x

b2n/augmentation/generator.py

@@ -2,7 +2,9 @@
 from scipy.stats import norm
 import cv2
 import random
-from tensorflow.keras.preprocessing.image import ImageDataGenerator
+#from tensorflow.keras.preprocessing.image import ImageDataGenerator
+import tensorflow as tf
+import tensorflow.python.keras.layers as layers
 import random
 
 def add_noise(img):
@@ -141,25 +143,30 @@ def invert_noise(img):
 
 def augmentation(x, y, Batch_Size = 32, Width_Shift_Range = 3, Height_Shift_Range = 0, 
                     Brightness_Range = [0.4,1.6],
-                    Rotation_Angle = 5,
+                    Rotation_Angle = 0.01,
                     ZoomRange_Out = 0.1,
                     ZoomRange_In = 0.1,
                     ShearRange= 3,
                     Channel_Shift=0.4,
                     preprocessing_function=preprocessing):
-    datagen = ImageDataGenerator(width_shift_range=Width_Shift_Range, 
-                             height_shift_range=Height_Shift_Range,
-                             brightness_range=Brightness_Range,
-                             #zoom_range=[1-ZoomRange_In, 1+ZoomRange_Out],
-                             rotation_range=Rotation_Angle,
-                             #channel_shift_range=Channel_Shift,
-                             fill_mode='nearest',
-                             shear_range=ShearRange
-                             ,preprocessing_function=preprocessing_function
-                             ,dtype=float)
-    #datagen.fit(x)
-    return datagen.flow(x, y, batch_size=Batch_Size)
+#    datagen = ImageDataGenerator(width_shift_range=Width_Shift_Range, 
+#                             height_shift_range=Height_Shift_Range,
+#                             brightness_range=Brightness_Range,
+#                             #zoom_range=[1-ZoomRange_In, 1+ZoomRange_Out],
+#                             rotation_range=Rotation_Angle,
+#                             #channel_shift_range=Channel_Shift,
+#                             fill_mode='nearest',
+#                             shear_range=ShearRange
+#                             ,preprocessing_function=preprocessing_function
+#                             ,dtype=float)
+
+    preprocessing = tf.keras.Sequential([
+        tf.keras.layers.RandomZoom([ZoomRange_In, ZoomRange_Out]),
+        tf.keras.layers.RandomRotation((-Rotation_Angle, Rotation_Angle))])
+
+    datagen = tf.data.Dataset.from_tensor_slices(x, y)
+    return datagen #datagen.map(preprocessing)   
+
 
 def no_augmentation(x, y, Batch_Size = 32):
-    datagen = ImageDataGenerator(dtype=float)
-    return datagen.flow(x, y, batch_size=Batch_Size)
+    return tf.data.Dataset.from_tensor_slices(x, y)

b2n/data/ziffer.py

@@ -1,7 +1,5 @@
 import os
-import shutil
 from PIL import Image 
-from tensorflow import keras
 import numpy as np
 
 

b2n/encodings/class_encoding.py

@@ -1,5 +1,3 @@
-import tensorflow.keras as keras
-import math
 import numpy as np
 
 def class_encoding(y_train, nb_classes):

b2n/plotfunctions.py

@@ -40,6 +40,24 @@ def plot_dataset_it(data_iter, columns=9, rows=5, nb_classes=100, classdecoding=
                 plt.axhline(y=y,color='yellow')
     plt.show()
 
+def plot_dataset_ds(data_set, columns=9, rows=5, nb_classes=100, classdecoding=class_decoding):
+
+    fig = plt.figure(figsize=(18, 11))
+    data_iter = iter(data_set)
+
+    for i in range(1, columns*rows +1):
+        img, label = next(data_iter)
+        fig.add_subplot(rows, columns, i)
+        plt.xticks([0.2, 0.4, 0.6, 0.8])
+        plt.title(str(class_decoding(np.array(label), nb_classes).reshape(-1)[0] ))  # set title
+        plt.imshow(np.array(img).astype(np.uint8), aspect='1.6', cmap='gray', extent=[0, 1, 0, 1], )
+        ax=plt.gca()
+        ax.get_xaxis().set_visible(False) 
+        # yellow lines
+        for y in np.arange(0.2, 0.8, 0.2):
+                plt.axhline(y=y,color='yellow')
+    plt.show()
+
 def plot_acc_loss(history, modelname="modelname"):
     fig, (ax1, ax2) = plt.subplots(1, 2)
     fig.suptitle(modelname)

