and feature maps
        one_like_mask = tf.ones_like(mask, dtype='int32')
        batch_shape = tf.concat(
                [[batches], [1], [1], [1]],
                axis=0)
        batch_range = tf.reshape(
                tf.range(output_shape[0], dtype='int32'),
                shape=batch_shape)
        b = one_like_mask * batch_range
        y = mask // (output_shape[2] * output_shape[3])
        x = (mask // output_shape[3]) % output_shape[2]
        feature_range = tf.range(output_shape[3], dtype='int32')
        f = one_like_mask * feature_range

        # transpose indices & reshape update values to one dimension
        updates_size = tf.size(updates)
        indices = tf.transpose(tf.reshape(
            tf.stack([b, y, x, f]),
            [4, updates_size]))
        values = tf.reshape(updates, [updates_size])
        ret = tf.scatter_nd(indices, values, output_shape)
        return ret

困扰我的事情是:

执行展开索引的操作 (MaxUnpooling2D) 与了解特定的批量大小严格相关，对于模型验证，我希望为 None 或未指定。

我不确定这段代码实际上是否与库的其余部分 100% 兼容。事实上在fit期间如果我使用 tf.keras.metrics.MeanIoU值收敛到 0.341并在除第一个以外的每个其他时期保持不变。相反，标准准确度指标工作得很好。

深度网络架构

以下是模型的完整定义。

import tensorflow as tf
import tensorflow.keras as keras
import tensorflow.keras.layers as layers
from tensorflow.keras.layers import Layer


class SegNet:
    def __init__(self, data_shape, classes = 3, batch_size = None):
        self.MODEL_NAME = 'SegNet'
        self.MODEL_VERSION = '0.2'

        self.classes = classes
        self.batch_size = batch_size

        self.build_model(data_shape)

    def build_model(self, data_shape):
        input_shape = (data_shape, data_shape, 3)

        inputs = keras.Input(shape=input_shape, batch_size=self.batch_size, name='Input')

        # Build sequential model

        # Encoding
        encoders = 5
        feature_maps = [64, 128, 256, 512, 512]
        n_convolutions = [2, 2, 3, 3, 3]
        eb_input = inputs
        eb_argmax_indices = []
        for encoder_index in range(encoders):
            encoder_block, argmax_indices = self.encoder_block(
                eb_input, encoder_index, feature_maps[encoder_index], n_convolutions[encoder_index])
            eb_argmax_indices.append(argmax_indices)
            eb_input = encoder_block

        # Decoding
        decoders = encoders
        db_input = encoder_block
        eb_argmax_indices.reverse()
        feature_maps.reverse()
        n_convolutions.reverse()
        d_feature_maps = [512, 512, 256, 128, 64]
        d_n_convolutions = n_convolutions
        for decoder_index in range(decoders):
            decoder_block = self.decoder_block(
                db_input, eb_argmax_indices[decoder_index], decoder_index, d_feature_maps[decoder_index], d_n_convolutions[decoder_index])
            db_input = decoder_block

        output = layers.Softmax()(decoder_block)

        self.model = keras.Model(inputs=inputs, outputs=output, name="SegNet")

    def encoder_block(self, x, encoder_index, feature_maps, n_convolutions):
        bank_input = x
        for conv_index in range(n_convolutions):
            bank = self.eb_layers_bank(
                bank_input, conv_index, feature_maps, encoder_index)
            bank_input = bank

        max_pool, indices = MaxPoolingWithArgmax2D(pool_size=(
            2, 2), strides=2, padding='same', name='EB_{}_MPOOL'.format(encoder_index + 1))(bank)

        return max_pool, indices

    def eb_layers_bank(self, x, bank_index, feature_maps, encoder_index):

        bank_input = x

        conv_l = layers.Conv2D(feature_maps, (3, 3), padding='same', name='EB_{}_BANK_{}_CONV'.format(
            encoder_index + 1, bank_index + 1))(bank_input)
        batch_norm = layers.BatchNormalization(
            name='EB_{}_BANK_{}_BN'.format(encoder_index + 1, bank_index + 1))(conv_l)
        relu = layers.ReLU(name='EB_{}_BANK_{}_RL'.format(
            encoder_index + 1, bank_index + 1))(batch_norm)

