Merge branch 'main' into pages

.gitignore

@@ -14,3 +14,4 @@ src/HowBadIsMyBatch.nbconvert.html
 src/__pycache__/
 src/intensivstationen/__pycache__/
 google-chrome-stable_current_amd64*
+src/captcha/__pycache__

environment.yml

@@ -3,7 +3,7 @@ channels:
   - defaults
   # - conda-forge
 dependencies:
-  - python=3
+  - python=3.9
   - ipykernel
   - numpy
   - pandas
@@ -12,7 +12,7 @@ dependencies:
   - bs4
   - lxml
   - jupyter
-  - tensorflow
+  - tensorflow=2.11
   - nb_conda_kernels
   - pillow
   - openpyxl

src/CaptchaReader.py

@@ -1,74 +0,0 @@
-import numpy as np
-import tensorflow as tf
-from tensorflow import keras
-from tensorflow.keras import layers
-from PIL import Image
-import numpy as np
-import io
-
-# copied from value of characters variable in captcha_ocr.ipynb or captcha_ocr_trainAndSaveModel.ipynb
-characters = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', 'A', 'B', 'C', 'D', 'E', 'F', 'a', 'b', 'c', 'd', 'e', 'f']
-
-img_width = 241
-img_height = 62
-
-downsample_factor = 4
-
-# copied from value of max_length variable in captcha_ocr.ipynb or captcha_ocr_trainAndSaveModel.ipynb
-max_length = 6
-
-char_to_num = layers.StringLookup(
-    vocabulary=list(characters),
-    mask_token=None)
-
-num_to_char = layers.StringLookup(
-    vocabulary=char_to_num.get_vocabulary(),
-    mask_token=None, invert=True)
-
-def encode_single_sample(img_path):
-    # 1. Read image
-    img = tf.io.read_file(img_path)
-    # 2. Decode and convert to grayscale
-    img = tf.io.decode_png(img, channels=1)
-    # 3. Convert to float32 in [0, 1] range
-    img = tf.image.convert_image_dtype(img, tf.float32)
-    # 4. Resize to the desired size
-    img = tf.image.resize(img, [img_height, img_width])
-    # 5. Transpose the image because we want the time
-    # dimension to correspond to the width of the image.
-    img = tf.transpose(img, perm=[1, 0, 2])
-    # 7. Return a dict as our model is expecting two inputs
-    return asSingleSampleBatch(img)
-
-def asSingleSampleBatch(img):
-    array = keras.utils.img_to_array(img)
-    array = np.expand_dims(array, axis=0)
-    return array
-
-def decode_batch_predictions(pred):
-    input_len = np.ones(pred.shape[0]) * pred.shape[1]
-    # Use greedy search. For complex tasks, you can use beam search
-    results = keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0][:, :max_length]
-    # Iterate over the results and get back the text
-    output_text = []
-    for res in results:
-        res = tf.strings.reduce_join(num_to_char(res)).numpy().decode("utf-8")
-        output_text.append(res)
-    return output_text
-
-def _getModel():
-    print("loading model...")
-    model = load_model()
-    model.summary()
-    return model
-
-def load_model():
-    model = keras.models.load_model('model')
-    return keras.models.Model(
-        model.get_layer(name="image").input,
-        model.get_layer(name="dense2").output)
-
-def getTextInCaptchaImage(captchaImageFile):
-    batchImages = encode_single_sample(captchaImageFile)
-    preds = _getModel().predict(batchImages)
-    return decode_batch_predictions(preds)[0]

src/HowBadIsMyBatch.ipynb

@@ -329,7 +329,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.8"
+   "version": "3.9.15"
   },
   "vscode": {
    "interpreter": {

src/VAERSFileDownloader.py

@@ -2,8 +2,9 @@ import os
 import time
 from WebDriver import getWebDriver, isCaptchaSolved, saveCaptchaImageAs
 from selenium.webdriver.common.by import By
-from CaptchaReader import getTextInCaptchaImage
+from captcha.CaptchaReader import CaptchaReader
 from zipUtils import unzipAndRemove
+from captcha.CaptchaShape import CaptchaShape
 