dig-class100-s2_bw.ipynb

@@ -975,7 +975,7 @@
     "\n",
     "#annealer = LearningRateScheduler(lambda x: 1e-3 * 0.98 ** x, verbose=0)\n",
     "earlystop = EarlyStopping(monitor='val_loss', mode='min' , patience=25, verbose=1,  restore_best_weights=True)\n",
-    "mc = ModelCheckpoint(output_tmp_dir + model_filename + '_best.h5', monitor='val_loss', mode='min', verbose=1, save_best_only=True, initial_value_threshold=.9)\n",
+    "mc = ModelCheckpoint(output_tmp_dir + model_filename + '_best.keras', monitor='val_loss', mode='min', verbose=1, save_best_only=True, initial_value_threshold=.9)\n",
     "reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.85, patience=3, min_lr=1e-5, verbose=1)\n",
     "history = model.fit(train_iterator, \n",
     "                validation_data=validation_iterator, \n",
@@ -985,7 +985,7 @@
     "                verbose=1)\n",
     "\n",
     "\n",
-    "#model.save(model_filename + '.h5')\n",
+    "#model.save(model_filename + '.keras')\n",
     "\n",
     "plot_acc_loss(history, \"Trainingsresults\")"
    ]
@@ -1039,7 +1039,7 @@
     }
    ],
    "source": [
-    "model = keras.models.load_model(output_tmp_dir + model_filename + '_best.h5')\n",
+    "model = keras.models.load_model(output_tmp_dir + model_filename + '_best.keras')\n",
     "false_predicted_files = predict_meter_digits(model, x_data=xz_data, y_data=yz_data, f_data=fz_data, \n",
     "                                            max_delta=0.11)\n",
     "fpf = pd.DataFrame(false_predicted_files)\n",

dig-class100_s0.ipynb

@@ -392,7 +392,7 @@
     "\n",
     "#annealer = LearningRateScheduler(lambda x: 1e-3 * 0.98 ** x, verbose=0)\n",
     "earlystop = EarlyStopping(monitor='val_accuracy', mode='max' , patience=25, verbose=1,  restore_best_weights=True)\n",
-    "mc = ModelCheckpoint(model_filename + '_best.h5', monitor='val_accuracy', mode='max', verbose=1, save_best_only=True)\n",
+    "mc = ModelCheckpoint(model_filename + '_best.keras', monitor='val_accuracy', mode='max', verbose=1, save_best_only=True)\n",
     "reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.8, patience=5, min_lr=1e-5, verbose=1)\n",
     "history = model.fit(train_iterator, \n",
     "                validation_data=validation_iterator, \n",
@@ -402,7 +402,7 @@
     "                verbose=1)\n",
     "\n",
     "\n",
-    "#model.save(model_filename + '.h5')\n",
+    "#model.save(model_filename + '.keras')\n",
     "\n",
     "plot_acc_loss(history, \"Trainingsresults\")"
    ]
@@ -424,7 +424,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "model = keras.models.load_model(model_filename + '_best.h5')\n",
+    "model = keras.models.load_model(model_filename + '_best.keras')\n",
     "xz_data, yz_data, fz_data = shuffle(xz_data, yz_data, fz_data)\n",
     "false_predicted_files = predict_meter_digits(model, x_data=xz_data, y_data=yz_data, f_data=fz_data, \n",
     "                                            max_delta=0.)\n",

dig_class100-s1.ipynb

@@ -1247,7 +1247,7 @@
     "                verbose=1)\n",
     "\n",
     "\n",
-    "#model.save(model_filename + '.h5')\n",
+    "#model.save(model_filename + '.keras')\n",
     "\n",
     "plot_acc_loss(history, \"Trainingsresults\")"
    ]

requirements.txt

@@ -5,9 +5,9 @@ lxml_html_clean
 scipy
 scikit-learn
 pandas
-tensorflow>=2.12.1; sys_platform != 'darwin'
-tensorflow-macos>=2.12.1; sys_platform == 'darwin'
-tensorflow-metal==0.8; sys_platform == 'darwin'
+tensorflow<=2.17.0; sys_platform != 'darwin'
+#tensorflow-macos>=2.12.1; sys_platform == 'darwin'
+#tensorflow-metal==0.8; sys_platform == 'darwin'
 tensorflow-model-optimization
 opencv-python
 notebook