        return relu

    def decoder_block(self, x, max_pooling_idices, decoder_index, feature_maps, n_convolutions):
        #bank_input = self.unpool_with_argmax(x, max_pooling_idices)
        bank_input = MaxUnpooling2D(name='DB_{}_UPSAMP'.format(decoder_index + 1))([x, max_pooling_idices])
        #bank_input = layers.UpSampling2D()(x)
        for conv_index in range(n_convolutions):
            if conv_index == n_convolutions - 1:
                last_l_banck = True
            else:
                last_l_banck = False
            bank = self.db_layers_bank(
                bank_input, conv_index, feature_maps, decoder_index, last_l_banck)
            bank_input = bank

        return bank

    def db_layers_bank(self, x, bank_index, feature_maps, decoder_index, last_l_bank):
        bank_input = x

        if (last_l_bank) & (decoder_index == 4):
            conv_l = layers.Conv2D(self.classes, (1, 1), padding='same', name='DB_{}_BANK_{}_CONV'.format(
                decoder_index + 1, bank_index + 1))(bank_input)
            #batch_norm = layers.BatchNormalization(
            #    name='DB_{}_BANK_{}_BN'.format(decoder_index + 1, bank_index + 1))(conv_l)
            return conv_l
        else:

            if (last_l_bank) & (decoder_index > 0):
                conv_l = layers.Conv2D(int(feature_maps / 2), (3, 3), padding='same', name='DB_{}_BANK_{}_CONV'.format(
                    decoder_index + 1, bank_index + 1))(bank_input)
            else:
                conv_l = layers.Conv2D(feature_maps, (3, 3), padding='same', name='DB_{}_BANK_{}_CONV'.format(
                    decoder_index + 1, bank_index + 1))(bank_input)
            batch_norm = layers.BatchNormalization(
                name='DB_{}_BANK_{}_BN'.format(decoder_index + 1, bank_index + 1))(conv_l)
            relu = layers.ReLU(name='DB_{}_BANK_{}_RL'.format(
                decoder_index + 1, bank_index + 1))(batch_norm)

            return relu

    def get_model(self):
        return self.model

这里是 model.summary() 的输出.

