Training loss is huge

[Primadual]

Hi, Thanks for the code. I am training the model using a small subset of the CASIA-WebFace dataset that was originally used for Facenet. However, the training loss after 100 epochs only comes down to 14.09 from 21! I am using your train code as is and the dataset as is. Do you know if I need to preprocess the data or something? Also, when I try to predict two different face images I get very similar values. Here is the code I use for predicting:

import tensorflow as tf
from PIL import Image
from tensorflow import keras
import numpy as np

gpus = tf.config.experimental.list_physical_devices(device_type='GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)
    
model = tf.keras.models.load_model('saved_model/FaceVerify')

# Check its architecture
model.summary()


# Load images
image_1 = "img/2_001.jpg"
try:
    image_1 = Image.open(image_1)
except:
    print('Image_1 Open Error! Try again!')

image_2 = "img/1_002.jpg"
try:
    image_2 = Image.open(image_2)
except:
    print('Image_2 Open Error! Try again!')


def letterbox_image(image, size):
        if input_shape[-1] == 1:
            image = image.convert("RGB")
        iw, ih  = image.size
        w, h    = size
        scale   = min(w/iw, h/ih)
        nw      = int(iw*scale)
        nh      = int(ih*scale)

        image   = image.resize((nw,nh), Image.BICUBIC)
        new_image = Image.new('RGB', size, (128,128,128))
        new_image.paste(image, ((w-nw)//2, (h-nh)//2))
        if input_shape[-1] == 1:
            new_image = new_image.convert("L")
        return new_image


# Resize the input image without distortion
input_shape = [250, 250, 3]
image_1 = letterbox_image(image_1, [input_shape[1], input_shape[0]])
image_2 = letterbox_image(image_2, [input_shape[1], input_shape[0]])

# Normalize the picture
image_1 = np.asarray(image_1).astype(np.float64) / 255
image_2 = np.asarray(image_2).astype(np.float64) / 255

# Add the batch dimension before it can be put into the network to predict
photo1 = np.expand_dims(image_1, 0)
photo2 = np.expand_dims(image_2, 0)

# The picture is transmitted to the network for prediction
output1 = model.predict(photo1)
output2 = model.predict(photo2)

# Calculate the distance between the two
l1 = np.sqrt(np.sum(np.square(output1 - output2), axis=-1))
print("probability = %5.3f" % l1)

[harxish]

How small is the subset ?