 
 #def getTextInCaptchaImage(captchaImageFile):
@@ -15,11 +16,16 @@ from zipUtils import unzipAndRemove
     
 def solveCaptchaAndStartFileDownload(driver, captchaImageFile):
     saveCaptchaImageAs(driver, captchaImageFile)
-    textInCaptchaImage = getTextInCaptchaImage(captchaImageFile)
+    textInCaptchaImage = _createCaptchaReader().getTextInCaptchaImage(captchaImageFile)
     print('textInCaptchaImage:', textInCaptchaImage)
     driver.find_element(By.ID, "verificationCode").send_keys(textInCaptchaImage)
     driver.find_element(By.CSS_SELECTOR, '[name="downloadbut"]').click()
 
+def _createCaptchaReader():
+    working_directory = os.path.dirname(__file__)
+    return CaptchaReader(modelFilepath = f'{working_directory}/captcha/MobileNetV3Small',
+                         captchaShape = CaptchaShape())
+
 def downloadFile(absoluteFile, driver, maxTries):
     def _downloadFile():
         driver.get('https://vaers.hhs.gov/eSubDownload/index.jsp?fn=' + os.path.basename(absoluteFile))
@@ -42,7 +48,7 @@ def _waitUntilDownloadHasFinished(file):
         time.sleep(2)
 
 def downloadVAERSFile(file, downloadDir):
-    driver = getWebDriver(downloadDir, isHeadless = True)
+    driver = getWebDriver(downloadDir, isHeadless = False)
     downloadedFile = downloadFile(
         absoluteFile = downloadDir + "/" + file,
         driver = driver,