Model: "SegNet"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
Input (InputLayer)              [(None, 416, 416, 3) 0                                            
__________________________________________________________________________________________________
EB_1_BANK_1_CONV (Conv2D)       (None, 416, 416, 64) 1792        Input[0][0]                      
__________________________________________________________________________________________________
EB_1_BANK_1_BN (BatchNormalizat (None, 416, 416, 64) 256         EB_1_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
EB_1_BANK_1_RL (ReLU)           (None, 416, 416, 64) 0           EB_1_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
EB_1_BANK_2_CONV (Conv2D)       (None, 416, 416, 64) 36928       EB_1_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
EB_1_BANK_2_BN (BatchNormalizat (None, 416, 416, 64) 256         EB_1_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
EB_1_BANK_2_RL (ReLU)           (None, 416, 416, 64) 0           EB_1_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
EB_1_MPOOL (MaxPoolingWithArgma ((None, 208, 208, 64 0           EB_1_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
EB_2_BANK_1_CONV (Conv2D)       (None, 208, 208, 128 73856       EB_1_MPOOL[0][0]                 
__________________________________________________________________________________________________
EB_2_BANK_1_BN (BatchNormalizat (None, 208, 208, 128 512         EB_2_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
EB_2_BANK_1_RL (ReLU)           (None, 208, 208, 128 0           EB_2_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
EB_2_BANK_2_CONV (Conv2D)       (None, 208, 208, 128 147584      EB_2_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
EB_2_BANK_2_BN (BatchNormalizat (None, 208, 208, 128 512         EB_2_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
EB_2_BANK_2_RL (ReLU)           (None, 208, 208, 128 0           EB_2_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
EB_2_MPOOL (MaxPoolingWithArgma ((None, 104, 104, 12 0           EB_2_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
EB_3_BANK_1_CONV (Conv2D)       (None, 104, 104, 256 295168      EB_2_MPOOL[0][0]                 
__________________________________________________________________________________________________
EB_3_BANK_1_BN (BatchNormalizat (None, 104, 104, 256 1024        EB_3_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
EB_3_BANK_1_RL (ReLU)           (None, 104, 104, 256 0           EB_3_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
EB_3_BANK_2_CONV (Conv2D)       (None, 104, 104, 256 590080      EB_3_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
EB_3_BANK_2_BN (BatchNormalizat (None, 104, 104, 256 1024        EB_3_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
EB_3_BANK_2_RL (ReLU)           (None, 104, 104, 256 0           EB_3_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
EB_3_BANK_3_CONV (Conv2D)       (None, 104, 104, 256 590080      EB_3_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
EB_3_BANK_3_BN (BatchNormalizat (None, 104, 104, 256 1024        EB_3_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
EB_3_BANK_3_RL (ReLU)           (None, 104, 104, 256 0           EB_3_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
EB_3_MPOOL (MaxPoolingWithArgma ((None, 52, 52, 256) 0           EB_3_BANK_3_RL[0][0]             
__________________________________________________________________________________________________
EB_4_BANK_1_CONV (Conv2D)       (None, 52, 52, 512)  1180160     EB_3_MPOOL[0][0]                 
__________________________________________________________________________________________________
EB_4_BANK_1_BN (BatchNormalizat (None, 52, 52, 512)  2048        EB_4_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
EB_4_BANK_1_RL (ReLU)           (None, 52, 52, 512)  0           EB_4_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
EB_4_BANK_2_CONV (Conv2D)       (None, 52, 52, 512)  2359808     EB_4_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
EB_4_BANK_2_BN (BatchNormalizat (None, 52, 52, 512)  2048        EB_4_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
EB_4_BANK_2_RL (ReLU)           (None, 52, 52, 512)  0           EB_4_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
EB_4_BANK_3_CONV (Conv2D)       (None, 52, 52, 512)  2359808     EB_4_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
EB_4_BANK_3_BN (BatchNormalizat (None, 52, 52, 512)  2048        EB_4_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
EB_4_BANK_3_RL (ReLU)           (None, 52, 52, 512)  0           EB_4_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
EB_4_MPOOL (MaxPoolingWithArgma ((None, 26, 26, 512) 0           EB_4_BANK_3_RL[0][0]             
__________________________________________________________________________________________________
EB_5_BANK_1_CONV (Conv2D)       (None, 26, 26, 512)  2359808     EB_4_MPOOL[0][0]                 
__________________________________________________________________________________________________
EB_5_BANK_1_BN (BatchNormalizat (None, 26, 26, 512)  2048        EB_5_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
EB_5_BANK_1_RL (ReLU)           (None, 26, 26, 512)  0           EB_5_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
EB_5_BANK_2_CONV (Conv2D)       (None, 26, 26, 512)  2359808     EB_5_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
EB_5_BANK_2_BN (BatchNormalizat (None, 26, 26, 512)  2048        EB_5_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
EB_5_BANK_2_RL (ReLU)           (None, 26, 26, 512)  0           EB_5_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
EB_5_BANK_3_CONV (Conv2D)       (None, 26, 26, 512)  2359808     EB_5_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
EB_5_BANK_3_BN (BatchNormalizat (None, 26, 26, 512)  2048        EB_5_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
EB_5_BANK_3_RL (ReLU)           (None, 26, 26, 512)  0           EB_5_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
EB_5_MPOOL (MaxPoolingWithArgma ((None, 13, 13, 512) 0           EB_5_BANK_3_RL[0][0]             
__________________________________________________________________________________________________
DB_1_UPSAMP (MaxUnpooling2D)    (1, 26, 26, 512)     0           EB_5_MPOOL[0][0]                 
                                                                 EB_5_MPOOL[0][1]                 
__________________________________________________________________________________________________
DB_1_BANK_1_CONV (Conv2D)       (1, 26, 26, 512)     2359808     DB_1_UPSAMP[0][0]                
__________________________________________________________________________________________________
DB_1_BANK_1_BN (BatchNormalizat (1, 26, 26, 512)     2048        DB_1_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
