search

GantMan / nsfw_model

Keras model of NSFW detector
Other
1.8k stars 279 forks source link

224*224 is not a good idea #89

Open yingshaoxo opened 3 years ago

yingshaoxo commented 3 years ago

To be honest, the input should be a vector extracted from the last layer of inceptionv3.

Only in this way, the model can get random sized image information right.

at the current version, even a woman that not porn, will be tagged as a porn picture with 0.9 confidence...

girl

yingshaoxo commented 3 years ago

It turns out to be the problem of input.

Here I got two functions, only the second function works well, it is strange:

The first method work but will get very bad accuracy.

    def detect(self, numpyImage):
        numpyImage = keras.preprocessing.image.smart_resize(numpyImage, size=(IMAGE_DIM, IMAGE_DIM), interpolation="nearest")
        numpyImage = numpyImage.astype(np.float)
        numpyImage /= 255
        probs = classify_nd(self.model, np.asarray([numpyImage]))
        return probs

    def classify(self, imagePath):
        image = keras.preprocessing.image.load_img(imagePath, target_size=(IMAGE_DIM, IMAGE_DIM))
        image = keras.preprocessing.image.img_to_array(image)
        image /= 255
        probs = classify_nd(self.model, np.asarray([image]))
        return probs
yingshaoxo commented 3 years ago

So is there a way to use the classify function without writing a .jpg file to disk first?

Help, please.

synodriver commented 3 years ago

So is there a way to use the classify function without writing a .jpg file to disk first?

Help, please.

By replacing it with io.BytesIO

AleksanderDanilian commented 1 year ago

Same with me - trying to use Keras Inception and Mobilenet models provided by the repo author - getting overwelming amount of "porn" classes, even for random pics like dogs, cars etc. Tried both using "from tensorflow.keras.applications.inception_v3 import preprocess_input" and same with Mobilenet. Also tried manually dividing arrays by 255. Still no luck. Help is much appreciated!

Tried resizing by different methods as well (embedded resize functions in tensorflow, PIL). Also tried loading imgs with cv2, PIL, Keras.

szeitlin commented 1 year ago

I'm also running into this problem, works fine from file on disk, does not work starting from bytearrays after resizing. My hypothesis is that it's something in what keras does when it loads an image, which is different: https://github.com/GantMan/nsfw_model/blob/master/nsfw_detector/predict.py#L42