'Conv2DSlowBackpropInput: Size of out_backprop doesn't match computed but network dimensions work
I am working with a segmentation problem and I am using an encoder-decoder architecture that uses dense blocks (https://arxiv.org/abs/1611.09326). I have rectangular images (1024,224,1). The custom generators make images and labels be [batch size, 1024,224,1]. When I load the architecture, the summary displays correct dimension both in the downsampling and upsampling. I have dome the math of the network manually as well and everything seems correct. However, when I use the model.fit method, it displays the following error:
InvalidArgumentError: Conv2DSlowBackpropInput: Size of out_backprop doesn't match computed: actual = 70, computed = 14 spatial_dim: 3 input: 28 filter: 3 output: 70 stride: 2 dilation: 1
Which makes no sense since the summary displays correct dimension. Could someone help me with this? Thank you!
Model: "fcn-densenet"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_4 (InputLayer) [(None, 1024, 224, 0 []
1)]
initial_conv2D (Conv2D) (None, 1024, 224, 4 432 ['input_4[0][0]']
8)
initial_bn (BatchNormalization (None, 1024, 224, 4 192 ['initial_conv2D[0][0]']
) 8)
activation_117 (Activation) (None, 1024, 224, 4 0 ['initial_bn[0][0]']
8)
dense_0_0_bn (BatchNormalizati (None, 1024, 224, 4 192 ['activation_117[0][0]']
on) 8)
activation_118 (Activation) (None, 1024, 224, 4 0 ['dense_0_0_bn[0][0]']
8)
dense_0_0_conv2D (Conv2D) (None, 1024, 224, 1 6912 ['activation_118[0][0]']
6)
dropout_100 (Dropout) (None, 1024, 224, 1 0 ['dense_0_0_conv2D[0][0]']
6)
concatenate_122 (Concatenate) (None, 1024, 224, 6 0 ['activation_117[0][0]',
4) 'dropout_100[0][0]']
dense_0_1_bn (BatchNormalizati (None, 1024, 224, 6 256 ['concatenate_122[0][0]']
on) 4)
activation_119 (Activation) (None, 1024, 224, 6 0 ['dense_0_1_bn[0][0]']
4)
dense_0_1_conv2D (Conv2D) (None, 1024, 224, 1 9216 ['activation_119[0][0]']
6)
dropout_101 (Dropout) (None, 1024, 224, 1 0 ['dense_0_1_conv2D[0][0]']
6)
concatenate_123 (Concatenate) (None, 1024, 224, 8 0 ['concatenate_122[0][0]',
0) 'dropout_101[0][0]']
tr_0_bn (BatchNormalization) (None, 1024, 224, 8 320 ['concatenate_123[0][0]']
0)
activation_120 (Activation) (None, 1024, 224, 8 0 ['tr_0_bn[0][0]']
0)
tr_0_conv2D (Conv2D) (None, 1024, 224, 8 6400 ['activation_120[0][0]']
0)
max_pooling2d_11 (MaxPooling2D (None, 512, 112, 80 0 ['tr_0_conv2D[0][0]']
) )
dense_1_0_bn (BatchNormalizati (None, 512, 112, 80 320 ['max_pooling2d_11[0][0]']
on) )
activation_121 (Activation) (None, 512, 112, 80 0 ['dense_1_0_bn[0][0]']
)
dense_1_0_conv2D (Conv2D) (None, 512, 112, 16 11520 ['activation_121[0][0]']
)
dropout_102 (Dropout) (None, 512, 112, 16 0 ['dense_1_0_conv2D[0][0]']
)
concatenate_124 (Concatenate) (None, 512, 112, 96 0 ['max_pooling2d_11[0][0]',
) 'dropout_102[0][0]']
dense_1_1_bn (BatchNormalizati (None, 512, 112, 96 384 ['concatenate_124[0][0]']
on) )
activation_122 (Activation) (None, 512, 112, 96 0 ['dense_1_1_bn[0][0]']
)
dense_1_1_conv2D (Conv2D) (None, 512, 112, 16 13824 ['activation_122[0][0]']
)
dropout_103 (Dropout) (None, 512, 112, 16 0 ['dense_1_1_conv2D[0][0]']
)
concatenate_125 (Concatenate) (None, 512, 112, 11 0 ['concatenate_124[0][0]',
2) 'dropout_103[0][0]']
tr_1_bn (BatchNormalization) (None, 512, 112, 11 448 ['concatenate_125[0][0]']
2)
activation_123 (Activation) (None, 512, 112, 11 0 ['tr_1_bn[0][0]']
2)
tr_1_conv2D (Conv2D) (None, 512, 112, 11 12544 ['activation_123[0][0]']
2)
max_pooling2d_12 (MaxPooling2D (None, 256, 56, 112 0 ['tr_1_conv2D[0][0]']
) )
