From e45076df91be18606d72efdbb753dfa99efd8e60 Mon Sep 17 00:00:00 2001
From: RoundRonin <roundronin8@gmail.com>
Date: Sun, 31 Mar 2024 16:37:00 +0300
Subject: [PATCH] Objecs for import and neural model created

Model still doesn't support custom structure
---
 .gitignore                                    |   4 +-
 .vscode/settings.json                         |   6 +-
 image_recognition/base.py                     |  20 ---
 image_recognition/cli.py                      | 157 +++++-------------
 image_recognition/modules/__init__.py         |   0
 image_recognition/modules/importer.py         |  74 +++++++++
 image_recognition/modules/model.py            |  71 ++++++++
 .../{ => modules}/visualization.py            | 127 +++++++-------
 8 files changed, 260 insertions(+), 199 deletions(-)
 delete mode 100644 image_recognition/base.py
 create mode 100644 image_recognition/modules/__init__.py
 create mode 100644 image_recognition/modules/importer.py
 create mode 100644 image_recognition/modules/model.py
 rename image_recognition/{ => modules}/visualization.py (81%)

diff --git a/.gitignore b/.gitignore
index aebf5f5..e0fc5ab 100644
--- a/.gitignore
+++ b/.gitignore
@@ -134,4 +134,6 @@ dmypy.json
 # Dataset
 
 Data*
-**/Data*
\ No newline at end of file
+**/Data*
+
+Save*
\ No newline at end of file
diff --git a/.vscode/settings.json b/.vscode/settings.json
index dcb1530..1363ec0 100644
--- a/.vscode/settings.json
+++ b/.vscode/settings.json
@@ -1,4 +1,4 @@
-{
-    "python.analysis.typeCheckingMode": "basic",
-    "python.analysis.autoImportCompletions": true
+{
+    "python.analysis.typeCheckingMode": "basic",
+    "python.analysis.autoImportCompletions": true
 }
\ No newline at end of file
diff --git a/image_recognition/base.py b/image_recognition/base.py
deleted file mode 100644
index 0e19fcf..0000000
--- a/image_recognition/base.py
+++ /dev/null
@@ -1,20 +0,0 @@
-"""
-image_recognition base module.
-
-This is the principal module of the image_recognition project.
-here you put your main classes and objects.
-
-Be creative! do whatever you want!
-
-If you want to replace this with a Flask application run:
-
-    $ make init
-
-and then choose `flask` as template.
-"""
-
-# example constant variable
-NAME = "image_recognition"
-
-def hello():
-    print("hello_there")
\ No newline at end of file
diff --git a/image_recognition/cli.py b/image_recognition/cli.py
index 68cd7ef..7a5065c 100644
--- a/image_recognition/cli.py
+++ b/image_recognition/cli.py
@@ -1,36 +1,23 @@
-"""CLI interface for image_recognition project.
 
-Be creative! do whatever you want!
-
-- Install click or typer and create a CLI app
-- Use builtin argparse
-- Start a web application
-- Import things from your .base module
-"""
-import image_recognition.base as base
-from image_recognition.visualization import plotter_evaluator
-
-import numpy as np
-
-import os
-import keras
-from keras import layers
-from tensorflow import data as tf_data
-
-from keras.models import Sequential
-from keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout
-from keras.optimizers import RMSprop
-
-from keras.callbacks import ReduceLROnPlateau
 
+from image_recognition.modules.importer import importer
+from image_recognition.modules.model import model
+from image_recognition.modules.visualization import plotter_evaluator
 
 def main():  # pragma: no cover
 
-    base.hello()
     ## Формирование классов на основе файловой структуры
 
     # Внутри указанной директории должны нахдиться папки, имя которых соотвествует классу.
-    # В папках -- изображения, соответсвующие классу.
+    # В папках -- изображения, соответсвующие классу. Называние изображений не имеет значения.
+
+    # data_directory/
+    # ...class_a/
+    # ......a_image_1.jpg
+    # ......a_image_2.jpg
+    # ...class_b/
+    # ......b_image_1.jpg
+    # ......b_image_2.jpg
 