src/captcha.ipynb

@@ -0,0 +1,561 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "UNKC5YSEIS_d"
+      },
+      "source": [
+        "# Captchas\n",
+        "\n",
+        "**see:** https://keras.io/examples/vision/captcha_ocr/<br>\n",
+        "**original:** https://colab.research.google.com/drive/1Olw2KMHfPlnGaYuzffl2zb6D1etlBGZf?usp=sharing<br>\n",
+        "**View Github version in Colab:** <a href=\"https://colab.research.google.com/github/KnollFrank/2captcha-worker-assistant-server/blob/master/captcha_ocr_trainAndSaveModel_colab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a><br>\n",
+        "**paper:** Simple and Easy: Transfer Learning-Based Attacks to Text CAPTCHA<br>"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "wRUsVuIiIS_s"
+      },
+      "source": [
+        "## Setup"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "zZSwQragIS_v"
+      },
+      "outputs": [],
+      "source": [
+        "import os\n",
+        "import numpy as np\n",
+        "from pathlib import Path\n",
+        "import tensorflow as tf\n",
+        "from captcha.GoogleDriveManager import GoogleDriveManager\n",
+        "from captcha.CaptchaGenerator import CaptchaGenerator\n",
+        "from captcha.CharNumConverter import CharNumConverter\n",
+        "from captcha.DataSplitter import DataSplitter\n",
+        "from captcha.DatasetFactory import DatasetFactory\n",
+        "from captcha.ModelFactory import ModelFactory\n",
+        "from captcha.PredictionsDecoder import PredictionsDecoder\n",
+        "from captcha.ModelDAO import ModelDAO\n",
+        "from captcha.CaptchaShape import CaptchaShape"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "0DZfMrbe3MGN"
+      },
+      "outputs": [],
+      "source": [
+        "def getImagesAndLabels(dataDir):\n",
+        "    fileSuffix = \".jpeg\"\n",
+        "    images = sorted(list(map(str, list(dataDir.glob(\"*\" + fileSuffix)))))\n",
+        "    labels = [image.split(os.path.sep)[-1].split(fileSuffix)[0] for image in images]\n",
+        "    return images, labels\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "kdL9_t03Mf3t"
+      },
+      "outputs": [],
+      "source": [
+        "def getTrainValidationTestDatasets(dataDir, datasetFactory):\n",
+        "    images, labels = getImagesAndLabels(dataDir)\n",
+        "    print(\"Number of images found:\", len(images))\n",
+        "    print(\"Characters:\", CaptchaGenerator.characters)\n",
+        "\n",
+        "    dataSplitter = DataSplitter(images, labels)\n",
+        "    \n",
+        "    return (\n",
+        "        datasetFactory.createDataset(*dataSplitter.getTrain()),\n",
+        "        datasetFactory.createDataset(*dataSplitter.getValid()),\n",
+        "        datasetFactory.createDataset(*dataSplitter.getTest())\n",
+        "        )"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "FqVSEuZp3MGT"
+      },
+      "outputs": [],
+      "source": [
+        "import matplotlib.pyplot as plt\n",
+        "import math\n",
+        "\n",
+        "def displayImagesInGrid(numGridCols, images, titles, titleColors):\n",
+        "    assert len(images) == len(titles) == len(titleColors)\n",
+        "    images = [image.numpy().astype(np.uint8) for image in images]\n",
+        "    numGridRows = math.ceil(len(images) / numGridCols)\n",
+        "    _, axs = plt.subplots(numGridRows, numGridCols, figsize=(15, 5))\n",
+        "    for row in range(numGridRows):\n",
+        "        for col in range(numGridCols):\n",
+        "            ax = axs[row, col]\n",
+        "            ax.axis(\"off\")\n",
+        "            i = row * numGridCols + col\n",
+        "            if(i < len(images)):\n",
+        "                ax.imshow(images[i])\n",
+        "                ax.set_title(titles[i], color=titleColors[i])\n",
+        "    plt.show()\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "apkeCHhP3MGU"
+      },
+      "outputs": [],
+      "source": [
+        "def display16Predictions(model, dataset, predictionsDecoder):\n",
+        "    for batch in dataset.take(1):\n",
+        "        numPredictions2Display = 16\n",
+        "        batch_images = batch[\"image\"][:numPredictions2Display]\n",
+        "        batch_labels = batch[\"label\"][:numPredictions2Display]\n",
+        "\n",
+        "        preds = model.predict(batch_images)\n",
+        "        pred_texts = predictionsDecoder.decode_batch_predictions(preds)\n",
+        "        orig_texts = predictionsDecoder.asStrings(batch_labels)\n",
+        "\n",
+        "        displayImagesInGrid(\n",
+        "            4,\n",
+        "            batch_images,\n",
+        "            [f\"Prediction/Truth: {pred_text}/{orig_text}\" for (pred_text, orig_text) in zip(pred_texts, orig_texts)],\n",
+        "            ['green' if pred_text == orig_text else 'red' for (pred_text, orig_text) in zip(pred_texts, orig_texts)])"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "def printLayers(model):\n",
+        "    for i, layer in enumerate(model.layers):\n",
+        "        print(i, layer.name)\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "S3X_SslH3MGY"
+      },
+      "outputs": [],
+      "source": [
+        "# FK-TODO: entferne die getAccuracy()-Methode. Implementiere stattdessen https://stackoverflow.com/questions/37657260/how-to-implement-custom-metric-in-keras oder https://keras.io/api/metrics/#custom-metrics\n",
+        "def getAccuracy(dataset, prediction_model, ctc_decode):\n",
+        "    accuracy = tf.keras.metrics.Accuracy()\n",
+        "\n",
+        "    for batch in dataset:\n",
+        "        accuracy.update_state(batch[\"label\"], ctc_decode(prediction_model.predict(batch[\"image\"], verbose=0)))\n",
+        "\n",
+        "    return accuracy.result().numpy()"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "94755hrNMf3w"
+      },
+      "source": [
+        "## Preparation"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "NZrKXF6P3MGY"
+      },
+      "outputs": [],
+      "source": [
+        "inColab = 'google.colab' in str(get_ipython())"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "7EsmTaF03MGZ"
+      },
+      "outputs": [],
+      "source": [
+        "if inColab:\n",
+        "    GoogleDriveManager.mount()"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "S_4hl4S4BmZK"
+      },
+      "outputs": [],
+      "source": [
+        "if inColab:\n",
+        "  !cp {GoogleDriveManager._baseFolder}/captchas.zip .\n",
+        "  !unzip captchas.zip"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "WmUghcQaMf3y"
+      },
+      "outputs": [],
+      "source": [
+        "modelDAO = ModelDAO(inColab)\n",
+        "charNumConverter = CharNumConverter(CaptchaGenerator.characters)\n",
+        "predictionsDecoder = PredictionsDecoder(CaptchaGenerator.captchaLength, charNumConverter.num_to_char)\n",
+        "captchaShape = CaptchaShape()\n",
+        "datasetFactory = DatasetFactory(captchaShape, charNumConverter.char_to_num, batch_size = 64)"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "lsLuSi7h3MGZ"
