📑 Exercises with Adversarial Regularization

🌐   DL_PP8_1_SMC.html

📑 Deep Learning. Practice Project 8_1:
Exercises with Adversarial Regularization

✒️ Code Modules, Helpful Functions, & Settings

xxxxxxxxxx

 

1

#!python3 -m pip install tensorflow==2.6.0 \

2

#--user --quiet --no-warn-script-location

3

!python3 -m pip install --upgrade neural_structured_learning \

4

--user --quiet --no-warn-script-location

5

path='/home/sc_work/.sage/local/lib/python3.9/site-packages'

6

import sys,warnings; sys.path.append(path)

7

warnings.filterwarnings('ignore')

8

import os,h5py,urllib,pandas as pd,numpy as np

9

import tensorflow as tf,pylab as pl,seaborn as sn

10

import neural_structured_learning as nsl

11

import tensorflow.keras.layers as tkl

12

import tensorflow.keras.callbacks as tkc

13

file_path='https://raw.githubusercontent.com/'+\

14

          'OlgaBelitskaya/data_kitchen/main/'

15

file_name='Pictograms64.h5'

16

model_weights='/tmp/checkpoint'

17

img_size=int(32)

18

from IPython.display import display,HTML

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

def pd_style():

2

    return [dict(selector='th',

3

                 props=[('font-size','12pt'),

4

                        ('min-width','150px')]),

5

            dict(selector='td',

6

                 props=[('padding','0em 0em'),

7

                        ('min-width','150px')]),

8

            dict(selector='tr:hover th:hover',

9

                 props=[('font-size','14pt'),

10

                        ('max-width','150px'),

11

                        ('text-shadow','3px 3px 3px #aaa')]),

12

            dict(selector='tr:hover td:hover',

13

                 props=[('font-size','12pt'),

14

                        ('max-width','150px'),

15

                        ('text-shadow','3px 3px 3px #aaa')])]

16

def display_images(images,labels,names,n):

17

    fig=pl.figure(figsize=(6,n/2))

18

    randch=np.random.choice(

19

        images.shape[0],size=n,replace=False)

20

    for i,idx in enumerate(randch):

21

        ax=fig.add_subplot(int(n//3)+1,3,i+1,

22

                           xticks=[],yticks=[])

23

        ax.imshow(images[idx],cmap='bone')

24

        label=[labels[:,idx]]

25

        name=[names[i][labels[i][idx]] 

26

                       for i in range(labels.shape[0])]

27

        ti='{} \n {}'.format(str(label),str(name))

28

        ax.set_title(ti,fontsize=8)

29

    pl.tight_layout(); pl.show()

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

def h5file2data(h5file,cmap,img_size,resize=True):

2

    with h5py.File(h5file,'r') as f:

3

        keys=list(f.keys())

4

        pretty_print('file keys: '+', '.join(keys))

5

        images=np.array(f[keys[int(0)]])

6

        labels=np.array(f[keys[int(1)]])

7

        names=[[el.decode('utf-8') for el in f[keys[i]]]

8

               for i in range(int(2),len(keys))]

9

        f.close()

10

    N=images.shape[int(0)]; n=int(.1*N)

11

    if resize:

12

        images=tf.image.resize(images,[img_size,img_size]).numpy()

13

    shuffle_ids=np.arange(N)

14

    np.random.RandomState(12).shuffle(shuffle_ids)

15

    images=images[shuffle_ids]

16

    labels=np.array([labels[i][shuffle_ids] 

17

                     for i in range(labels.shape[int(0)])])

18

    x_test,x_valid,x_train=\

19

    images[:n],images[n:2*n],images[2*n:]

20

    y_test,y_valid,y_train=\

21

    labels[:,:n],labels[:,n:2*n],labels[:,2*n:]

22

    pretty_print('data outputs: ')

23

    df=pd.DataFrame([[x_train.shape,x_valid.shape,x_test.shape],

24

                     [x_train.dtype,x_valid.dtype,x_test.dtype],

25

                     [y_train.shape,y_valid.shape,y_test.shape],

26

                     [y_train.dtype,y_valid.dtype,y_test.dtype]],

27

                    columns=['train','valid','test'],

28

                    index=['image shape','image type',

29

                           'label shape','label type'])

30

    display(df.style.set_table_styles(pd_style()))

31

    pretty_print('distribution of labels: ')

32

    idx=['labels %d'%(i+int(1)) for i in range(labels.shape[int(0)])]

33

    df=pd.DataFrame(labels,index=idx).T

34

    for i in range(labels.shape[int(0)]):

35

        df['name %d'%(i+int(1))]=[names[i][l] for l in labels[i]]

36

    fig=pl.figure(figsize=(6,6))    

37

    for i in range(labels.shape[int(0)]):

38

        ax=fig.add_subplot(labels.shape[int(0)],int(1),i+int(1))

39

        sn.countplot(y='name %s'%(i+1),data=df,

40

                     palette=cmap,alpha=.5,ax=ax)

