janpalasek / ulozto-captcha-breaker Goto Github PK

import tensorflow as tf


def all_correct_acc(y_true: tf.Tensor, y_pred: tf.Tensor):
    """
    Computes accuracy between y_true and y_pred in the following manner:

    - If i-th sample has all values on y_pred same as on y_true, then 1.
    - Otherwise 0.

    It is hence more restricting then a typical accuracy.

    Args:
        y_true (tf.Tensor): 2D tensor of shape (N, L), where N is the number of samples and L is length of the vector (number of characters).
        y_pred: 2D tensor of shape (N, L), where N is the number of samples and L is length of the vector (number of characters)

    Returns:
        Accuracy: number between [0, 1] denoting how many codes were predicted correctly.
    """
    if y_true.shape[0] is None and y_true.shape[1] is None and y_true.shape[2] is None:
        return tf.convert_to_tensor(0)

    # cast to int64 so we can compare it
    y_true = tf.cast(y_true, tf.dtypes.int64)

    if len(y_pred.shape) <= 2:
        y_pred = tf.expand_dims(y_pred, axis=1)
    if len(y_true.shape) <= 1:
        y_true = tf.expand_dims(y_true, axis=1)
    y_pred = tf.argmax(y_pred, axis=2)
    correct = y_true == y_pred
    # tf.print(f"Pred shape: {y_true.shape}", output_stream=sys.stdout)

    all_correct = tf.reduce_all(correct, axis=1)
    all_correct = tf.cast(all_correct, tf.dtypes.float32)

    return tf.reduce_mean(all_correct)

import numpy as np
from PIL import Image


class ConvertToGrayscalePreprocessor:
    """
    Converts image to grayscale.
    """
    def __call__(self, img: np.ndarray):
        r, g, b = img[:, :, 0], img[:, :, 1], img[:, :, 2]
        output = 0.299 * r + 0.587 * g + 0.114 * b
        return output


class ImageCutPreprocessor:
    def __init__(self, pieces_count: int):
        self._pieces_count = pieces_count

    def __call__(self, image: np.ndarray):
        images = np.split(image, self._pieces_count, axis=1)

        return np.array(images)


class NormalizeImagePreprocessor:
    """
    Converts image from byte format (values are integers in {0, ..., 255} to normalized float format (values are
    floats in the interval [0, 1].
    """
    def __init__(self):
        pass

    def __call__(self, image):
        image = image.astype(np.float32) / 255
        image = np.expand_dims(image, axis=len(image.shape))
        return image


class ResizePreprocessor:
    """
    Resizes image to target width and height.
    """
    def __init__(self, target_height, target_width):
        self._target_height = target_height
        self._target_width = target_width

    def __call__(self, img: np.ndarray):
        return img.resize(img, (self._target_width, self._target_height))

git clone [email protected]:JanPalasek/ulozto-captcha-breaker.git
cd ulozto-captcha-breaker
python -m venv "venv"
source venv/bin/activate
python -m pip install --upgrade pip
python -m pip install --upgrade wheel setuptools pip-tools
python -m piptools sync
python -m pip install -e .

python bin/simple_captcha_generate.py --height=70 --width=175 --available_chars="0123456789" --captcha_length=4 --dataset_size=10000
python bin/captcha_annotate.py --val_split=0.1 --test_split=0.1

python bin/train.py --available_chars="0123456789" --captcha_length=4

python bin/test.py --available_chars="0123456789" --captcha_length=4 --weights_file=/home/manas/Projects/2.Other/ulozto-captcha-breaker/out/logs/train.py-2023-09-10_163320-ac=0123456789,bs=32,cl=4,e=1500,fl=0,l=1e-05,od=out,pm=None,rl=False,smp=None,s=42,tih=None,tiw=None,wf=None/cp-07.h5

32/32 [==============================] - 2s 40ms/step
Test acc: 0.92

python bin/create_tflite.py --pretrained_model=out/model

2023-09-11 18:07:01.736468: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:362] Ignored output_format.
2023-09-11 18:07:01.736497: W tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.cc:365] Ignored drop_control_dependency.
2023-09-11 18:07:01.737129: I tensorflow/cc/saved_model/reader.cc:45] Reading SavedModel from: out/model
2023-09-11 18:07:01.753180: I tensorflow/cc/saved_model/reader.cc:89] Reading meta graph with tags { serve }
2023-09-11 18:07:01.753227: I tensorflow/cc/saved_model/reader.cc:130] Reading SavedModel debug info (if present) from: out/model
2023-09-11 18:07:01.785731: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:354] MLIR V1 optimization pass is not enabled
2023-09-11 18:07:01.796363: I tensorflow/cc/saved_model/loader.cc:229] Restoring SavedModel bundle.
2023-09-11 18:07:01.967260: I tensorflow/cc/saved_model/loader.cc:213] Running initialization op on SavedModel bundle at path: out/model
2023-09-11 18:07:02.019981: I tensorflow/cc/saved_model/loader.cc:305] SavedModel load for tags { serve }; Status: success: OK. Took 282855 microseconds.
2023-09-11 18:07:02.140917: I tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc:268] disabling MLIR crash reproducer, set env var `MLIR_CRASH_REPRODUCER_DIRECTORY` to enable.

python bin/predict.py --model_path=out/model.tflite --available_chars="0123456789" --image_path=test1.png
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Decoded label is the following:
1117

python bin/predict.py --model_path=out/model.tflite --available_chars="0123456789" --image_path=test2.png
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Decoded label is the following:
1197

python bin/predict.py --model_path=out/model.tflite --available_chars="0123456789" --image_path=test3.png
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Decoded label is the following:
1147

python bin/predict.py --model_path=out/model.tflite --available_chars="0123456789" --image_path=test4.png
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Decoded label is the following:
1147

python bin/predict.py --model_path=out/model.tflite --available_chars="0123456789" --image_path=test5.png
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Decoded label is the following:
1147