     # Размер по вертикали, размер по горизонтали. К этим значениям будут приведены все изображения (сжаты/растянуты, не обрезаны)
     height = 140
@@ -41,64 +28,32 @@ def main():  # pragma: no cover
     batch_size = 32
 
     # Вышеописанная директория
-    path_to_data = "Data50"
-
-    train_ds, val_ds = keras.utils.image_dataset_from_directory(
-        path_to_data,
-        validation_split=0.2,
-        subset="both",
-        seed=1337,
-        label_mode="categorical",
-        shuffle=True,
-        image_size=image_size,
-        batch_size=batch_size,
-        color_mode="grayscale",
-        crop_to_aspect_ratio=True,
-    )
+    data_directory = "Data50"
+
+    # Разделение на тренировочные и тестовые данные в долях. Указывается доля тестовых данны (.2 по-умолчанию)
+    validation_split = 0.2
+
+    i = importer(image_size, batch_size, data_directory, validation_split)
     
     # Получение имён классов, числа классов.
-    # TODO: Rework
-    labels =  np.array([])
-    for x, y in val_ds:
-        labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)])
-
-    class_names = set(labels)
-    num_classes = len(class_names)
 
-    print(len(labels))
+    class_names = i.class_names
+    num_classes = i.num_classes
 
     ## Обработка данных
 
     # Вносится рандомизация (ротация, зум, перемещение). Также приводится яркость к понятному нейросети формату (вместо 0-255, 0-1).
 
-    data_augmentation_layers = [
-        layers.RandomRotation(0.08),
-        layers.RandomZoom(
-            height_factor=[-0.2,0.2],
-            width_factor=[-0.2,0.2],
-            fill_mode="constant",
-            fill_value=255.0,
-        ),
-        layers.RandomTranslation(
-            height_factor = [-0.1, 0.1],
-            width_factor = [-0.1, 0.1],
-            fill_mode="constant",
-            fill_value=255.0,
-        ),
-        layers.Rescaling(1.0 / 255)
-    ]
-
-    def data_augmentation(images):
-        for layer in data_augmentation_layers:
-            images = layer(images)
-        return images
+    i.generate_augmentation_layers(0.2, 0.1, 0.08)
 
     ### Применение слоёв обработки данных
 
-    train_ds = train_ds.map(
-        lambda img, label: (data_augmentation(img), label),
-        num_parallel_calls=tf_data.AUTOTUNE,
-    )
+    i.apply_augmentation()
+
+    ## Получение данных для модели
+
+    train_ds = i.train_ds
+    val_ds = i.validation_ds
 
     ## Формирование модели
 
@@ -108,58 +63,38 @@ def data_augmentation(images):
 
     # Последний слой имеет число нейронов, равное количеству классов.
 
-    model = Sequential()
-
-    model.add(keras.Input(shape=(height,width,1)))
-
-    model.add(Conv2D(16, (3,3), 1, activation='relu'))
-    model.add(MaxPooling2D())
-    # model.add(Dropout(0.25))
-
-    model.add(Conv2D(32, (3,3), 1, activation='relu'))
-    model.add(MaxPooling2D())
-    # model.add(Dropout(0.25))
-
-    model.add(Conv2D(16, (3,3), 1, activation='relu'))
-    model.add(MaxPooling2D())
-    # model.add(Dropout(0.25))
-
-    model.add(Flatten())
-    model.add(Dense(256, activation='relu'))
-    # model.add(Dropout(0.25))
-
-    model.add(Dense(num_classes, activation = "softmax"))
-
-    # Компиляция модели
-    optimizer = RMSprop(learning_rate=0.001, rho=0.9, epsilon=1e-08)
-    model.compile(optimizer = optimizer , loss = "categorical_crossentropy", metrics=["accuracy"])
+    # Первое число -- количество фильтров, второе -- окно в пискислях (3 на 3, напрмиер), которым алгоритм проходит по изображению.
+    # Каждый новый tuple -- новый слой Conv2D. Можно эксперементировать с числами.
+    conv_descriptor = (
+        (16, (3,3)),
+        (32, (3,3)),
+        (16, (3,3))
+    )
+    # Количсетво простых слоёв
+    dense_layer_number = 1
 
