1
$\begingroup$

I am trying to build face landmark detection model using simple regression.I used celeba dataset which has 5 points hence 10 output units.I used grayscale and normalized image as input. Here is my model

    self.model = models.Sequential()
    self.model.add(layers.Conv2D(32, (4, 4), activation='relu', input_shape=(218, 178,1)))#l1
    self.model.add(layers.Conv2D(64, (4, 4), activation='relu'))#l2
    self.model.add(layers.Conv2D(32, (4, 4), activation='relu'))#l3
    self.model.add(layers.Conv2D(16, (4, 4), activation='relu'))#l4
    self.model.add(layers.Conv2D(16, (3, 3), activation='relu'))
    self.model.add(layers.Conv2D(16, (3, 3), activation='relu'))
    self.model.add(layers.Conv2D(32, (3, 3), activation='relu'))
    self.model.add(layers.Conv2D(64, (3, 3), activation='relu'))
    self.model.add(layers.Conv2D(128, (3, 3), activation='relu'))
    self.model.add(layers.AveragePooling2D((3, 3)))
    self.model.add(layers.Flatten())
    self.model.add(layers.Dense(16))
    self.model.add(layers.Dense(10)) 

Here is my loss function

model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001),loss='mse')

Here loss is stuck around 18 and doesn't go below that. I tried various configuration of CNN architecture like adding and removing layers and I also tried with changing learning rate but no use.

Please anyone point me in right direction. How can I debug this network. (For this I only used first 100 images of dataset)

$\endgroup$
1
  • $\begingroup$ Make the network bigger i.e. try 64 throughout. Also, the last Dense of 16 is too less. First, overfit the train set. Preprocess the image too. $\endgroup$
    – 10xAI
    Apr 20 at 13:23
0
$\begingroup$

Use a standard pretrained network like ResNet, Effnet or even VGG16 for feature extraction and put a regression layer in the end.

Now, talking about your network. FIrst of all, the image size is not square, make it! Don't use even valued kernels (3 and 5 are mostly used). Number of filters increase with depth, yours don't follow a pattern.

Lastly, 100 images are nothing for these behemoths, give it more data.

Better stick to a pretrained model.

$\endgroup$
3
  • $\begingroup$ I am doing this for practice not for any actual implementation. So I want to do it from scratch $\endgroup$
    – LOLs
    Apr 20 at 12:27
  • $\begingroup$ And also making it square made worse $\endgroup$
    – LOLs
    Apr 20 at 12:59
  • $\begingroup$ Make smallest side 256 and take 224x224 center crop. That's how you need to do it. Also, did you correct your network first. Correct that then make observations about the images. $\endgroup$ Apr 20 at 13:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.