41

    pl.tight_layout(); pl.show()       

42

    return [names,x_train,x_valid,x_test,y_train,y_valid,y_test]

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

def cb(mw):

2

    early_stopping=tf.keras.callbacks.EarlyStopping(

3

        monitor='val_sparse_categorical_accuracy',

4

        patience=int(20),verbose=int(2),mode='max')

5

    checkpointer=tf.keras.callbacks.ModelCheckpoint(

6

        save_best_only=True,save_weights_only=True,

7

        monitor='val_sparse_categorical_accuracy',

8

        filepath=mw,verbose=int(2),mode='max')

9

    lr_reduction=tf.keras.callbacks.ReduceLROnPlateau(

10

        monitor='val_sparse_categorical_accuracy',

11

        verbose=int(2),patience=int(5),factor=float(.9),mode='max')

12

    return [checkpointer,early_stopping,lr_reduction]

13

def keras_history_plot(

14

    fit_history,fig_size=6,color='#348ABD'):

15

    keys=list(fit_history.history.keys())

16

    list_history=[fit_history.history[keys[i]] 

17

                  for i in range(len(keys))]

18

    dfkeys=pd.DataFrame(list_history).T

19

    dfkeys.columns=keys

20

    fig=pl.figure(figsize=(fig_size,fig_size))

21

    ax1=fig.add_subplot(3,1,1)

22

    dfkeys.iloc[:,[0,4]].plot(

23

        ax=ax1,color=['slategray',color],grid=True)

24

    ax2=fig.add_subplot(3,1,2)

25

    dfkeys.iloc[:,[3,7]].plot(

26

        ax=ax2,color=['slategray',color],grid=True)

27

    ax3=fig.add_subplot(3,1,3)

28

    dfkeys.iloc[:,[2,6]].plot(

29

        ax=ax3,color=['slategray',color],grid=True)

30

    pl.tight_layout(); pl.show()

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

✒️ Data Loading & Processing

xxxxxxxxxx

 

1

input_file=urllib.request.urlopen(file_path+file_name)

2

output_file=open(file_name,'wb')

3

output_file.write(input_file.read())

4

output_file.close(); input_file.close()

5

[names,x_train,x_valid,x_test,y_train,y_valid,y_test]=\

6

h5file2data(file_name,'tab20',img_size)

7

num_classes=len(names[int(1)])

8

os.remove(file_name)

9

display_images(x_train,y_train,names,6)

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

batch_size=int(16)

2

train=tf.data.Dataset.from_tensor_slices(

3

    {'input':x_train,

4

     'label':np.array(y_train[1],dtype='float32')})\

5

     .batch(batch_size)

6

valid=tf.data.Dataset.from_tensor_slices(

7

    {'input':x_valid,

8

     'label':np.array(y_valid[1],dtype='float32')})\

9

     .batch(batch_size)

10

valid_steps=x_valid.shape[0]//batch_size

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

✒️ Models with Adversarial Regularization => CNN

xxxxxxxxxx

 

1

base_model=tf.keras.Sequential([

2

    tf.keras.Input((img_size,img_size,int(3)),name='input'),

3

    tf.keras.layers.Conv2D(int(32),(int(5),int(5)),padding='same'),

4

    tf.keras.layers.Activation('relu'),

5

    tf.keras.layers.MaxPooling2D(pool_size=(int(2),int(2))),

6

    tf.keras.layers.Dropout(float(.25)),

7

    tf.keras.layers.Conv2D(int(96),(int(5),int(5))),

8

    tf.keras.layers.Activation('relu'),    

9

    tf.keras.layers.MaxPooling2D(pool_size=(int(2),int(2))),

10

    tf.keras.layers.Dropout(float(.25)),

11

    tf.keras.layers.GlobalMaxPooling2D(),    

12

    tf.keras.layers.Dense(int(256)),

13

    tf.keras.layers.Activation('relu'),

14

    tf.keras.layers.Dropout(float(.5)),

15

    tf.keras.layers.Dense(num_classes,activation='softmax')

16

])

17

adv_config=nsl.configs\

18

.make_adv_reg_config(multiplier=float(.2),adv_step_size=float(.05))

19

adv_model=nsl.keras\

20

.AdversarialRegularization(base_model,adv_config=adv_config)

21

adv_model.compile(optimizer='adam',metrics=['accuracy'],

22

                  loss='sparse_categorical_crossentropy')

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

history=adv_model.fit(

2

    train,validation_data=valid,validation_steps=valid_steps,

3

    verbose=int(2),epochs=int(3),callbacks=cb(model_weights));

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

keras_history_plot(history)

2

adv_model.load_weights(model_weights)

3

dict(zip(adv_model.metrics_names,adv_model.evaluate(

4

    {'input':x_test,

5

     'label':np.array(y_test[int(1)],dtype='float32')},

6

     verbose=int(0))))

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages

xxxxxxxxxx

 

1

​

Language: Sage Gap GP HTML Macaulay2 Maxima Octave Python R Singular

---

Messages