TensorFlow Transfer Learning, Functional API, model.predict returns always same class.

[AdrianKazii]

I have an issue with model.predict. I run feature extraction model, then fine tuning with Functional API. Training looks good, loss and accuracy curves show that model is learning, evaluation in a range of 87-91%. But then when I try to plot the image either with pred_and_plot function from helper_functions from up here https://github.com/mrdbourke/tensorflow-deep-learning/blob/main/extras/helper_functions.py:

from helper_functions import pred_and_plot, load_and_prep_image
pred_and_plot(model, 'cb29a074cf97c5c05924ece2db29a134-recipe_main.jpg',train_data_10_percent.class_names)

or I use custom code:

import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import os
import random

# dataset = input('Type in "test" or "train" to pick dataset of image: ')

folder = '10_food_classes_10_percent'
dataset = 'train'
target_class = random.choice(train_data_10_percent.class_names)

target_dir = os.path.join(folder + '/' + dataset + '/' + target_class + '/')
random_image = random.choice(os.listdir(target_dir))

random_image_path = target_dir + random_image

# Read in the random image
img = mpimg.imread(random_image_path)
resized_img = tf.image.resize(img, [224, 224])
resized_img = resized_img / 255.
pred = model_0.predict(tf.expand_dims(resized_img, axis = 0)) #### STAD BLAD!!!!
predicted_class = train_data_10_percent.class_names[pred.argmax()]
print(pred)
print(pred.argmax())
plt.imshow(resized_img)
plt.title(f"Prediction: {predicted_class}")
plt.axis(False);

For custom code where image is randomly picked I always get 1 class as prediction but always new image on a plot. I've checked the pred, pred.argmax() and I always get almost the same predictions for 10 class classification:

pred.argmax():

[[0.07503171 0.11896327 0.12848163 0.09343361 0.04384859 0.02784183 0.26499084 0.11767029 0.09404188 0.03569638]]

pred:

6

"Almost the same" means that the prediction is always done, because I get slightly different numbers, e.g. instead 0.0751232 I get 0.0703171 in 0 index. Thus, it seems that model constantly predicts one image and display another one. Did anyone have the same issue, is it something with my code, Functional API or...? Just to add I used transfer learning previously with Sequential and didn't have that issue.

[nazzal5448]

having the same issue with EfficentNetB0. the metrices are good but still it predicts meso_soup everytime