DB_1_BANK_1_RL (ReLU)           (1, 26, 26, 512)     0           DB_1_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
DB_1_BANK_2_CONV (Conv2D)       (1, 26, 26, 512)     2359808     DB_1_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
DB_1_BANK_2_BN (BatchNormalizat (1, 26, 26, 512)     2048        DB_1_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
DB_1_BANK_2_RL (ReLU)           (1, 26, 26, 512)     0           DB_1_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
DB_1_BANK_3_CONV (Conv2D)       (1, 26, 26, 512)     2359808     DB_1_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
DB_1_BANK_3_BN (BatchNormalizat (1, 26, 26, 512)     2048        DB_1_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
DB_1_BANK_3_RL (ReLU)           (1, 26, 26, 512)     0           DB_1_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
DB_2_UPSAMP (MaxUnpooling2D)    (1, 52, 52, 512)     0           DB_1_BANK_3_RL[0][0]             
                                                                 EB_4_MPOOL[0][1]                 
__________________________________________________________________________________________________
DB_2_BANK_1_CONV (Conv2D)       (1, 52, 52, 512)     2359808     DB_2_UPSAMP[0][0]                
__________________________________________________________________________________________________
DB_2_BANK_1_BN (BatchNormalizat (1, 52, 52, 512)     2048        DB_2_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
DB_2_BANK_1_RL (ReLU)           (1, 52, 52, 512)     0           DB_2_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
DB_2_BANK_2_CONV (Conv2D)       (1, 52, 52, 512)     2359808     DB_2_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
DB_2_BANK_2_BN (BatchNormalizat (1, 52, 52, 512)     2048        DB_2_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
DB_2_BANK_2_RL (ReLU)           (1, 52, 52, 512)     0           DB_2_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
DB_2_BANK_3_CONV (Conv2D)       (1, 52, 52, 256)     1179904     DB_2_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
DB_2_BANK_3_BN (BatchNormalizat (1, 52, 52, 256)     1024        DB_2_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
DB_2_BANK_3_RL (ReLU)           (1, 52, 52, 256)     0           DB_2_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
DB_3_UPSAMP (MaxUnpooling2D)    (1, 104, 104, 256)   0           DB_2_BANK_3_RL[0][0]             
                                                                 EB_3_MPOOL[0][1]                 
__________________________________________________________________________________________________
DB_3_BANK_1_CONV (Conv2D)       (1, 104, 104, 256)   590080      DB_3_UPSAMP[0][0]                
__________________________________________________________________________________________________
DB_3_BANK_1_BN (BatchNormalizat (1, 104, 104, 256)   1024        DB_3_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
DB_3_BANK_1_RL (ReLU)           (1, 104, 104, 256)   0           DB_3_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
DB_3_BANK_2_CONV (Conv2D)       (1, 104, 104, 256)   590080      DB_3_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
DB_3_BANK_2_BN (BatchNormalizat (1, 104, 104, 256)   1024        DB_3_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
DB_3_BANK_2_RL (ReLU)           (1, 104, 104, 256)   0           DB_3_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
DB_3_BANK_3_CONV (Conv2D)       (1, 104, 104, 128)   295040      DB_3_BANK_2_RL[0][0]             
__________________________________________________________________________________________________
DB_3_BANK_3_BN (BatchNormalizat (1, 104, 104, 128)   512         DB_3_BANK_3_CONV[0][0]           
__________________________________________________________________________________________________
DB_3_BANK_3_RL (ReLU)           (1, 104, 104, 128)   0           DB_3_BANK_3_BN[0][0]             
__________________________________________________________________________________________________
DB_4_UPSAMP (MaxUnpooling2D)    (1, 208, 208, 128)   0           DB_3_BANK_3_RL[0][0]             
                                                                 EB_2_MPOOL[0][1]                 
__________________________________________________________________________________________________
DB_4_BANK_1_CONV (Conv2D)       (1, 208, 208, 128)   147584      DB_4_UPSAMP[0][0]                
__________________________________________________________________________________________________
DB_4_BANK_1_BN (BatchNormalizat (1, 208, 208, 128)   512         DB_4_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
DB_4_BANK_1_RL (ReLU)           (1, 208, 208, 128)   0           DB_4_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
DB_4_BANK_2_CONV (Conv2D)       (1, 208, 208, 64)    73792       DB_4_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
DB_4_BANK_2_BN (BatchNormalizat (1, 208, 208, 64)    256         DB_4_BANK_2_CONV[0][0]           
__________________________________________________________________________________________________
DB_4_BANK_2_RL (ReLU)           (1, 208, 208, 64)    0           DB_4_BANK_2_BN[0][0]             
__________________________________________________________________________________________________
DB_5_UPSAMP (MaxUnpooling2D)    (1, 416, 416, 64)    0           DB_4_BANK_2_RL[0][0]             
                                                                 EB_1_MPOOL[0][1]                 
__________________________________________________________________________________________________
DB_5_BANK_1_CONV (Conv2D)       (1, 416, 416, 64)    36928       DB_5_UPSAMP[0][0]                
__________________________________________________________________________________________________
DB_5_BANK_1_BN (BatchNormalizat (1, 416, 416, 64)    256         DB_5_BANK_1_CONV[0][0]           
__________________________________________________________________________________________________
DB_5_BANK_1_RL (ReLU)           (1, 416, 416, 64)    0           DB_5_BANK_1_BN[0][0]             
__________________________________________________________________________________________________
DB_5_BANK_2_CONV (Conv2D)       (1, 416, 416, 3)     195         DB_5_BANK_1_RL[0][0]             
__________________________________________________________________________________________________
softmax (Softmax)               (1, 416, 416, 3)     0           DB_5_BANK_2_CONV[0][0]           
==================================================================================================
Total params: 29,459,075
Trainable params: 29,443,203
Non-trainable params: 15,872
__________________________________________________________________________________________________

