👀 Views: 3 💬 Answers: 1 📅 Created: 2025-06-10

tensorflow machine-learning model-prediction python

I've encountered a strange issue with I'm trying to debug I'm learning this framework and I'm dealing with Hey everyone, I'm running into an issue that's driving me crazy. I'm currently working on a binary classification question using TensorFlow 2.8, and I’m running into an scenario where the model's predictions seem completely off. I’ve trained a simple CNN model on a dataset of 10,000 images. The training seemed to go well with an accuracy of around 95% on the validation set. However, when I try to use `model.predict()` on a batch of new images, I’m getting results that don’t match my expectations – predictions that are all around 0.5 for positive and negative classes. Here’s a snippet of my code where I load the images and call `predict`: ```python import numpy as np import tensorflow as tf from tensorflow.keras.preprocessing import image # Load and preprocess new images new_images = [] for img_path in ['path/to/image1.jpg', 'path/to/image2.jpg']: img = image.load_img(img_path, target_size=(128, 128)) img_array = image.img_to_array(img) new_images.append(img_array) new_images = np.array(new_images) / 255.0 # Normalizing # Predicting predictions = model.predict(new_images) print(predictions) ``` I have ensured that the images are preprocessed in the same way as the training images. I even checked the shapes and types of the input images, and they all seem correct. The model architecture is also quite simple: ```python model = tf.keras.Sequential([ tf.keras.layers.Conv2D(32, (3, 3), activation='relu', input_shape=(128, 128, 3)), tf.keras.layers.MaxPooling2D(pool_size=(2, 2)), tf.keras.layers.Flatten(), tf.keras.layers.Dense(1, activation='sigmoid') ]) model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy']) ``` Additionally, I’m not getting any errors, but the output predictions seem suspiciously similar for different images. I’ve tried loading images with different formats (PNG, JPEG) and made sure they are not corrupted. Could this be an scenario with the model’s learning capacity, or might I be missing something crucial in the prediction phase? Any insights would be immensely helpful! I'm working on a web app that needs to handle this. Is there a better approach? Thanks for taking the time to read this! Has anyone dealt with something similar? What are your experiences with this? I'd love to hear your thoughts on this. The project is a CLI tool built with Python.