+      },
+      "source": [
+        "## Create And Train Base Model"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "oRcemcbG3MGa"
+      },
+      "outputs": [],
+      "source": [
+        "if inColab:\n",
+        "    !sudo apt install ttf-mscorefonts-installer\n",
+        "    !sudo fc-cache -f\n",
+        "    !fc-match Arial"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "P7myCt7e2h6A"
+      },
+      "outputs": [],
+      "source": [
+        "# \"We generate 200,000 images for base model pre-training\"\n",
+        "captchaGenerator = CaptchaGenerator(\n",
+        "    numCaptchas = 50, # 50, # 200000,\n",
+        "    dataDir = Path(\"captchas/generated/VAERS/\"))"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "j9apYsyI3MGb"
+      },
+      "outputs": [],
+      "source": [
+        "captchaGenerator.createAndSaveCaptchas()"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "AgN4skCkMf31"
+      },
+      "outputs": [],
+      "source": [
+        "train_dataset, validation_dataset, test_dataset = getTrainValidationTestDatasets(captchaGenerator.dataDir, datasetFactory)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "RcgWHXVSNsa7"
+      },
+      "outputs": [],
+      "source": [
+        "for batch in train_dataset.take(1):\n",
+        "    numImages2Display = 16\n",
+        "    images = batch[\"image\"][:numImages2Display]\n",
+        "    labels = batch[\"label\"][:numImages2Display]\n",
+        "    displayImagesInGrid(4, images, predictionsDecoder.asStrings(labels), ['black'] * len(labels))"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "zDoFYKM2hdEW"
+      },
+      "outputs": [],
+      "source": [
+        "modelFactory = ModelFactory(captchaShape, charNumConverter.char_to_num)\n",
+        "model = modelFactory.createMobileNetV3Small()\n",
+        "model.summary()"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "ltXYrpjIITAb"
+      },
+      "outputs": [],
+      "source": [
+        "# \"the success rates became stable after the base-model training epochs exceeded 20\"\n",
+        "history = model.fit(\n",
+        "    train_dataset,\n",
+        "    validation_data=validation_dataset,\n",
+        "    epochs=20)\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "fPG-Yl1SJfF7"
+      },
+      "outputs": [],
+      "source": [
+        "modelDAO.saveModel(model)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "NnNHMtIGITAe"
+      },
+      "outputs": [],
+      "source": [
+        "prediction_model = ModelFactory.createPredictionModel(model)\n",
+        "prediction_model.summary()\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "YW651ztD8sKI"
+      },
+      "outputs": [],
+      "source": [
+        "display16Predictions(prediction_model, test_dataset, predictionsDecoder)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "V5gqMBIwBmZU"
+      },
+      "outputs": [],
+      "source": [
+        "getAccuracy(test_dataset, prediction_model, predictionsDecoder.ctc_decode)"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "UYxiYTH9BmZU"
+      },
+      "source": [
+        "## Transfer learning"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "WV8IS4KrBmZU"
+      },
+      "outputs": [],
+      "source": [
+        "# \"we collected 1,500 real CAPTCHAs from the websites. Note that only 500 of them are used for fine-tuning, and another 1,000 are applied to calculate the test accuracy\"\n",
+        "# FK-TODO: lade das pre-trainierte model und trainiere es mit 500 real-world-Daten aus dem Ordner captchas/VAERS/, die restlichen 540 (es sollten nach obigem Zitat aber 1,000 sein) sind dann die Test-Daten.\n",
+        "# see https://keras.io/guides/transfer_learning/\n",
+        "# see https://www.tensorflow.org/tutorials/images/transfer_learning\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "modelName, numTrainableLayers = 'MobileNetV3Small', 104\n",
+        "# modelName, numTrainableLayers = 'ResNet101', 348"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "D7ogEQmB3MGj"
+      },
+      "outputs": [],
+      "source": [
+        "# FK-TODO: DRY with VAERSFileDownloader\n",
+        "modelFilepath = f'{os.getcwd()}/captcha/{modelName}'\n",
+        "model = modelDAO.loadModel(modelFilepath)\n",
+        "model.summary(show_trainable=True)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "gbPigogKNFrD"
+      },
+      "outputs": [],
+      "source": [
+        "# printLayers(model)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "59quw8o3Mf34"
+      },
+      "outputs": [],
+      "source": [
+        "model.trainable = True\n",
+        "for layer in model.layers[:numTrainableLayers]:\n",
+        "    layer.trainable = False"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "acGczax3Mf34"
+      },
+      "outputs": [],
+      "source": [
+        "model.summary(show_trainable=True)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "q7_MjUO0BmZV"
+      },
+      "outputs": [],
+      "source": [
+        "train_dataset, validation_dataset, test_dataset = getTrainValidationTestDatasets(Path(\"captcha/captchas/VAERS/\"), datasetFactory)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "dZsCpibkBmZX"
+      },
+      "outputs": [],
+      "source": [
+        "# \"The model is optimized by a stochastic gradient descent (SGD) strategy with an initial learning rate of 0.004, weight decay of 0.00004 and momentum of 0.9.\"\n",
+        "from tensorflow.keras.optimizers import SGD\n",
+        "# model.compile(optimizer=SGD(learning_rate=0.0001, momentum=0.9))\n",
+        "model.compile(optimizer='adam')\n",
+        "\n",
+        "# \"Therefore, in our experiments, we chose 1 epoch for the fine-tuning stage.\"\n",
+        "history = model.fit(\n",
+        "    train_dataset,\n",
+        "    validation_data=validation_dataset,\n",
+        "    epochs=20)\n"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "TRbJigbH3MGl"
+      },
+      "outputs": [],
+      "source": [
+        "prediction_model = ModelFactory.createPredictionModel(model)\n",
+        "prediction_model.summary()"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "rPszfhJ4BmZX"
+      },
+      "outputs": [],
+      "source": [
+        "getAccuracy(test_dataset, prediction_model, predictionsDecoder.ctc_decode)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "collapsed": true,
+        "id": "hfmRY1qC7aVV"
+      },
+      "outputs": [],
+      "source": [
+        "display16Predictions(prediction_model, test_dataset, predictionsDecoder)"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "modelDAO.saveModel(model)"
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "collapsed_sections": [],
+      "name": "captcha.ipynb",
+      "private_outputs": true,
+      "provenance": []
+    },
+    "gpuClass": "standard",
+    "kernelspec": {
+      "display_name": "howbadismybatch-venv-kernel",
+      "language": "python",
+      "name": "howbadismybatch-venv-kernel"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.9.15"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