如您所见，我被迫在 MaxUnpooling2D 中指定批量大小，否则我会收到无法执行操作的错误，因为有 None值和形状无法正确转换。

当我尝试预测图像时，我不得不指定正确的批处理尺寸，否则会出现如下错误:

InvalidArgumentError:  Shapes of all inputs must match: values[0].shape = [4,208,208,64] != values[1].shape = [1,208,208,64]
     [[{{node SegNet/DB_5_UPSAMP/PartitionedCall/PartitionedCall/DB_5_UPSAMP/stack}}]] [Op:__inference_predict_function_70839]

这是由从最大池操作中解开索引所需的实现引起的。

训练图

这是 20 个 epoch 训练的引用。

如您所见，MeanIoU 指标是线性的，除了 epoch 1 之外没有任何进展，没有更新。

另一个指标工作正常，损失正确减少。

––––––––––

结论

有更好的方法，与最新版本的 TF 更兼容，使用最大池操作中的索引实现解散和上采样？

如果实现是正确的，为什么我的指标会停留在特定的值上？我在模型中做错了吗？

谢谢!

最佳答案

您可以通过两种方式在自定义层中进行批量大小未知的 reshape 。

如果您知道形状的其余部分，请使用 -1 作为批量大小进行整形:

假设您知道预期数组的大小:

import tensorflow.keras.backend as K
reshaped = K.reshape(original, (-1, x, y, channels))

假设你不知道大小，那么使用 K.shape将形状作为张量:

inputs_shape = K.shape(inputs)
batch_size = inputs_shape[:1]
x = inputs_shape[1:2]
y = inputs_shape[2:3]
ch = inputs_shape[3:]

#you can then concatenate these and operate them (notice I kept them as 1D vector, not as scalar)
newShape = K.concatenate([batch_size, x, y, ch]) #of course you will make your operations

一旦我做了自己的 Segnet 版本，我就没有使用索引，而是保留了一个热门版本。确实需要额外的操作，但它可能运行良好:

def get_indices(original, unpooled):
    is_equal = K.equal(original, unpooled)
    return K.cast(is_equal, K.floatx())

previous_output = ...
pooled = MaxPooling2D()(previous_output)
unpooled = UpSampling2D()(pooled)

one_hot_indices = Lambda(get_indices)([previous_output, unpooled])

然后在上采样之后，我连接这些索引并传递一个新的 conv:

some_output = ...
upsampled = UpSampling2D()(some_output)
with_indices = Concatenate([upsampled, one_hot_indices])
upsampled = Conv2D(...)(with_indices)

关于tensorflow - TF2.1 : SegNet model architecture problem. 度量计算错误，保持不变并收敛到确定的值，我们在Stack Overflow上找到一个类似的问题： https://stackoverflow.com/questions/61099167/

