Low accuracy about face-clustering HOT 4 CLOSED

Hi, I've trained my model with Google Colab's GPU for almost 8-10 epochs. I did different trials achieving almost the same accuracy (around 90%).

Anyway, You can try to:

1. Debug your loaded images: check the pixels' values, see if channels are switched. (are you training on CelebA, right?)
2. Pretrain on ImageNet:

def mobilenet_model(num_features):
  base = MobileNetV2(input_shape=IMG_SHAPE,
                     weights='imagenet',     # <- 
                     include_top=False,
                     pooling='avg')
  
  # model top
  x = base.output
  x = Dense(1536, activation='relu')(x)
  x = BatchNormalization()(x)
  x = Dropout(0.3)(x)

  top = Dense(num_features, activation='sigmoid')(x)
  return Model(inputs=base.input, outputs=top)

3. Reduce the strength of data augmentation (or avoid it):

# try with a factor of '0.1' instead of '0.2'
train_datagen = ImageDataGenerator(rotation_range=10, 
                                   rescale=1./255, 
                                   width_shift_range=0.1, 
                                   height_shift_range=0.1, 
                                   shear_range=0.1, 
                                   zoom_range=0.1, 
                                   horizontal_flip=True, 
                                   fill_mode='nearest')

4. If none of the above works, try a different optimizer (e.g. adam) and/or increase the batch_size to 96 or 128.

Hope it helps.

from face-clustering.

Hi, I think that cosine_proximity should be the same as cosine_similarity.

I've noticed you're using keras stuff from tensorflow (I mean tf.keras). So, that's why you cannot find cosine_similarity because it's not defined under the tf.keras namespace. But that's fine.

You can also try to use binary_crossentropy as loss function and see if accuracy improves.

Note: if you import stuff from tf.keras make sure to use it for every keras-related stuff. Because there are some inconsistencies between plain keras imports and tf.keras ones. Also, I'm not aware of the existence of potential performance discrepancies between keras and tf.keras.

from face-clustering.

Yes! Thank you! This was an issue - after installing keras everythng works fine. I had only to replace 'other' with 'raw', and change batch size to 32:

valid_generator = valid_datagen.flow_from_dataframe(
dataframe=valid_split,
directory=celeba.images_folder,
x_col='image_id',
y_col=celeba.features_name,
target_size=TARGET_SIZE,
batch_size=batch_size,
#class_mode='other'
class_mode='raw'
)

It seems that tf.keras and keras works differently.

from face-clustering.

Thank you for the response! :-)

(1) Yes
(2) I have changed weights='imagenet'
(3) I have removed data augmentation:

train_datagen = tf.keras.preprocessing.image.ImageDataGenerator(
rescale=1. / 255,
)

(4) Bigger batch than 64 does not fit into my video memory (as long as I correctly read the exception)

After 32 epoches I obtained 86%.
I run tensorflow 2.1 on GeForce 2600 GTX.

In model.compile I do not have 'cosine_proximity' - this option is unavilable. Instead I use 'cosine_similarity' - is this the same?

from face-clustering.