src/captcha/CTCLayer.py

@@ -0,0 +1,27 @@
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import layers
+
+
+# see https://keras.io/guides/making_new_layers_and_models_via_subclassing/
+class CTCLayer(layers.Layer):
+    
+    def __init__(self, name=None):
+        super().__init__(name=name)
+        self.loss_fn = keras.backend.ctc_batch_cost
+
+    def call(self, y_true, y_pred):
+        # Compute the training-time loss value and add it
+        # to the layer using `self.add_loss()`.
+        batch_len = tf.cast(tf.shape(y_true)[0], dtype="int64")
+        input_length = tf.cast(tf.shape(y_pred)[1], dtype="int64")
+        label_length = tf.cast(tf.shape(y_true)[1], dtype="int64")
+
+        input_length = input_length * tf.ones(shape=(batch_len, 1), dtype="int64")
+        label_length = label_length * tf.ones(shape=(batch_len, 1), dtype="int64")
+
+        loss = self.loss_fn(y_true, y_pred, input_length, label_length)
+        self.add_loss(loss)
+
+        # At test time, just return the computed predictions
+        return y_pred

src/captcha/CaptchaGenerator.py

@@ -0,0 +1,43 @@
+from PIL import Image, ImageDraw, ImageFont
+import random
+import string
+import shutil
+
+
+class CaptchaGenerator:
+
+    characters = sorted(set(list(string.ascii_letters + string.digits)))
+    captchaLength = 6
+
+    def __init__(self, numCaptchas, dataDir):
+        self.numCaptchas = numCaptchas
+        self.dataDir = dataDir
+
+    def createAndSaveCaptchas(self):
+        self._prepareDataDir()
+        for _ in range(self.numCaptchas):
+            self._createAndSaveCaptcha()
+
+    def _prepareDataDir(self):
+        shutil.rmtree(self.dataDir, ignore_errors = True)
+        self.dataDir.mkdir(parents=True, exist_ok=True)
+
+    def _createAndSaveCaptcha(self):
+        captchaString = self._createCaptchaString()
+        captcha = self._createCaptcha(captchaString)
+        captcha.save(f"{str(self.dataDir)}/{captchaString}.jpeg")
+
+    def _createCaptchaString(self):
+        return ''.join(random.choice(CaptchaGenerator.characters) for _ in range(CaptchaGenerator.captchaLength))
+
+    def _createCaptcha(self, word):
+        image = Image.new("RGB", (360, 96), "#373737")
+        draw = ImageDraw.Draw(image)
+        font = ImageFont.truetype("ariali.ttf", size=40)
+        draw.text((30, 10), word[0], font=font)
+        draw.text((80, 30), word[1], font=font)
+        draw.text((135, 10), word[2], font=font)
+        draw.text((190, 30), word[3], font=font)
+        draw.text((250, 10), word[4], font=font)
+        draw.text((295, 30), word[5], font=font)
+        return image

