dl-lab-experiments
August 11, 2024
1 3. Implement a feed forward neural network with three hidden
layers for classification on cifar-10 dataset
[ ]: import tensorflow as tf
from keras import models,layers
from keras.datasets import cifar10
from keras.utils import to_categorical
[ ]: # Load CIFAR-10 dataset
(X_train, y_train), (X_test, y_test) = cifar10.load_data()
[ ]: # Normalize pixel values to be between 0 and 1
X_train, X_test= X_train / 255.0, X_test / 255.0
[ ]: print(f'X_train shape: {X_train.shape}\ny_train shape: {y_train.shape}')
print(f'X_test shape: {X_test.shape}\ny_test shape: {y_test.shape}')
X_train shape: (50000, 32, 32, 3)
y_train shape: (50000, 1)
X_test shape: (10000, 32, 32, 3)
y_test shape: (10000, 1)
[ ]: # Convert labels to one-hot encoding
y_train = to_categorical(y_train, 10)
y_test = to_categorical(y_test, 10)
[ ]: print(f'X_train shape: {X_train.shape}\ny_train shape: {y_train.shape}')
print(f'X_test shape: {X_test.shape}\ny_test shape: {y_test.shape}')
X_train shape: (50000, 32, 32, 3)
y_train shape: (50000, 10)
X_test shape: (10000, 32, 32, 3)
y_test shape: (10000, 10)
[ ]: # Define the model
model = models.Sequential()
1
# Flatten the input for the fully connected layer
model.add(layers.Flatten(input_shape=(32, 32, 3)))
# Three hidden layers with ReLU activation
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(256, activation='relu'))
model.add(layers.Dense(128, activation='relu'))
# Output layer with softmax activation for classification
model.add(layers.Dense(10, activation='softmax'))
[ ]: # Compile the model
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
[ ]: # Display the model summary
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 3072) 0
dense (Dense) (None, 512) 1573376
dense_1 (Dense) (None, 256) 131328
dense_2 (Dense) (None, 128) 32896
dense_3 (Dense) (None, 10) 1290
=================================================================
Total params: 1738890 (6.63 MB)
Trainable params: 1738890 (6.63 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
[ ]: # Train the model
history = model.fit(X_train, y_train, epochs=15, validation_data=(X_test,␣
↪y_test))
Epoch 1/15
1563/1563 [==============================] - 13s 5ms/step - loss: 1.8592 -
accuracy: 0.3278 - val_loss: 1.6996 - val_accuracy: 0.3868
Epoch 2/15
2
1563/1563 [==============================] - 7s 4ms/step - loss: 1.6749 -
accuracy: 0.3996 - val_loss: 1.6449 - val_accuracy: 0.4069
Epoch 3/15
1563/1563 [==============================] - 7s 5ms/step - loss: 1.5880 -
accuracy: 0.4330 - val_loss: 1.5731 - val_accuracy: 0.4367
Epoch 4/15
1563/1563 [==============================] - 7s 4ms/step - loss: 1.5302 -
accuracy: 0.4517 - val_loss: 1.5233 - val_accuracy: 0.4575
Epoch 5/15
1563/1563 [==============================] - 6s 4ms/step - loss: 1.4919 -
accuracy: 0.4660 - val_loss: 1.5045 - val_accuracy: 0.4708
Epoch 6/15
1563/1563 [==============================] - 7s 5ms/step - loss: 1.4612 -
accuracy: 0.4775 - val_loss: 1.5188 - val_accuracy: 0.4588
Epoch 7/15
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4365 -
accuracy: 0.4868 - val_loss: 1.4675 - val_accuracy: 0.4818
Epoch 8/15
1563/1563 [==============================] - 7s 4ms/step - loss: 1.4100 -
accuracy: 0.4952 - val_loss: 1.4518 - val_accuracy: 0.4887
Epoch 9/15
1563/1563 [==============================] - 7s 4ms/step - loss: 1.3890 -
accuracy: 0.5034 - val_loss: 1.4769 - val_accuracy: 0.4796
Epoch 10/15
