How to predict on directory of images

Part 1 (2018)
1

If anyone has a code snippet on how to predict on a full directory of images please post it here.

I have a learner that is trained and I want to predict an a large set of images.

I’m looking for code snippets to do this.
I’ll work on it in the meantime as well starting from the posts on predicting on a single image.

At the same time i’m wondering why we need to apply transforms at all to images we want to predict on? I could understand if we are doing TTA but why just blindly apply transforms?

Thanks

2

Got this mostly working:

def load_images(directory):
    images = []
    image_names = glob.glob(os.path.join(directory, "*.jpg"))
    for filename in image_names:
        images.append(scipy.misc.imread(filename))
    images = np.array(images)
    return images


def predict_directory(learn, input_directory, val_tfms):
    learn.precompute = False
    images = load_images(input_directory)
    image_names = glob.glob(os.path.join(input_directory, "*.jpg"))
    for index, input_image in enumerate(images):
        transformed_image = val_tfms(input_image)
        pred = to_np(learn.models.model(V(T(transformed_image[None]).cuda())))
        print(pred, np.exp(pred), image_names[index])
        if np.exp(pred)[1][1] > 0.5:
            plt.figure()
            plt.imshow(images[index])
    print(data.classes)
3

Not working yet as my validation set doesn’t get predicted correctly even though I have 91% accuracy…