C3W3_Assignment Model Structure

Course Q&A TensorFlow: Advanced Techniques Specialization Advanced Computer Vision with TensorFlow
1

I did not get the same number of parameters in the FCN8 model. I don’t know whether any layer is set wrong. Could anybody help me check it?

The output shape are the same, but the parameter numbers do not match the expected ones.

 Layer (type)                Output Shape              Param #   
=================================================================
 input_10 (InputLayer)       [(None, 64, 84, 1)]       0         
                                                                 
 zero_padding2d_8 (ZeroPaddi  (None, 64, 96, 1)        0         
 ng2D)                                                           
                                                                 
 conv2d_74 (Conv2D)          (None, 64, 96, 32)        224       
                                                                 
 leaky_re_lu_74 (LeakyReLU)  (None, 64, 96, 32)        0         
                                                                 
 conv2d_75 (Conv2D)          (None, 64, 96, 32)        6176      
                                                                 
 leaky_re_lu_75 (LeakyReLU)  (None, 64, 96, 32)        0         
                                                                 
 max_pooling2d_37 (MaxPoolin  (None, 32, 48, 32)       0         
 g2D)                                                            
                                                                 
 batch_normalization_37 (Bat  (None, 32, 48, 32)       128       
 chNormalization)                                                
                                                                 
 conv2d_76 (Conv2D)          (None, 32, 48, 64)        12352     
                                                                 
 leaky_re_lu_76 (LeakyReLU)  (None, 32, 48, 64)        0         
                                                                 
 conv2d_77 (Conv2D)          (None, 32, 48, 64)        24640     
                                                                 
 leaky_re_lu_77 (LeakyReLU)  (None, 32, 48, 64)        0         
                                                                 
 max_pooling2d_38 (MaxPoolin  (None, 16, 24, 64)       0         
 g2D)                                                            
                                                                 
 batch_normalization_38 (Bat  (None, 16, 24, 64)       256       
 chNormalization)                                                
                                                                 
 conv2d_78 (Conv2D)          (None, 16, 24, 128)       49280     
                                                                 
 leaky_re_lu_78 (LeakyReLU)  (None, 16, 24, 128)       0         
                                                                 
 conv2d_79 (Conv2D)          (None, 16, 24, 128)       98432     
                                                                 
 leaky_re_lu_79 (LeakyReLU)  (None, 16, 24, 128)       0         
                                                                 
 max_pooling2d_39 (MaxPoolin  (None, 8, 12, 128)       0         
 g2D)                                                            
                                                                 
 batch_normalization_39 (Bat  (None, 8, 12, 128)       512       
 chNormalization)                                                
                                                                 
 conv2d_80 (Conv2D)          (None, 8, 12, 256)        196864    
                                                                 
 leaky_re_lu_80 (LeakyReLU)  (None, 8, 12, 256)        0         
                                                                 
 conv2d_81 (Conv2D)          (None, 8, 12, 256)        393472    
                                                                 
 leaky_re_lu_81 (LeakyReLU)  (None, 8, 12, 256)        0         
                                                                 
 max_pooling2d_40 (MaxPoolin  (None, 4, 6, 256)        0         
 g2D)                                                            
                                                                 
 batch_normalization_40 (Bat  (None, 4, 6, 256)        1024      
 chNormalization)                                                
                                                                 
 conv2d_82 (Conv2D)          (None, 4, 6, 256)         393472    
                                                                 
 leaky_re_lu_82 (LeakyReLU)  (None, 4, 6, 256)         0         
                                                                 
 conv2d_83 (Conv2D)          (None, 4, 6, 256)         393472    
                                                                 
 leaky_re_lu_83 (LeakyReLU)  (None, 4, 6, 256)         0         
                                                                 
 max_pooling2d_41 (MaxPoolin  (None, 2, 3, 256)        0         
 g2D)                                                            
                                                                 
 batch_normalization_41 (Bat  (None, 2, 3, 256)        1024      
 chNormalization)                                                
                                                                 
=================================================================
Total params: 1,571,328
Trainable params: 1,569,856
Non-trainable params: 1,472

Besides, I tried to change the kernel_size from the Conv2DTranspose in the fcn8_decoder function. However, if I change it to a non-square shape then the model will not work. I might not quite understand how the kernel size being calculated here. Could anybody tell me how you set the kernel_size here? Thank you.

2

Ok let me give you some guidelines:

First you should not post your code its against the rules.

If you dont get the same number of parameteres then surely there is a problem on your implementation, so you need to recheck it. You need to see how many arguments are in the conv_block and call that function properly.

If you dont understand what kernel size is you should google: conv2d, tensorflow, it will give you the details about that tensorflow function.

3

Okay, now I know that I misunderstood the hint above the code block. If I had used the parameters in the ungraded lab, the error would not have occurred. But it is good that I could understand how it works now. Since the code works, I remove the code in the question above. Thank you.

1 Like
4

hey m having the same issue, I checked the ungraded lab also, but that is not giving the desired expected output. can you tell me in the first ungraded lab, I used this for the 5 blocks, am I doing right?

x = block(x,n_convs=2, filters=128, kernel_size=(3,3), activation=‘relu’,pool_size=(2,2), pool_stride=(2,2), block_name=‘block2’)
p2 = x
p.s. this is not the assignment code(it is only for reference)

thank you
DP