src/captcha/CaptchaReader.py

@@ -0,0 +1,31 @@
+from captcha.ModelDAO import ModelDAO
+from captcha.ModelFactory import ModelFactory
+from captcha.PredictionsDecoder import PredictionsDecoder
+from captcha.CaptchaGenerator import CaptchaGenerator
+from captcha.CharNumConverter import CharNumConverter
+from captcha.DatasetFactory import DatasetFactory
+import numpy as np
+from tensorflow import keras
+
+
+class CaptchaReader:
+
+    def __init__(self, modelFilepath, captchaShape):
+        self.modelFilepath = modelFilepath
+        self.captchaShape = captchaShape
+
+    def getTextInCaptchaImage(self, captchaImageFile):
+        return self._getTextsInCaptchaImage(self._getCaptchaImage(captchaImageFile))[0]
+
+    def _getCaptchaImage(self, captchaImageFile):
+        return self._asSingleSampleBatch(DatasetFactory.encodeImage(captchaImageFile, self.captchaShape))
+
+    def _asSingleSampleBatch(self, img):
+        return np.expand_dims(keras.utils.img_to_array(img), axis=0)
+
+    def _getTextsInCaptchaImage(self, captchaImage):
+        preds = self._createPredictionModel().predict(captchaImage)
+        return PredictionsDecoder(CaptchaGenerator.captchaLength, CharNumConverter(CaptchaGenerator.characters).num_to_char).decode_batch_predictions(preds)
+
+    def _createPredictionModel(self):
+        return ModelFactory.createPredictionModel(ModelDAO(inColab=False).loadModel(self.modelFilepath))

src/captcha/CaptchaReaderTest.py

@@ -0,0 +1,22 @@
+import unittest
+from captcha.CaptchaReader import CaptchaReader
+from captcha.CaptchaShape import CaptchaShape
+import os
+
+class CaptchaReaderTest(unittest.TestCase):
+
+    def setUp(self):
+        self.working_directory = os.path.dirname(__file__)
+
+    def test_getTextInCaptchaImage(self):
+        # Given
+        textInCaptchaImage = '1Ad47a'
+        captchaReader = CaptchaReader(
+            modelFilepath = f'{self.working_directory}/MobileNetV3Small',
+            captchaShape = CaptchaShape())
+
+        # When
+        textInCaptchaImageActual = captchaReader.getTextInCaptchaImage(f'{self.working_directory}/captchas/VAERS/{textInCaptchaImage}.jpeg')
+
+        # Then
+        self.assertEqual(textInCaptchaImageActual, textInCaptchaImage)

src/captcha/CaptchaShape.py

@@ -0,0 +1,5 @@
+class CaptchaShape:
+
+    def __init__(self):
+        self.width = 241
+        self.height = 62

src/captcha/CharNumConverter.py

@@ -0,0 +1,10 @@
+from tensorflow.keras import layers
+
+class CharNumConverter:
+
+    def __init__(self, characters):
+        self.char_to_num = layers.StringLookup(vocabulary=list(characters), mask_token=None)
+        self.num_to_char = layers.StringLookup(
+            vocabulary=self.char_to_num.get_vocabulary(),
+            mask_token=None,
+            invert=True)

src/captcha/DataSplitter.py

@@ -0,0 +1,28 @@
+import numpy as np
+
+
+class DataSplitter:
+
+    def __init__(self, x, y):
+        (self.x_train, self.y_train), (x_valid_test, y_valid_test) = DataSplitter._splitData(np.array(x), np.array(y), train_size=0.7)
+        (self.x_valid, self.y_valid), (self.x_test, self.y_test) = DataSplitter._splitData(x_valid_test, y_valid_test, train_size=0.5)
+
+    def getTrain(self):
+        return (self.x_train, self.y_train)
+
+    def getValid(self):
+        return (self.x_valid, self.y_valid)
+
+    def getTest(self):
+        return (self.x_test, self.y_test)
+
+    @staticmethod
+    def _splitData(x, y, train_size=0.9, shuffle=True):
+        size = len(x)
+        indices = np.arange(size)
+        if shuffle:
+            np.random.shuffle(indices)
+        train_samples = int(size * train_size)
+        x_train, y_train = x[indices[:train_samples]], y[indices[:train_samples]]
+        x_test, y_test = x[indices[train_samples:]], y[indices[train_samples:]]
+        return (x_train, y_train), (x_test, y_test)