1563/1563 [==============================] - 7s 5ms/step - loss: 1.3650 -
accuracy: 0.5095 - val_loss: 1.4687 - val_accuracy: 0.4818
Epoch 11/15
1563/1563 [==============================] - 11s 7ms/step - loss: 1.3455 -
accuracy: 0.5212 - val_loss: 1.4320 - val_accuracy: 0.4926
Epoch 12/15
1563/1563 [==============================] - 9s 6ms/step - loss: 1.3239 -
accuracy: 0.5271 - val_loss: 1.4626 - val_accuracy: 0.4830
Epoch 13/15
1563/1563 [==============================] - 8s 5ms/step - loss: 1.3052 -
accuracy: 0.5326 - val_loss: 1.5024 - val_accuracy: 0.4751
Epoch 14/15
1563/1563 [==============================] - 7s 4ms/step - loss: 1.2937 -
accuracy: 0.5364 - val_loss: 1.4824 - val_accuracy: 0.4759
Epoch 15/15
1563/1563 [==============================] - 6s 4ms/step - loss: 1.2708 -
accuracy: 0.5444 - val_loss: 1.4649 - val_accuracy: 0.4905
[ ]: # Evaluate the model
score = model.evaluate(X_test, y_test)
313/313 [==============================] - 1s 3ms/step - loss: 1.4649 -
accuracy: 0.4905
3
[ ]: import matplotlib.pyplot as plt
plt.plot(history.history['accuracy'], label='Training Accuracy')
plt.plot(history.history['val_accuracy'], label='Validation Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()
plt.show()
2 4. Analyzing the impact of optimization and weight initializa-
tion techniques on neural networks
[ ]: import tensorflow as tf
import numpy as np
from keras import models,layers,optimizers
from keras.datasets import cifar10
from keras.utils import to_categorical
[ ]: (X_train,y_train), (X_test,y_test) = cifar10.load_data()
Downloading data from https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
4
170498071/170498071 [==============================] - 6s 0us/step
[ ]: X_train = X_train.astype('float32')/ 255.0
X_test = X_test.astype('float32')/255.0
[ ]: X_train.shape
[ ]: (50000, 32, 32, 3)
[ ]: y_train = to_categorical(y_train,10)
y_test = to_categorical(y_test,10)
[ ]: #Xavier Initialization
model1 = models.Sequential()
model1.add(layers.Flatten(input_shape=(32,32,3)))
model1.add(layers.
↪Dense(256,activation='relu',kernel_initializer='glorot_uniform'))
model1.add(layers.
↪Dense(256,activation='relu',kernel_initializer='glorot_uniform'))
model1.add(layers.
↪Dense(10,activation='softmax',kernel_initializer='glorot_uniform'))
[ ]: #Kaiming Initialization
model2 = models.Sequential()
model2.add(layers.Flatten(input_shape=(32,32,3)))
model2.add(layers.Dense(256,activation='relu',kernel_initializer='he_normal'))
model2.add(layers.Dense(128,activation='relu',kernel_initializer='he_normal'))
model2.add(layers.Dense(10,activation='softmax',kernel_initializer='he_normal'))
[ ]: #With dropout Layer
model3 = models.Sequential()
model3.add(layers.Flatten(input_shape=(32,32,3)))
model3.add(layers.
↪Dense(256,activation='relu',kernel_initializer='glorot_uniform'))
model3.add(layers.Dropout(0.25))
model3.add(layers.Dense(128,activation='relu'))
model3.add(layers.Dense(10,activation='softmax'))
5
[ ]: # with batch normalization
model4 = models.Sequential()
model4.add(layers.Flatten(input_shape=(32,32,3)))
model4.add(layers.Dense(256,activation='relu'))
model4.add(layers.BatchNormalization())
model4.add(layers.Activation('relu'))
model4.add(layers.Dense(10,activation='softmax'))
[ ]: sgd_optimizer = optimizers.SGD(learning_rate=0.01, momentum=0.9)
model1.compile(optimizer=sgd_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model1.summary())
Xavier_history = model1.