-    model.summary()
+    m = model(image_size, num_classes, dense_layer_number, conv_descriptor)
+    m.compile()
 
     # Обучение нейросети
 
     # Число проходов по набору данных. Не всегда улучшает результат. Надо смотреть на графики. (50 по умолчанию, при малом наборе данных)
     epochs = 50
 
-    learning_rate_reduction = ReduceLROnPlateau(monitor='accuracy',
-                                            patience=3, 
-                                            verbose=1, 
-                                            factor=0.5, 
-                                            min_lr=0.00001)
-    callbacks = [
-        # keras.callbacks.ModelCheckpoint("save_at_{epoch}.keras"),
-        learning_rate_reduction
-    ]
+    m.init_learning_rate_reduction()
+    # m.init_save_at_epoch()
+
+    m.train(train_ds, epochs, val_ds)
 
-    history = model.fit(train_ds, epochs = epochs, validation_data = val_ds, callbacks=callbacks)
+    history = m.history
+    model_instance = m.model
 
     # Визуализация
 
-    pe = plotter_evaluator(history, model, class_names)
+    pe = plotter_evaluator(history, model_instance, class_names)
     pe.calc_predictions(val_ds)
 
     ## Графики потерь и точности
-
     # Высокой должна быть и accuracy и val_accuracy. Первая -- точность на обучающей выборке, вторая -- на тестовой. 
     # Когда/если точность на обучающей выборке начинает превосходить точность на тестовой, продолжать обучение не следует.
 
@@ -168,13 +103,11 @@ def data_augmentation(images):
     pe.plot_loss_accuracy()
 
     ## Вычисление отчёта о качестве классификации
-
     # Значения accuracy, recall, f1 должны быть высокими.
 
     pe.print_report()
 
