v_cnn.py.txt

# V-CNN Versatile CNN model - 
# includes L-CNN, NL-CNN and XNL-CNN as particular cases. 
# The basic unit is the "macro-layer" as in the XNL-CNN but here one can independently choose the 
# filter size (fil) and 
# nonlinearity nl (0 means "linear" convolution) 
# It allows any number of additional dense layers e.g. hid=[] (no hidden dense) or hid =[100, 100] (two additional). 
# Copyright Radu and Ioana DOGARU - correspondence: radu.dogaru@upb.ro 
# Last update June 20, 2023 

======================================
from tensorflow.keras.models import Sequential

from tensorflow.keras.layers import Dense, Dropout, Flatten, Activation, BatchNormalization
from tensorflow.keras.layers import Conv2D, DepthwiseConv2D, MaxPooling2D, AveragePooling2D, GlobalAveragePooling2D, SeparableConv2D  # straturi convolutionale si max-pooling 
from tensorflow.keras.optimizers import RMSprop, SGD, Adadelta, Adam, Nadam

#-------------------------- num_classes and IMAGE_SIZE should be adapted to ypur dataset 
# optimizers, losses as well 
#------------------------------ code 

#num_classes=104; 
#input_shape=[*IMAGE_SIZE, 3]; 
#----------------------------------------------------------------

# -- in a future version to be included as arguments of the model function 

flat=1  # Global Avg (flat=0) vs. Flatten (flat=1)... 
fil=[80, 100, 110, 120, 130]  # number of filters per each macrolayer , here 5 macro-layers 
nl=[1,1,0,1,1] # nonlin factors (should have same size as fil )
hid=[]  # Hidden layers in the end (empty)
#----------------------  code 

csize=3; stri=2; psiz=4; pad='same';
drop1=0.5    
nfilmax=np.shape(np.array(fil))[0]
model = Sequential()
    # First macrolayer - connected to input  ----------------
layer=0
if nl[layer]>0:
        model.add(Conv2D(fil[layer], padding=pad, kernel_size=(csize, csize), input_shape=input_shape ) )
        model.add(Activation('relu'))
        for nonlin in range(1,nl[0]):
            model.add(Conv2D(fil[layer], padding=pad, kernel_size=(csize, csize)) )
            model.add(Activation('relu'))
        
        model.add(Conv2D(fil[0], padding=pad, kernel_size=(csize, csize) ) ) # Activ NL-CNN-2
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size=(psiz, psiz),strides=(stri,stri),padding=pad))
        #model.add(Dropout(drop1))
  
else:
        model.add(Conv2D(fil[0], padding=pad, kernel_size=(csize, csize), input_shape=input_shape ) ) # Activ NL-CNN-2
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size=(psiz, psiz),strides=(stri,stri),padding=pad))
        #model.add(Dropout(drop1))
    # Other macro-layers 
    
for layer in range(1,nfilmax):
        #------------------ nonlin layers -----------------
        for nonlin in range(nl[layer]):
            model.add(Conv2D(fil[layer], padding=pad, kernel_size=(csize, csize)) )
            model.add(Activation('relu'))
            
        #----------------- default macrolayer output 
        
        model.add(Conv2D(fil[layer], padding=pad, kernel_size=(csize, csize)) )
        model.add(BatchNormalization())
        model.add(MaxPooling2D(pool_size=(psiz, psiz),strides=(stri,stri),padding=pad))
        #model.add(Dropout(drop1))
        
    # Exit classifier 
    # INPUT TO DENSE LAYER (FLATTEN - more data can overfit / GLOBAL - less data - may be a good choice ) 
if flat==1:
    model.add(Flatten())  # alternanta cu GlobalAv ..
elif flat==0:
    model.add(GlobalAveragePooling2D()) # pare sa fie mai Ok la cifar         
nhid=np.shape(np.array(hid))[0]
if nhid>0:
        for lay in range(nhid):
            model.add(Dense(hid[lay], activation='relu'))
            model.add(Dropout(drop1))
model.add(Dense(num_classes, activation='softmax'))

# END OF MODEL DESCRIPTION 
model.compile(
    optimizer=tf.keras.optimizers.Adam(),
    loss = 'categorical_crossentropy',
    metrics=['accuracy']
)

model.summary()