↪fit(X_train,y_train,epochs=15,batch_size=32,validation_split=0.2)
Xavier_score = model1.evaluate(X_test,y_test,batch_size=32)
print(Xavier_score)
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 3072) 0
dense (Dense) (None, 256) 786688
dense_1 (Dense) (None, 256) 65792
dense_2 (Dense) (None, 10) 2570
=================================================================
Total params: 855050 (3.26 MB)
Trainable params: 855050 (3.26 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
None
Epoch 1/15
1250/1250 [==============================] - 11s 5ms/step - loss: 1.8755 -
accuracy: 0.3203 - val_loss: 1.8879 - val_accuracy: 0.3179
Epoch 2/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.7116 -
accuracy: 0.3821 - val_loss: 1.7003 - val_accuracy: 0.3947
Epoch 3/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6399 -
accuracy: 0.4110 - val_loss: 1.6479 - val_accuracy: 0.4036
Epoch 4/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.5987 -
6
accuracy: 0.4257 - val_loss: 1.6595 - val_accuracy: 0.4075
Epoch 5/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.5721 -
accuracy: 0.4345 - val_loss: 1.6283 - val_accuracy: 0.4223
Epoch 6/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.5427 -
accuracy: 0.4479 - val_loss: 1.6270 - val_accuracy: 0.4249
Epoch 7/15
1250/1250 [==============================] - 8s 6ms/step - loss: 1.5200 -
accuracy: 0.4543 - val_loss: 1.6115 - val_accuracy: 0.4251
Epoch 8/15
1250/1250 [==============================] - 8s 6ms/step - loss: 1.5047 -
accuracy: 0.4614 - val_loss: 1.5817 - val_accuracy: 0.4326
Epoch 9/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.4787 -
accuracy: 0.4676 - val_loss: 1.6318 - val_accuracy: 0.4289
Epoch 10/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.4709 -
accuracy: 0.4707 - val_loss: 1.5859 - val_accuracy: 0.4364
Epoch 11/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.4467 -
accuracy: 0.4812 - val_loss: 1.6028 - val_accuracy: 0.4405
Epoch 12/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.4289 -
accuracy: 0.4861 - val_loss: 1.5789 - val_accuracy: 0.4499
Epoch 13/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.4155 -
accuracy: 0.4927 - val_loss: 1.5831 - val_accuracy: 0.4351
Epoch 14/15
1250/1250 [==============================] - 4s 4ms/step - loss: 1.4029 -
accuracy: 0.4967 - val_loss: 1.5780 - val_accuracy: 0.4483
Epoch 15/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.3904 -
accuracy: 0.5012 - val_loss: 1.5149 - val_accuracy: 0.4675
313/313 [==============================] - 1s 3ms/step - loss: 1.5040 -
accuracy: 0.4652
[1.5040026903152466, 0.4652000069618225]
[ ]: sgd_optimizer = optimizers.SGD(learning_rate=0.01, momentum=0.9)
model2.compile(optimizer=sgd_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model2.summary())
Kaiming_history = model2.
↪fit(X_train,y_train,epochs=15,batch_size=32,validation_split=0.2)
Kaiming_score = model2.evaluate(X_test,y_test,batch_size=128)
print(Kaiming_score)
7
Model: "sequential_1"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten_1 (Flatten) (None, 3072) 0
dense_3 (Dense) (None, 256) 786688
dense_4 (Dense) (None, 128) 32896
dense_5 (Dense) (None, 10) 1290
=================================================================
Total params: 820874 (3.13 MB)
Trainable params: 820874 (3.13 MB)
Non-trainable params: 0 (0.00 Byte)
_________________________________________________________________
None
Epoch 1/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.8932 -
accuracy: 0.3129 - val_loss: 1.8377 - val_accuracy: 0.3323
Epoch 2/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7308 -
accuracy: 0.3795 - val_loss: 1.7861 - val_accuracy: 0.3530
Epoch 3/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6724 -
accuracy: 0.4001 - val_loss: 1.6722 - val_accuracy: 0.3971
Epoch 4/15
1250/1250 [==============================] - 7s 5ms/step - loss: 1.6274 -
accuracy: 0.4169 - val_loss: 1.6683 - val_accuracy: 0.4111
Epoch 5/15
1250/1250 [==============================] - 6s 4ms/step - loss: 1.5974 -
accuracy: 0.4286 - val_loss: 1.6234 - val_accuracy: 0.4245
Epoch 6/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.5605 -
accuracy: 0.4411 - val_loss: 1.6466 - val_accuracy: 0.4192
Epoch 7/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.5411 -
accuracy: 0.4503 - val_loss: 1.5860 - val_accuracy: 0.4396
Epoch 8/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.5239 -
accuracy: 0.4536 - val_loss: 1.7091 - val_accuracy: 0.4008
Epoch 9/15
1250/1250 [==============================] - 7s 5ms/step - loss: 1.5152 -
accuracy: 0.4567 - val_loss: 1.5768 - val_accuracy: 0.4461
Epoch 10/15
1250/1250 [==============================] - 10s 8ms/step - loss: 1.4989 -
accuracy: 0.4622 - val_loss: 1.5809 - val_accuracy: 0.4467
8
Epoch 11/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.4892 -
accuracy: 0.4690 - val_loss: 1.5822 - val_accuracy: 0.4462
Epoch 12/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.4619 -
accuracy: 0.4761 - val_loss: 1.5610 - val_accuracy: 0.4498
Epoch 13/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.4540 -
accuracy: 0.4814 - val_loss: 1.5693 - val_accuracy: 0.4469
Epoch 14/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.4407 -
accuracy: 0.4827 - val_loss: 1.5491 - val_accuracy: 0.4594
Epoch 15/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.4353 -
accuracy: 0.4843 - val_loss: 1.5748 - val_accuracy: 0.4488
79/79 [==============================] - 0s 3ms/step - loss: 1.5467 - accuracy:
0.4579
[1.54667329788208, 0.4578999876976013]
[ ]: sgd_optimizer = optimizers.SGD(learning_rate=0.01, momentum=0.9)
model3.compile(optimizer=sgd_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model3.summary())
dropout_history = model3.
