Chapter 4. CNN for images¶
This notebook illustrates the use of CNN models for the identification of images of filtered resonant arguments for the $s-s_6-g_5 + g_6$ node secular resonance. To save disk space, we are providing a limited number of images, so the results are purely illustrative. Below there are the necessary commands to run the model.
Perform the necessary imports
In [1]:
Copied!
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import Flatten
from tensorflow.keras.layers import Conv2D
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import Dropout
from tensorflow.keras.optimizers import SGD
from tensorflow.keras.callbacks import ModelCheckpoint
import numpy as np
import pandas as pd
from PIL import Image
from sys import getsizeof
import copy
import time
import datetime
import tracemalloc
import tensorflow as tf
import matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.layers import Flatten
from tensorflow.keras.layers import Conv2D
from tensorflow.keras.layers import MaxPooling2D
from tensorflow.keras.layers import Dropout
from tensorflow.keras.optimizers import SGD
from tensorflow.keras.callbacks import ModelCheckpoint
import numpy as np
import pandas as pd
from PIL import Image
from sys import getsizeof
import copy
import time
import datetime
import tracemalloc
2024-03-19 10:52:51.156832: I tensorflow/core/util/port.cc:113] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`. 2024-03-19 10:52:51.181113: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered 2024-03-19 10:52:51.181141: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered 2024-03-19 10:52:51.182163: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered 2024-03-19 10:52:51.188511: I tensorflow/core/platform/cpu_feature_guard.cc:182] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations. To enable the following instructions: AVX2 AVX512F AVX512_VNNI FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags. 2024-03-19 10:52:51.687617: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT
Start tracking time and define the orbits classes
In [2]:
Copied!
start_time = time.time()
# Starting monitoring the process
tracemalloc.start()
class_names = ['circulation state', 'switching orbits', 'libration state']
start_time = time.time()
# Starting monitoring the process
tracemalloc.start()
class_names = ['circulation state', 'switching orbits', 'libration state']
Define a function to read the images
In [3]:
Copied!
# read data
def read_data(filename, Images_loc, Labels, set_type):
data_ast = pd.read_csv(filename,
skiprows=0,
header=None,
delim_whitespace=True,
index_col=None,
names = Labels,
low_memory=False,
dtype={'id': np.int64,
'a': np.float64,
'e': np.float64,
'sin_i': np.float64,
'Mag': np.float64,
'freq': np.float64,
'label': np.int64
}
)
n_lines =int(len(data_ast))
data= data_ast.iloc[0:n_lines, :]
data_id = list(data_ast.id)
img = [Images_loc + str("{:07d}".format(ast_id)) + '.png' for ast_id in data_id]
width, height = Image.open(img[0]).convert('1').size
images = [np.array(Image.open(x).convert('1').getdata()).reshape(width, height) for x in img]
im_labels = data_ast.label
print(f'The size of the variable images is : {getsizeof(images)} bytes')
print(f'We have {len(images)} images ({width} X {height} pixels) in the ', set_type, ' set, belonging to '
f'{len(set(im_labels))} classes:')
for i in range(len(set(im_labels))):
print(f' {len([x for x in im_labels if x == i])} asteroids in {class_names[i]} (label: {i})')
print()
return images, im_labels, data, data_id
# read data
def read_data(filename, Images_loc, Labels, set_type):
data_ast = pd.read_csv(filename,
skiprows=0,
header=None,
delim_whitespace=True,
index_col=None,
names = Labels,
low_memory=False,
dtype={'id': np.int64,
'a': np.float64,
'e': np.float64,
'sin_i': np.float64,
'Mag': np.float64,
'freq': np.float64,
'label': np.int64
}
)
n_lines =int(len(data_ast))
data= data_ast.iloc[0:n_lines, :]
data_id = list(data_ast.id)
img = [Images_loc + str("{:07d}".format(ast_id)) + '.png' for ast_id in data_id]
width, height = Image.open(img[0]).convert('1').size
images = [np.array(Image.open(x).convert('1').getdata()).reshape(width, height) for x in img]
im_labels = data_ast.label
print(f'The size of the variable images is : {getsizeof(images)} bytes')
print(f'We have {len(images)} images ({width} X {height} pixels) in the ', set_type, ' set, belonging to '
f'{len(set(im_labels))} classes:')
for i in range(len(set(im_labels))):
print(f' {len([x for x in im_labels if x == i])} asteroids in {class_names[i]} (label: {i})')
print()
return images, im_labels, data, data_id
Load training images and labels
In [4]:
Copied!