dense_2_0_bn (BatchNormalizati (None, 256, 56, 112 448 ['max_pooling2d_12[0][0]']
on) )
activation_124 (Activation) (None, 256, 56, 112 0 ['dense_2_0_bn[0][0]']
)
dense_2_0_conv2D (Conv2D) (None, 256, 56, 16) 16128 ['activation_124[0][0]']
dropout_104 (Dropout) (None, 256, 56, 16) 0 ['dense_2_0_conv2D[0][0]']
concatenate_126 (Concatenate) (None, 256, 56, 128 0 ['max_pooling2d_12[0][0]',
) 'dropout_104[0][0]']
dense_2_1_bn (BatchNormalizati (None, 256, 56, 128 512 ['concatenate_126[0][0]']
on) )
activation_125 (Activation) (None, 256, 56, 128 0 ['dense_2_1_bn[0][0]']
)
dense_2_1_conv2D (Conv2D) (None, 256, 56, 16) 18432 ['activation_125[0][0]']
dropout_105 (Dropout) (None, 256, 56, 16) 0 ['dense_2_1_conv2D[0][0]']
concatenate_128 (Concatenate) (None, 256, 56, 32) 0 ['dropout_104[0][0]',
'dropout_105[0][0]']
tr_up_0_conv2DT (Conv2DTranspo (None, 512, 112, 32 9248 ['concatenate_128[0][0]']
se) )
concatenate_129 (Concatenate) (None, 512, 112, 14 0 ['tr_up_0_conv2DT[0][0]',
4) 'concatenate_125[0][0]']
dense_3_0_bn (BatchNormalizati (None, 512, 112, 14 576 ['concatenate_129[0][0]']
on) 4)
activation_126 (Activation) (None, 512, 112, 14 0 ['dense_3_0_bn[0][0]']
4)
dense_3_0_conv2D (Conv2D) (None, 512, 112, 16 20736 ['activation_126[0][0]']
)
dropout_106 (Dropout) (None, 512, 112, 16 0 ['dense_3_0_conv2D[0][0]']
)
concatenate_130 (Concatenate) (None, 512, 112, 16 0 ['concatenate_129[0][0]',
0) 'dropout_106[0][0]']
dense_3_1_bn (BatchNormalizati (None, 512, 112, 16 640 ['concatenate_130[0][0]']
on) 0)
activation_127 (Activation) (None, 512, 112, 16 0 ['dense_3_1_bn[0][0]']
0)
dense_3_1_conv2D (Conv2D) (None, 512, 112, 16 23040 ['activation_127[0][0]']
)
dropout_107 (Dropout) (None, 512, 112, 16 0 ['dense_3_1_conv2D[0][0]']
)
concatenate_132 (Concatenate) (None, 512, 112, 32 0 ['dropout_106[0][0]',
) 'dropout_107[0][0]']
tr_up_1_conv2DT (Conv2DTranspo (None, 1024, 224, 3 9248 ['concatenate_132[0][0]']
se) 2)
concatenate_133 (Concatenate) (None, 1024, 224, 1 0 ['tr_up_1_conv2DT[0][0]',
12) 'concatenate_123[0][0]']
dense_4_0_bn (BatchNormalizati (None, 1024, 224, 1 448 ['concatenate_133[0][0]']
on) 12)
activation_128 (Activation) (None, 1024, 224, 1 0 ['dense_4_0_bn[0][0]']
12)
dense_4_0_conv2D (Conv2D) (None, 1024, 224, 1 16128 ['activation_128[0][0]']
6)
dropout_108 (Dropout) (None, 1024, 224, 1 0 ['dense_4_0_conv2D[0][0]']
6)
concatenate_134 (Concatenate) (None, 1024, 224, 1 0 ['concatenate_133[0][0]',
28) 'dropout_108[0][0]']
dense_4_1_bn (BatchNormalizati (None, 1024, 224, 1 512 ['concatenate_134[0][0]']
on) 28)
activation_129 (Activation) (None, 1024, 224, 1 0 ['dense_4_1_bn[0][0]']
28)
dense_4_1_conv2D (Conv2D) (None, 1024, 224, 1 18432 ['activation_129[0][0]']
6)
dropout_109 (Dropout) (None, 1024, 224, 1 0 ['dense_4_1_conv2D[0][0]']
6)
concatenate_135 (Concatenate) (None, 1024, 224, 1 0 ['concatenate_134[0][0]',
44) 'dropout_109[0][0]']
conv2d_3 (Conv2D) (None, 1024, 224, 3 432 ['concatenate_135[0][0]']
)
reshape_6 (Reshape) (None, 229376, 3) 0 ['conv2d_3[0][0]']
activation_130 (Activation) (None, 229376, 3) 0 ['reshape_6[0][0]']
reshape_7 (Reshape) (None, 1024, 224, 3 0 ['activation_130[0][0]']
)
==================================================================================================
Total params: 197,920
Trainable params: 195,296
Non-trainable params: 2,624
__________________________________________________________________________________________________
Sources
This article follows the attribution requirements of Stack Overflow and is licensed under CC BY-SA 3.0.
Source: Stack Overflow
| Solution | Source |
|---|