↪fit(X_train,y_train,epochs=15,batch_size=32,validation_split=0.2)
dropout_score = model3.evaluate(X_test,y_test,batch_size=128)
print(dropout_score)
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten_2 (Flatten) (None, 3072) 0
dense_6 (Dense) (None, 256) 786688
dropout (Dropout) (None, 256) 0
dense_7 (Dense) (None, 128) 32896
dense_8 (Dense) (None, 10) 1290
=================================================================
Total params: 820874 (3.13 MB)
Trainable params: 820874 (3.13 MB)
Non-trainable params: 0 (0.00 Byte)
9
_________________________________________________________________
None
Epoch 1/15
1250/1250 [==============================] - 9s 5ms/step - loss: 2.0002 -
accuracy: 0.2613 - val_loss: 1.8789 - val_accuracy: 0.3180
Epoch 2/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.8879 -
accuracy: 0.3088 - val_loss: 1.8183 - val_accuracy: 0.3398
Epoch 3/15
1250/1250 [==============================] - 7s 6ms/step - loss: 1.8400 -
accuracy: 0.3270 - val_loss: 1.7877 - val_accuracy: 0.3499
Epoch 4/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7914 -
accuracy: 0.3453 - val_loss: 1.7410 - val_accuracy: 0.3655
Epoch 5/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7679 -
accuracy: 0.3561 - val_loss: 1.7466 - val_accuracy: 0.3774
Epoch 6/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7416 -
accuracy: 0.3689 - val_loss: 1.6831 - val_accuracy: 0.3946
Epoch 7/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7284 -
accuracy: 0.3720 - val_loss: 1.6930 - val_accuracy: 0.3958
Epoch 8/15
1250/1250 [==============================] - 6s 5ms/step - loss: 1.7120 -
accuracy: 0.3805 - val_loss: 1.6957 - val_accuracy: 0.3885
Epoch 9/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.7019 -
accuracy: 0.3849 - val_loss: 1.6806 - val_accuracy: 0.4138
Epoch 10/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6828 -
accuracy: 0.3924 - val_loss: 1.6229 - val_accuracy: 0.4289
Epoch 11/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6775 -
accuracy: 0.3933 - val_loss: 1.6534 - val_accuracy: 0.4150
Epoch 12/15
1250/1250 [==============================] - 6s 4ms/step - loss: 1.6607 -
accuracy: 0.3996 - val_loss: 1.6522 - val_accuracy: 0.4134
Epoch 13/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6433 -
accuracy: 0.4073 - val_loss: 1.6365 - val_accuracy: 0.4357
Epoch 14/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6384 -
accuracy: 0.4091 - val_loss: 1.6059 - val_accuracy: 0.4261
Epoch 15/15
1250/1250 [==============================] - 5s 4ms/step - loss: 1.6356 -
accuracy: 0.4095 - val_loss: 1.6004 - val_accuracy: 0.4362
79/79 [==============================] - 0s 5ms/step - loss: 1.5753 - accuracy:
10
0.4425
[1.575345754623413, 0.4424999952316284]
[ ]: sgd_optimizer = optimizers.SGD(learning_rate=0.01, momentum=0.9)
model4.compile(optimizer=sgd_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
print(model4.summary())
BN_history = model4.