filename_train = './TRAINING/res_status_all'
Images_loc_train = './TRAINING/res_'
names_train = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'training'
train_images, train_labels, train_data, train_id = read_data(filename_train,Images_loc_train, names_train, set_type)
min_pixel = min(list(map(min, train_images[0])))
max_pixel = max(list(map(max, train_images[0])))
print(f'The pixel values of each image vary from {min_pixel} to {max_pixel}')
filename_train = './TRAINING/res_status_all'
Images_loc_train = './TRAINING/res_'
names_train = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'training'
train_images, train_labels, train_data, train_id = read_data(filename_train,Images_loc_train, names_train, set_type)
min_pixel = min(list(map(min, train_images[0])))
max_pixel = max(list(map(max, train_images[0])))
print(f'The pixel values of each image vary from {min_pixel} to {max_pixel}')
The size of the variable images is : 792 bytes We have 78 images (100 X 100 pixels) in the training set, belonging to 3 classes: 32 asteroids in circulation state (label: 0) 14 asteroids in switching orbits (label: 1) 32 asteroids in libration state (label: 2) The pixel values of each image vary from 0 to 255
Load test images and labels
In [5]:
Copied!
filename_test = './TEST/res_status_all'
Images_loc_test = './TEST/res_'
names_test = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'testing'
test_images, test_labels, test_data, test_id = read_data(filename_test,Images_loc_test, names_test, set_type)
filename_test = './TEST/res_status_all'
Images_loc_test = './TEST/res_'
names_test = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'testing'
test_images, test_labels, test_data, test_id = read_data(filename_test,Images_loc_test, names_test, set_type)
The size of the variable images is : 472 bytes We have 50 images (100 X 100 pixels) in the testing set, belonging to 3 classes: 24 asteroids in circulation state (label: 0) 4 asteroids in switching orbits (label: 1) 22 asteroids in libration state (label: 2)
Load validation set images and labels
In [6]:
Copied!
filename_val = './VALIDATION/res_status_all'
Images_loc_val = './VALIDATION/res_'
names_val = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'validation'
val_images, val_labels, val_data, val_id = read_data(filename_val, Images_loc_val, names_val, set_type)
# preprocessing the data: rescale the pixels value to range from 0 to 1
train_images = train_images / max_pixel
test_images = test_images / max_pixel
val_images = val_images / max_pixel
filename_val = './VALIDATION/res_status_all'
Images_loc_val = './VALIDATION/res_'
names_val = ['id', 'a', 'e', 'sin_i', 'Mag', 'freq', 'label']
set_type = 'validation'
val_images, val_labels, val_data, val_id = read_data(filename_val, Images_loc_val, names_val, set_type)
# preprocessing the data: rescale the pixels value to range from 0 to 1
train_images = train_images / max_pixel
test_images = test_images / max_pixel
val_images = val_images / max_pixel
The size of the variable images is : 312 bytes We have 29 images (100 X 100 pixels) in the validation set, belonging to 2 classes: 14 asteroids in circulation state (label: 0) 0 asteroids in switching orbits (label: 1)
Define data augmentation
In [7]:
Copied!
# Define data augmentation
data_augmentation = keras.Sequential(
[tf.keras.layers.experimental.preprocessing.RandomFlip("horizontal", input_shape=(100,100,1)),
tf.keras.layers.experimental.preprocessing.RandomRotation(0.1),
tf.keras.layers.experimental.preprocessing.RandomZoom(0.1),
]
)
# Define data augmentation
data_augmentation = keras.Sequential(
[tf.keras.layers.experimental.preprocessing.RandomFlip("horizontal", input_shape=(100,100,1)),
tf.keras.layers.experimental.preprocessing.RandomRotation(0.1),
tf.keras.layers.experimental.preprocessing.RandomZoom(0.1),
]
)
Set up the model: Here is the VGG model with data augmentation (DA) and dropout (DO).
In [8]:
Copied!
# Set up the model
def define_model_VGG():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(train_images.shape[1], (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(train_images.shape[1], (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
model.add(MaxPooling2D((2, 2)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
# compile model
opt = SGD(learning_rate=0.001, momentum=0.9)
return model
# Set up the model
def define_model_VGG():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(train_images.shape[1], (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(train_images.shape[1], (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same'))
model.add(MaxPooling2D((2, 2)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
# compile model
opt = SGD(learning_rate=0.001, momentum=0.9)
return model
Here is the Inception model with DA and DO.