     ## Матрица запутанности
-
     # Хорший способ понять, как именно нейросеть ошибается
 
     pe.plot_confusion_matrix()
diff --git a/image_recognition/modules/__init__.py b/image_recognition/modules/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/image_recognition/modules/importer.py b/image_recognition/modules/importer.py
new file mode 100644
index 0000000..b8fd568
--- /dev/null
+++ b/image_recognition/modules/importer.py
@@ -0,0 +1,74 @@
+
+
+import keras
+from tensorflow import data as tf_data
+import numpy as np
+
+class importer:
+
+    __SEED = 42
+    __data_augmentation_layers: list
+
+    train_ds: tf_data.Dataset
+    validation_ds: tf_data.Dataset
+
+    class_names: set
+    num_classes: int
+
+    def __init__(self, image_size, batch_size, data_directory, validation_split):
+       
+        self.train_ds, self.validation_ds = keras.utils.image_dataset_from_directory(
+            data_directory,
+            validation_split=validation_split,
+            subset="both",
+            seed=self.__SEED,
+            label_mode="categorical",
+            shuffle=True,
+            image_size=image_size,
+            batch_size=batch_size,
+            color_mode="grayscale",
+            crop_to_aspect_ratio=True,
+        )
+
+        self.__get_stats()
+
+    def generate_augmentation_layers(self, zoom_factor: float, move_factor: float, rotation_factor: float):
+        
+        self.__data_augmentation_layers = [
+            keras.layers.RandomRotation(rotation_factor),
+            keras.layers.RandomZoom(
+                height_factor=[-zoom_factor,zoom_factor],
+                width_factor=[-zoom_factor,zoom_factor],
+                fill_mode="constant",
+                fill_value=255.0
+            ),
+            keras.layers.RandomTranslation(
+                height_factor = [-move_factor, move_factor],
+                width_factor = [-move_factor, move_factor],
+                fill_mode="constant",
+                fill_value=255.0,
+            ),
+            keras.layers.Rescaling(1.0 / 255)
+        ]
+    
+    def apply_augmentation(self):
+        
+        self.train_ds = self.train_ds.map(
+            lambda img, label: (self.__data_augmentation(img), label),
+            num_parallel_calls=tf_data.AUTOTUNE,
+        )
+
+    def __data_augmentation(self, images):
+        for layer in self.__data_augmentation_layers:
+            images = layer(images)
+        return images
+
+    def __get_stats(self):
+
+        # TODO: Rework
+        labels = np.array([])
+        for _, y in self.validation_ds: # type: ignore
+            labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)])
+
+        self.class_names = set(labels)
+        self.num_classes = len(self.class_names)
\ No newline at end of file
diff --git a/image_recognition/modules/model.py b/image_recognition/modules/model.py
new file mode 100644
index 0000000..f41c04d
--- /dev/null
+++ b/image_recognition/modules/model.py
@@ -0,0 +1,71 @@
+
+
+import keras
+from keras.models import Sequential
+from keras.layers import Conv2D, MaxPooling2D, Dense, Flatten, Dropout
+from keras.optimizers import RMSprop
+from keras.callbacks import History
+
+from keras.callbacks import ReduceLROnPlateau
+from tensorflow import data as tf_data
+
+class model:
+
+    model: keras.Model
+    history: History
+    callbacks = []
+
+    def __init__(self, image_size: tuple[int,int], num_classes: int, dense_layers_number: int, conv_descriptor: tuple):
+        
+        model = Sequential()
+
+        model.add(keras.Input(shape=image_size + (1, )))
+
+        model.add(Conv2D(16, (3,3), 1, activation='relu'))
+        model.add(MaxPooling2D())
+        # model.add(Dropout(0.25))
+
+        model.add(Conv2D(32, (3,3), 1, activation='relu'))
+        model.add(MaxPooling2D())
+        # model.add(Dropout(0.25))
+
+        model.add(Conv2D(16, (3,3), 1, activation='relu'))
+        model.add(MaxPooling2D())
+        # model.add(Dropout(0.25))
+
+        model.add(Flatten())
+        model.add(Dense(256, activation='relu'))
+        # model.add(Dropout(0.25))
+
+        model.add(Dense(num_classes, activation = "softmax"))
+
+        self.model = model
+
+        model.summary()
+
+    def compile(self):
+
+        # Компиляция модели
+        optimizer = RMSprop(learning_rate=0.001, rho=0.9, epsilon=1e-08)
+        self.model.compile(optimizer = optimizer , loss = "categorical_crossentropy", metrics=["accuracy"])
+        # self.model.summary()
+
+    def train(self, train_ds: tf_data.Dataset, epochs: int, validation_data: tf_data.Dataset):
+        
+        self.history = self.model.fit(train_ds, epochs = epochs, validation_data = validation_data, callbacks=self.callbacks)
+    
+    def init_learning_rate_reduction(self):
+        