↪fit(X_train,y_train,epochs=15,batch_size=128,validation_split=0.2)
BN_score = model4.evaluate(X_test,y_test,batch_size=128)
print(BN_score)
Model: "sequential_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten_3 (Flatten) (None, 3072) 0
dense_9 (Dense) (None, 256) 786688
batch_normalization (Batch (None, 256) 1024
Normalization)
activation (Activation) (None, 256) 0
dense_10 (Dense) (None, 10) 2570
=================================================================
Total params: 790282 (3.01 MB)
Trainable params: 789770 (3.01 MB)
Non-trainable params: 512 (2.00 KB)
_________________________________________________________________
None
Epoch 1/15
313/313 [==============================] - 4s 7ms/step - loss: 1.7435 -
accuracy: 0.3929 - val_loss: 1.9077 - val_accuracy: 0.3212
Epoch 2/15
313/313 [==============================] - 3s 9ms/step - loss: 1.5787 -
accuracy: 0.4527 - val_loss: 2.0264 - val_accuracy: 0.3129
Epoch 3/15
313/313 [==============================] - 2s 8ms/step - loss: 1.4976 -
accuracy: 0.4791 - val_loss: 1.7964 - val_accuracy: 0.3873
Epoch 4/15
313/313 [==============================] - 2s 5ms/step - loss: 1.4604 -
accuracy: 0.4917 - val_loss: 1.7981 - val_accuracy: 0.3686
Epoch 5/15
11
313/313 [==============================] - 2s 5ms/step - loss: 1.4433 -
accuracy: 0.4959 - val_loss: 1.6776 - val_accuracy: 0.4188
Epoch 6/15
313/313 [==============================] - 2s 5ms/step - loss: 1.4147 -
accuracy: 0.5046 - val_loss: 1.6728 - val_accuracy: 0.4098
Epoch 7/15
313/313 [==============================] - 2s 8ms/step - loss: 1.3822 -
accuracy: 0.5164 - val_loss: 1.6894 - val_accuracy: 0.4140
Epoch 8/15
313/313 [==============================] - 2s 8ms/step - loss: 1.3502 -
accuracy: 0.5282 - val_loss: 1.6909 - val_accuracy: 0.4284
Epoch 9/15
313/313 [==============================] - 4s 11ms/step - loss: 1.3272 -
accuracy: 0.5384 - val_loss: 1.6696 - val_accuracy: 0.4245
Epoch 10/15
313/313 [==============================] - 2s 6ms/step - loss: 1.3183 -
accuracy: 0.5387 - val_loss: 1.7419 - val_accuracy: 0.4188
Epoch 11/15
313/313 [==============================] - 2s 5ms/step - loss: 1.2873 -
accuracy: 0.5481 - val_loss: 1.6096 - val_accuracy: 0.4474
Epoch 12/15
313/313 [==============================] - 2s 5ms/step - loss: 1.2667 -
accuracy: 0.5526 - val_loss: 1.8200 - val_accuracy: 0.4028
Epoch 13/15
313/313 [==============================] - 2s 5ms/step - loss: 1.2532 -
accuracy: 0.5611 - val_loss: 2.0636 - val_accuracy: 0.3672
Epoch 14/15
313/313 [==============================] - 2s 5ms/step - loss: 1.2398 -
accuracy: 0.5660 - val_loss: 1.7374 - val_accuracy: 0.4327
Epoch 15/15
313/313 [==============================] - 2s 6ms/step - loss: 1.2253 -
accuracy: 0.5719 - val_loss: 1.8646 - val_accuracy: 0.3977
79/79 [==============================] - 0s 4ms/step - loss: 1.8478 - accuracy:
0.4013
[1.8477566242218018, 0.40130001306533813]
[ ]: import matplotlib.pyplot as plt
plt.plot(Xavier_history.history['val_accuracy'],label='Xavier Initialization')
plt.plot(Kaiming_history.history['val_accuracy'],label='Kaiming Initialization')
plt.plot(dropout_history.history['val_accuracy'],label='Dropout')
plt.plot(BN_history.history['val_accuracy'],label='Batch_Normalization')
plt.legend()
plt.show()
12