In [9]:
Copied!
# Set up the model
def define_model_Inception():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(64, (1, 1), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (5, 5), activation='relu', kernel_initializer='he_uniform', padding='same'))
model.add(MaxPooling2D((3, 3)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
return model
# Set up the model
def define_model_Inception():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(64, (1, 1), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (5, 5), activation='relu', kernel_initializer='he_uniform', padding='same'))
model.add(MaxPooling2D((3, 3)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
return model
The ResNet model with DA and DO.
In [10]:
Copied!
# Set up the model
def define_model_ResNet():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (3, 3), activation='linear', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
return model
# Set up the model
def define_model_ResNet():
model = Sequential()
model.add(data_augmentation)
model.add(Conv2D(64, (3, 3), activation='relu', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Conv2D(64, (3, 3), activation='linear', kernel_initializer='he_uniform', padding='same', input_shape=(100, 100,1)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation='relu', kernel_initializer='he_uniform'))
model.add(Dense(3, activation='softmax'))
return model
This checkpoint object will store the model parameters in the file "weights.hdf5"
In [12]:
Copied!
checkpoint = ModelCheckpoint('./weights.hdf5',
save_weights_only=True,
monitor='accuracy',
mode='max',
verbose=1,
save_best_only=True,)
checkpoint = ModelCheckpoint('./weights.hdf5',
save_weights_only=True,
monitor='accuracy',
mode='max',
verbose=1,
save_best_only=True,)
Start monitoring time
In [13]:
Copied!
start_time = time.time()
# starting the monitoring
tracemalloc.start()
start_time = time.time()
# starting the monitoring
tracemalloc.start()
Compile the module, we are using Inception here, but feel free to explore the other models
In [14]:
Copied!
model=define_model_Inception()
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics='accuracy')
model=define_model_Inception()
model.summary()
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics='accuracy')
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
sequential (Sequential) (None, 100, 100, 1) 0
conv2d (Conv2D) (None, 100, 100, 64) 128
conv2d_1 (Conv2D) (None, 100, 100, 64) 36928
conv2d_2 (Conv2D) (None, 100, 100, 64) 102464
max_pooling2d (MaxPooling2 (None, 33, 33, 64) 0
D)
dropout (Dropout) (None, 33, 33, 64) 0
flatten (Flatten) (None, 69696) 0
dense (Dense) (None, 128) 8921216
dense_1 (Dense) (None, 3) 387
=================================================================
Total params: 9061123 (34.57 MB)
Trainable params: 9061123 (34.57 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
Fit the model, using the checkpoint as a callback. This usually takes up to 15 minutes, depending on your machine.
In [15]:
Copied!
x = model.fit(train_images, train_labels, epochs=12, callbacks=[checkpoint],
verbose = 0, validation_data=(val_images, val_labels))
model.load_weights('./weights.hdf5')
end_time = time.time()
exec_time = datetime.timedelta(seconds=(end_time - start_time))
print(f'\n --- The execution time was: {exec_time} (h:m:s) ---')
# displaying the memory: The output is given in form of (current, peak), i.e, current memory is the memory the code is currently using, Peak memory is the maximum space the program used while executing.
print(tracemalloc.get_traced_memory())
x = model.fit(train_images, train_labels, epochs=12, callbacks=[checkpoint],
verbose = 0, validation_data=(val_images, val_labels))
model.load_weights('./weights.hdf5')
end_time = time.time()
exec_time = datetime.timedelta(seconds=(end_time - start_time))
print(f'\n --- The execution time was: {exec_time} (h:m:s) ---')
# displaying the memory: The output is given in form of (current, peak), i.e, current memory is the memory the code is currently using, Peak memory is the maximum space the program used while executing.
