2021-01-14 08:07:24 +01:00
|
|
|
import os
|
|
|
|
import pydload
|
|
|
|
import numpy as np
|
|
|
|
import tensorflow as tf
|
|
|
|
|
|
|
|
from .image_utils import load_images
|
|
|
|
|
|
|
|
|
|
|
|
class LiteClassifier:
|
|
|
|
def __init__(self):
|
|
|
|
url = "https://github.com/notAI-tech/NudeNet/releases/download/v0/classifier.tflite"
|
|
|
|
home = os.path.expanduser("~")
|
|
|
|
model_folder = os.path.join(home, ".NudeNet/")
|
|
|
|
if not os.path.exists(model_folder):
|
|
|
|
os.mkdir(model_folder)
|
|
|
|
|
|
|
|
model_path = os.path.join(model_folder, "lite_classifier")
|
|
|
|
|
|
|
|
if not os.path.exists(model_path):
|
|
|
|
print("Downloading the checkpoint to", model_path)
|
|
|
|
pydload.dload(url, save_to_path=model_path, max_time=None)
|
|
|
|
|
|
|
|
self.interpreter = tf.lite.Interpreter(model_path=model_path)
|
|
|
|
self.interpreter.allocate_tensors()
|
|
|
|
|
|
|
|
def classify(self, image_paths, size=(256, 256)):
|
|
|
|
if isinstance(image_paths, str):
|
|
|
|
image_paths = [image_paths]
|
|
|
|
|
|
|
|
input_details = self.interpreter.get_input_details()
|
|
|
|
output_details = self.interpreter.get_output_details()
|
|
|
|
|
|
|
|
loaded_images, _ = load_images(image_paths, size, image_paths)
|
|
|
|
|
|
|
|
result = {}
|
|
|
|
for image_path, img in zip(image_paths, loaded_images):
|
|
|
|
img = np.expand_dims(img, axis=0)
|
|
|
|
input_data = np.array(img, dtype=np.float32)
|
|
|
|
self.interpreter.set_tensor(input_details[0]["index"], input_data)
|
|
|
|
|
|
|
|
self.interpreter.invoke()
|
|
|
|
|
|
|
|
# The function `get_tensor()` returns a copy of the tensor data.
|
|
|
|
# Use `tensor()` in order to get a pointer to the tensor.
|
|
|
|
output_data = self.interpreter.get_tensor(output_details[0]["index"])
|
|
|
|
result[image_path] = {
|
|
|
|
"unsafe": output_data[0][0],
|
|
|
|
"safe": output_data[0][1],
|
|
|
|
}
|
|
|
|
|
|
|
|
return result
|