src/captcha/DatasetFactory.py

@@ -0,0 +1,27 @@
+import tensorflow as tf
+
+
+class DatasetFactory:
+    
+    def __init__(self, captchaShape, char_to_num, batch_size):
+        self.captchaShape = captchaShape
+        self.char_to_num = char_to_num
+        self.batch_size = batch_size
+
+    def createDataset(self, x, y):
+        dataset = tf.data.Dataset.from_tensor_slices((x, y))
+        dataset = dataset.map(self._encodeImageAndLabel, num_parallel_calls=tf.data.AUTOTUNE)
+        dataset = dataset.batch(self.batch_size).prefetch(buffer_size=tf.data.AUTOTUNE)
+        return dataset
+
+    def _encodeImageAndLabel(self, imageFilename, label):
+        return {
+            "image": DatasetFactory.encodeImage(imageFilename, self.captchaShape),
+            "label": self.char_to_num(tf.strings.unicode_split(label, input_encoding="UTF-8"))}
+    
+    @staticmethod
+    def encodeImage(imageFilename, captchaShape):
+        img = tf.io.read_file(imageFilename)
+        img = tf.io.decode_jpeg(img, channels=3)
+        img = tf.image.resize(img, [captchaShape.height, captchaShape.width])
+        return img

src/captcha/GoogleDriveManager.py

@@ -0,0 +1,27 @@
+from pathlib import Path
+
+
+class GoogleDriveManager:
+  
+  _googleDriveFolder = Path('/content/gdrive')
+  _baseFolder = _googleDriveFolder / 'MyDrive/CAPTCHA/models/'
+
+  @staticmethod
+  def mount():
+    from google.colab import drive
+    drive.mount(str(GoogleDriveManager._googleDriveFolder))
+
+  @staticmethod
+  def uploadFolderToGoogleDrive(folder):
+    pass
+    # FK-FIXME:
+    # !zip -r {folder}.zip {folder}/
+    # !cp {folder}.zip {GoogleDriveManager._baseFolder}
+
+  @staticmethod
+  def downloadFolderFromGoogleDrive(folder):
+    pass
+    # FK-FIXME:
+    # !cp {GoogleDriveManager._baseFolder}/{folder}.zip .
+    # !rm -rf {folder}
+    # !unzip {folder}.zip

src/captcha/ModelDAO.py

@@ -0,0 +1,20 @@
+from tensorflow import keras
+from captcha.GoogleDriveManager import GoogleDriveManager
+import shutil
+
+
+class ModelDAO:
+
+    def __init__(self, inColab):
+        self.inColab = inColab
+
+    def saveModel(self, model):
+        shutil.rmtree(model.name, ignore_errors = True)
+        model.save(model.name)
+        if self.inColab:
+            GoogleDriveManager.uploadFolderToGoogleDrive(model.name)
+
+    def loadModel(self, modelFilepath):
+        if self.inColab:
+            GoogleDriveManager.downloadFolderFromGoogleDrive(modelFilepath)
+        return keras.models.load_model(modelFilepath)