print(tracemalloc.get_traced_memory())
Epoch 1: accuracy improved from -inf to 0.25641, saving model to ./weights.hdf5 Epoch 2: accuracy improved from 0.25641 to 0.37179, saving model to ./weights.hdf5 Epoch 3: accuracy improved from 0.37179 to 0.42308, saving model to ./weights.hdf5 Epoch 4: accuracy improved from 0.42308 to 0.47436, saving model to ./weights.hdf5 Epoch 5: accuracy did not improve from 0.47436 Epoch 6: accuracy did not improve from 0.47436 Epoch 7: accuracy did not improve from 0.47436 Epoch 8: accuracy improved from 0.47436 to 0.61538, saving model to ./weights.hdf5 Epoch 9: accuracy improved from 0.61538 to 0.66667, saving model to ./weights.hdf5 Epoch 10: accuracy did not improve from 0.66667 Epoch 11: accuracy did not improve from 0.66667 Epoch 12: accuracy improved from 0.66667 to 0.70513, saving model to ./weights.hdf5 --- The execution time was: 0:00:23.423197 (h:m:s) --- (19634630, 91573820)
Plot diagnostic learning curves
In [17]:
Copied!
def summarize_diagnostics(x):
fig = plt.figure()
figure = fig.add_subplot(211)
figure.plot(x.epoch, x.history['loss'], color='blue', label='train')
figure.plot(x.epoch, x.history['val_loss'],color='orange', label='val')
#plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.title('Cross Entropy Loss: VGG')
figure = fig.add_subplot(212)
figure.plot(x.epoch, x.history['accuracy'], color='blue', label='train')
figure.plot(x.epoch, x.history['val_accuracy'],color='orange', label='val')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
plt.title('Classification Accuracy: VGG')
#plt.show()
plt.tight_layout()
fig.savefig('history_model_VGG.png', format='png', dpi=300)
plt.show()
plt.close(fig)
def summarize_diagnostics(x):
fig = plt.figure()
figure = fig.add_subplot(211)
figure.plot(x.epoch, x.history['loss'], color='blue', label='train')
figure.plot(x.epoch, x.history['val_loss'],color='orange', label='val')
#plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()
plt.title('Cross Entropy Loss: VGG')
figure = fig.add_subplot(212)
figure.plot(x.epoch, x.history['accuracy'], color='blue', label='train')
figure.plot(x.epoch, x.history['val_accuracy'],color='orange', label='val')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
plt.title('Classification Accuracy: VGG')
#plt.show()
plt.tight_layout()
fig.savefig('history_model_VGG.png', format='png', dpi=300)
plt.show()
plt.close(fig)
Make the predictions
In [18]:
Copied!
summarize_diagnostics(x)
predictions = model.predict(test_images)
predict_label = [int(np.argmax(x)) for x in predictions]
predict_acc = [100 * max(x) for x in predictions]
predicted_data = copy.deepcopy(test_data)
predicted_data['predicted_label'] = list(predict_label)
predicted_data.to_csv(r's_pred_data.csv', index=False, header=False, sep=' ', float_format='%.7f')
print()
summarize_diagnostics(x)
predictions = model.predict(test_images)
predict_label = [int(np.argmax(x)) for x in predictions]
predict_acc = [100 * max(x) for x in predictions]
predicted_data = copy.deepcopy(test_data)
predicted_data['predicted_label'] = list(predict_label)
predicted_data.to_csv(r's_pred_data.csv', index=False, header=False, sep=' ', float_format='%.7f')
print()
2/2 [==============================] - 0s 62ms/step
Show the first images in the test data
In [19]:
Copied!
def show_images():
fig = plt.figure(figsize=(8, 12))
for i in range(50):
plt.subplot(10, 5, i + 1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
plt.imshow(test_images[i])
color = 'blue'
plt.xlabel("{} ({:2.0f}%)".format(predict_label[i], predict_acc[predict_label[i]]),color=color, fontsize=10)
plt.ylabel("{}".format(test_id[i]), color=color, fontsize=10)
plt.subplots_adjust(hspace=0.3, wspace=0)
# plt.show()
fig.savefig('predicted_data.png', format='png', dpi=300)
plt.show()
plt.close(fig)
show_images()
print("End")
def show_images():
fig = plt.figure(figsize=(8, 12))
for i in range(50):
plt.subplot(10, 5, i + 1)
plt.xticks([])
plt.yticks([])
plt.grid(False)
plt.imshow(test_images[i])
color = 'blue'
plt.xlabel("{} ({:2.0f}%)".format(predict_label[i], predict_acc[predict_label[i]]),color=color, fontsize=10)
plt.ylabel("{}".format(test_id[i]), color=color, fontsize=10)
plt.subplots_adjust(hspace=0.3, wspace=0)
# plt.show()
fig.savefig('predicted_data.png', format='png', dpi=300)
plt.show()
plt.close(fig)
show_images()
print("End")
End
In [ ]:
Copied!