+        learning_rate_reduction = ReduceLROnPlateau(
+            monitor='accuracy',
+            patience=3, 
+            verbose=1, 
+            factor=0.5, 
+            min_lr=0.00001
+        )
+
+        self.callbacks.append(learning_rate_reduction)
+    
+    def init_save_at_epoch(self):
+        
+        self.callbacks.append(keras.callbacks.ModelCheckpoint("save_at_{epoch}.keras"))
\ No newline at end of file
diff --git a/image_recognition/visualization.py b/image_recognition/modules/visualization.py
similarity index 81%
rename from image_recognition/visualization.py
rename to image_recognition/modules/visualization.py
index e2ddffb..5e254b6 100644
--- a/image_recognition/visualization.py
+++ b/image_recognition/modules/visualization.py
@@ -1,64 +1,65 @@
-
-
-import clang
-import matplotlib.pyplot as plt
-import numpy as np
-from sklearn.metrics import classification_report
-import keras.callbacks as cb
-import keras as k
-from scikitplot.metrics import plot_confusion_matrix 
-
-class plotter_evaluator:
-
-    model: k.Model
-    history: cb.History
-    class_names: list
-    labels: np.ndarray
-    pred_labels: list
-
-    def __init__(self, history: cb.History, model: k.Model, class_names: set):
-        self.history = history
-        self.model = model
-
-        if type(list(class_names)[0]) is not str:
-            self.class_names=list(map(str,class_names))
-        else:
-            self.class_names = list(class_names)
-
-    def calc_predictions(self, test_values: list):
-
-        labels =  np.array([])
-        for _, y in test_values:
-            labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)])
-
-        self.labels = labels
-
-        Predictions = self.model.predict(test_values)
-        self.pred_labels = np.argmax(Predictions, axis = 1)
-
-    def plot_loss_accuracy(self):
-        
-        history = self.history
-        fig = plt.figure(figsize=(9,5))
-
-        plt.subplot(211)
-        plt.plot(history.history['loss'], color='teal', label='loss')
-        plt.plot(history.history['val_loss'], color='orange', label='val_loss')
-        plt.legend(loc="upper left")
-
-        plt.subplot(212)
-        plt.plot(history.history['accuracy'], color='teal', label='accuracy')
-        plt.plot(history.history['val_accuracy'], color='orange', label='val_accuracy')
-        plt.legend(loc="upper left")
-
-        fig.suptitle('Loss and Accuracy', fontsize=19)
-        plt.show()
-
-    def print_report(self):
-
-        print(classification_report(self.labels, self.pred_labels, target_names=self.class_names))
-
-    def plot_confusion_matrix(self):
-        
-        fig, ax = plt.subplots(1, 2, figsize = (25,  8))
+
+
+import clang
+import matplotlib.pyplot as plt
+import numpy as np
+from sklearn.metrics import classification_report
+from keras.callbacks import History
+import keras
+from scikitplot.metrics import plot_confusion_matrix 
+from tensorflow import data as tf_data
+
+class plotter_evaluator:
+
+    model: keras.Model
+    history: History
+    class_names: list
+    labels: np.ndarray
+    pred_labels: list
+
+    def __init__(self, history: History, model: keras.Model, class_names: set):
+        self.history = history
+        self.model = model
+
+        if type(list(class_names)[0]) is not str:
+            self.class_names=list(map(str,class_names))
+        else:
+            self.class_names = list(class_names)
+
+    def calc_predictions(self, test_values: tf_data.Dataset):
+
+        labels =  np.array([])
+        for _, y in test_values: # type: ignore
+            labels = np.concatenate([labels, np.argmax(y.numpy(), axis=-1)])
+
+        self.labels = labels
+
+        Predictions = self.model.predict(test_values)
+        self.pred_labels = np.argmax(Predictions, axis = 1)
+
+    def plot_loss_accuracy(self):
+        
+        history = self.history
+        fig = plt.figure(figsize=(9,5))
+
+        plt.subplot(211)
+        plt.plot(history.history['loss'], color='teal', label='loss')
+        plt.plot(history.history['val_loss'], color='orange', label='val_loss')
+        plt.legend(loc="upper left")
+
+        plt.subplot(212)
+        plt.plot(history.history['accuracy'], color='teal', label='accuracy')
+        plt.plot(history.history['val_accuracy'], color='orange', label='val_accuracy')
+        plt.legend(loc="upper left")
+
+        fig.suptitle('Loss and Accuracy', fontsize=19)
+        plt.show()
+
+    def print_report(self):
+
+        print(classification_report(self.labels, self.pred_labels, target_names=self.class_names))
+
+    def plot_confusion_matrix(self):
+        
+        fig, ax = plt.subplots(1, 2, figsize = (25,  8))
         ax = plot_confusion_matrix(self.labels, self.pred_labels, ax = ax[0], cmap= 'YlGnBu')
\ No newline at end of file