src/captcha/ModelFactory.py

@@ -0,0 +1,102 @@
+from captcha.CTCLayer import CTCLayer
+import tensorflow as tf
+from tensorflow import keras
+from tensorflow.keras import layers
+
+
+class ModelFactory:
+    
+    predictionModelInputLayerName = "image"
+    predictionModelOutputLayerName = "dense2"
+
+    def __init__(self, captchaShape, char_to_num):
+        self.captchaShape = captchaShape
+        self.char_to_num = char_to_num
+
+    # see https://www.tensorflow.org/api_docs/python/tf/keras/applications/resnet/ResNet101
+    def createResNet101(self):
+        return self._createModel(
+            baseModelFactory = lambda input_tensor: tf.keras.applications.resnet.ResNet101(
+                input_tensor = input_tensor,
+                weights = 'imagenet',
+                include_top = False),
+            preprocess_input = tf.keras.applications.resnet.preprocess_input,
+            name = 'ResNet101')
+
+    def createMobileNetV2(self):
+        return self._createModel(
+            baseModelFactory = lambda input_tensor: tf.keras.applications.MobileNetV2(
+                input_tensor = input_tensor,
+                weights = 'imagenet',
+                include_top = False),
+            preprocess_input = tf.keras.applications.mobilenet_v2.preprocess_input,
+            name = 'MobileNetV2')
+
+    def createMobileNetV3Small(self):
+        return self._createModel(
+            baseModelFactory = lambda input_tensor: tf.keras.applications.MobileNetV3Small(
+                input_tensor = input_tensor,
+                minimalistic = True,
+                weights = 'imagenet',
+                include_top = False),
+            preprocess_input = tf.keras.applications.mobilenet_v3.preprocess_input,
+            name = 'MobileNetV3Small')
+            
+    @staticmethod
+    def createPredictionModel(model):
+        return keras.models.Model(
+            model.get_layer(name=ModelFactory.predictionModelInputLayerName).input,
+            model.get_layer(name=ModelFactory.predictionModelOutputLayerName).output)
+
+    def _createModel(self, baseModelFactory, preprocess_input, name):
+        # Inputs to the model
+        input_image = layers.Input(
+            shape = (self.captchaShape.height, self.captchaShape.width, 3),
+            name = ModelFactory.predictionModelInputLayerName,
+            dtype = "float32")
+        labels = layers.Input(name="label", shape=(None,), dtype="float32")
+        
+        image = preprocess_input(input_image)
+        # Transpose the image because we want the time dimension to correspond to the width of the image.
+        image = tf.keras.layers.Permute(dims=[2, 1, 3])(image)
+        base_model = baseModelFactory(image)
+        x = layers.Reshape(
+            target_shape=(base_model.output_shape[1], base_model.output_shape[2] * base_model.output_shape[3]),
+            name="reshape")(base_model.output)
+        x = layers.Dense(64, activation="relu", name="dense1")(x)
+        x = layers.Dropout(0.2)(x)
+
+        # RNNs
+        x = layers.Bidirectional(
+            layers.LSTM(
+                128,
+                return_sequences=True,
+                dropout=0.25,
+                unroll=False,
+                name="LSTM1"))(x)
+        x = layers.Bidirectional(
+            layers.LSTM(
+                64,
+                return_sequences=True,
+                dropout=0.25,
+                unroll=False,
+                name="LSTM2"))(x)
+
+        # Output layer
+        x = layers.Dense(
+            len(self.char_to_num.get_vocabulary()) + 1,
+            activation="softmax",
+            name=ModelFactory.predictionModelOutputLayerName)(x)
+
+        # Add CTC layer for calculating CTC loss at each step
+        output = CTCLayer(name="ctc_loss")(labels, x)
+
+        model = keras.models.Model(
+            inputs=[input_image, labels],
+            outputs=output,
+            name=name)
+        # "The model is optimized by a stochastic gradient descent (SGD) strategy with an initial learning rate of 0.004, weight decay of 0.00004 and momentum of 0.9."
+        # from tensorflow.keras.optimizers import SGD
+        # model.compile(optimizer=SGD(learning_rate=0.004, "weight_decay=0.00004," momentum=0.9)
+        model.compile(optimizer=keras.optimizers.Adam())
+        return model

src/captcha/PredictionsDecoder.py

@@ -0,0 +1,24 @@
+import tensorflow as tf
+from tensorflow import keras
+import numpy as np
+
+
+class PredictionsDecoder:
+
+    def __init__(self, captchaLength, num_to_char):
+        self.captchaLength = captchaLength
+        self.num_to_char = num_to_char
+
+    def decode_batch_predictions(self, pred):
+        return self.asStrings(self.ctc_decode(pred))
+
+    def ctc_decode(self, pred):
+        input_len = np.ones(pred.shape[0]) * pred.shape[1]
+        # Use greedy search. For complex tasks, you can use beam search
+        return keras.backend.ctc_decode(pred, input_length=input_len, greedy=True)[0][0][:, :self.captchaLength]
+
+    def asStrings(self, labels):
+        return [self.asString(label) for label in labels]
+
+    def asString(self, label):
+        return tf.strings.reduce_join(self.num_to_char(label)).numpy().decode("utf-8")