I've been trying to deploy a very simple toy Keras model to Cloud Functions, which would predict the class of an image, but for reasons unknown, when the execution gets to the predict method, it gets stuck, does not throw any error, and eventually times out. import functions_framework import io import numpy as np import tensorflow as tf from tensorflow.keras.models import load_model from PIL import Image model = load_model("gs://<my-bucket>/cifar10_model.keras") class_names = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'] def preprocess_image(image_file): img = Image.open(io.BytesIO(image_file.read())) img = img.resize((32, 32)) img = np.array(img) img = img / 255.0 img = img.reshape(1, 32, 32, 3) return img @functions_framework.http def predict(request): image = preprocess_image(request.files['image_file']) print(image.shape) # this prints OK prediction = model.predict(image) print(prediction) # this never prints predicted_class = class_names[np.argmax(prediction)] return f"Predicted class: {predicted_class}" Debugging locally works fine, the prediction is fast as expected (the model weights file is 2MB). I also added several prints along the way (removed from the snippet above) and the execution works fine until the predict method. Even though the minimal compute configuration should work, I tried reserving more memory and CPU, but nothing worked. The model is hosted at Storage, I tried downloading it first, but that didn't work either. I did also try making the prediction inside a tf.device('/cpu:0') context, passing a step=1 parameter and converting the image array to a Keras Dataset first, as suggested by ChatGPT, with the same results. Actually, nothing prints as a result of invoking predict at all. Calling call instead of predict got me nowhere. What am I missing? Continue reading...