文章推荐： android - 无法执行 apkanalyzer

文章推荐： Tensorflow 服务性能与直接推理相比非常慢

文章推荐： javascript - 从 PanResponder 内部更新状态

文章推荐： r - 在 R 中生成综合报告格式

ssas - 如何根据当前行和切片器上下文编写 DAX 度量？
背景信息:对于国际销售表中的每一行，我需要检索过去特定日期的美元汇率，以便分析人员确定汇率变化的影响关于销售数字。然后，我将使用今天的汇率与过去的汇率之间的差值，并将其乘以销售额来确定影响。实际问题
powerbi - 根据切片器值选择 DAX 度量
是否可以通过切片器值动态选取表中定义的适当 DAX 度量？源表: +----------------+------------+ | col1 | col2 | +-
未显示子项的 Android ViewFlipper 度量
我有一个 ViewFlipper在我的主要 Activity View 上。在 onCreate 我实例化添加到 ViewFlipper 的 View 。之后，我将显示的 child 设置为第一个。当
python - 如何计算多标签分类中的 F1 度量？
我正在研究句子类别检测问题。每个句子可以属于多个类别例如: "It has great sushi and even better service." True Label: [[ 0. 0.
cluster-analysis - 计算聚类的 F 度量
谁能帮我一起计算F-measure？我知道如何计算召回率和准确率，但不知道对于给定的算法如何计算一个 F-measure 值。例如，假设我的算法创建了 m 个集群，但我知道相同数据有 n 个集群(由
machine-learning - 为什么不能从宏精度和召回率计算宏 F1 度量？
我对通过宏精度和手动召回计算宏 f1-score 感兴趣。但结果并不相等。代码中 f1 和 f1_new 的最终公式有什么区别？ from sklearn.metrics import precisi
select - 基于最后一列值和筛选器的 Power BI 度量
我有一张记录了一些人体重的表格: Year Person Weight 2010 Mike 75 2010 Laura 60 2011 Mike 80 201
train() 中的 ROC 度量，插入符包
df分为训练数据帧和测试数据帧。训练数据帧分为训练数据帧和测试数据帧。因变量Y是二进制(因子)，值为 0 和 1。我试图用此代码(神经网络，插入符号包)预测概率: library(caret) mod
python - 如何在 Keras 中使用 Hausdorff 度量？
我想使用 Hausdorff 距离作为训练指标，但我刚刚找到了 Weighted_Hausdorff_loss并将其用作医学图像分割的指标。 import math import numpy as n
r - R 中的精度、召回率和 f 度量
我有一段时间没有使用 R，所以也许我只是不习惯它，但是..我在 R 中有一个表，有两个列，第一个有预测值(值可以是 0 或 1 )，第二个具有实际值(也是 0 或 1)。我需要找到召回率、精度和 f
metrics - 使用 Statsd 聚合 Collectd 度量
我正在使用 Collectd 收集系统指标。我正在小范围内收集测量值以获得准确的值。但是我想使用 Statsd 在本地聚合这些值。 Statsd 应该聚合这些值并以更长的时间间隔将它们发送到 libr
machine-learning - 在多类分类中，如何计算每个类的 F1 度量？
我使用SciKit作为一个库来处理分类算法，例如:NB、SVM。这是一个非常漂亮的binary classification implementation对于“垃圾邮件和HAM”电子邮件:
algorithm - 度量 TSP 的 MST 启发式的严格示例
我正在寻找 MST 启发式算法的严格示例，它是度量旅行商问题的 2 近似算法。这个算法在网上很容易找到，但我找不到具体的例子。我所说的严格示例是指给定算法返回的解决方案比最佳解决方案差 2 倍的示例
looker-studio - Data Studio 案例函数度量/维度错误
我使用 Data Studio 中的 Case 函数来确定某个值是否高于或低于 6,000 英镑，并根据输出呈现两个数字之一。这两个数字是计算字段。第一个案例陈述: (大于或小于)- CASE WH
rrdtool - Ganglia/RRD 工具中的 load_one 度量 Y 轴是什么意思？
我正在使用 Ganglia + RRDTool为 monitoring a web farm .很多图很清楚，但是当我看到load_one metric , 我 don't have Y-axis l
list - Power BI - 用于引用值列表的 DAX 度量。计算每个经纪人的净营业额 - 更优雅的解决方案？
以下是股票交易数据的简化版本。 StockData = DATATABLE ( "STOCK", STRING, "Date", DATETIME, "Buyer", STRI
ASP.NET Core 7: using OpenTelemetry to send traces/metrics to Grafana(ASP.NET Core 7：使用开放遥测向Grafana发送跟踪/度量)
我正在尝试将ASP.NET Core 7应用程序中的度量/跟踪发送到Grafana。。这是我的《码头工人》作文文件。。下面是我的收集器配置：。下面是配置OpenTelemeter的服务集合扩展方法。。
ASP.NET Core 7: using OpenTelemetry to send traces/metrics to Grafana(ASP.NET Core 7：使用开放遥测向Grafana发送跟踪/度量)
我正试图从我的ASP.NET Core 7应用程序向Grafana发送度量/跟踪。。这是我的《码头工人》作文文件。。下面是我的收集器配置：。下面是配置OpenTelemeter的服务集合扩展方法。。首

行者123

个人简介

我是一名优秀的程序员,十分优秀！

作者热门文章

滴滴打车优惠券免费领取

全站热门文章

首页

博学

6Ren·AI

商城

tensorflow - TF2.1 : SegNet model architecture problem. 度量计算错误，保持不变并收敛到确定的值