def DiffSamp(samples):
for samples in samples:
model = get_model()
opt = RMSprop(lr=0.0001, decay=1e-6)
optName = "RMSProp"
initName = "dataInit"
lr = 0.0001
epochs = 50
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
model = jan_initialize(model, X_train,
samples) #check the working function :)
hist = model.fit(X_train,
Y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(X_test, Y_test))
#
HistoryPath = os.path.join(
save_dir, optName + "_" + initName + "_Hist_" +
"lr_{:1}_epoch_{:1d}_Jan28_".format(lr, epochs) + str(samples) +
"_Samples")
saveHistory(HistoryPath, hist)
return
def trainOptInit(X_train,
X_test,
Y_train,
Y_test,
initDict,
optDict={'RMSProp': RMSprop(lr=0.0001, decay=1e-6)},
mode='full',
saveHist=True):
if mode == 'test': #Just test over a few values
X_train = X_train[0:200]
Y_train = Y_train[0:200]
X_test = X_test[0:40]
Y_test = Y_test[0:40]
for optName, opt in optDict.items():
#---------------------------------------------- For Different Init and Optimizers paras
for key, initName in initDict.items():
print(" Training for {:s} initialization with {:s} as optimizer".
format(key, optName))
model = get_model(initName,
input_shape=X_train.shape[1:],
num_classes=num_classes)
# lr = 0.0001
# opt = RMSprop(lr=lr, decay=1e-6)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
# hist = model.fit(X_train,Y_train, batch_size=batch_size, epochs=epochs, validation_data=(X_test,Y_test),shuffle=True); #hist will store everything
#earlyStop = EarlyStopping(monitor='val_acc', min_delta=0, patience=10, verbose=0, mode='max')
hist = model.fit(X_train,
Y_train,
batch_size=batch_size,
epochs=epochs,
validation_data=(X_test, Y_test),
shuffle=True,
callbacks=[earlyStop])
#With Callback
if (saveHist == True):
#---------------------------------------------- Save History & Model
model.save(
os.path.join(
save_dir, optName + initName +
"_lr_{:1}_epoch_{:1d}_".format(lr, epochs) +
model_name))
HistoryPath = os.path.join(
save_dir, optName + "_" + initName + "_Hist_" +
"lr_{:1}_epoch_{:1d}_".format(lr, epochs) + model_name)
saveHistory(HistoryPath, hist)
return 1
base_dir = "/home/ankit/Desktop/Dataset/JanResults/CIFAR10"; #save results here...
data_folder = "/home/ankit/Desktop/Dataset/CIFAR10Dataset/cifar-10-batches-py/"; #Server
img_size = 32;
batch_size = 32;
num_classes = 10;
epochs = 10;
#------------------- Get the X_train and Model for temporary use
X_train,Y_train,X_test,Y_test = load_data_full(data_folder=data_folder, batch_no = 5,img_size=img_size)
model = get_model();
del Y_train, X_test,Y_test; #Freeing up the space
#---------------------------------------------------------------------------------------------------
def jan_initialize(model,X_train,samp_size = 8):
# samp_size = 8192;
totalFiles = list(np.arange(X_train.shape[0]))
random.shuffle(totalFiles);
layers = model.layers;
img_size = 32
batch_size = 32
num_classes = 10
epochs = 100
save_dir = os.path.join(base_dir, 'Jan18')
model_name = 'Jan18'
# history_name = 'Keras_CIFAR10_HeInit_v1'
X_train, Y_train, X_test, Y_test = load_data_full(data_folder=data_folder,
batch_no=5,
img_size=img_size)
# X_train = X_train[0:200];Y_train = Y_train[0:200];
# X_test = X_test[0:40];Y_test = Y_test[0:40];
model = get_model()
samples = 1024
opt = RMSprop(lr=0.0001, decay=1e-6)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=['accuracy'])
model = jan_initialize(model, X_train, samples) #check the working function :)
#
#
hist = model.fit(X_train,
Y_train,
batch_size=batch_size,
epochs=epochs,
img_size = 32
batch_size = 32
num_classes = 10
epochs = 150
save_dir = os.path.join(base_dir,'SavedModels/ProposedModel_try1')
model_name = 'Keras_CIFAR10_proposed_v1.h5'
history_name = 'Keras_CIFAR10_proposed_v1'
#load Train/Test
X_train,Y_train,X_test,Y_test = load_data_full(data_folder=data_folder, batch_no = 5,img_size=img_size)
#ModelKeys
initDict = {'LeCun':'lecun_uniform', 'GlorotNorm':'glorot_normal','GlorotUni':'glorot_uniform','HeNorm':'he_normal','HeUni':'he_uniform'}
model = get_model(initDict['HeNorm'],input_shape= X_train.shape[1:],num_classes = num_classes)
model = scaledInitialize(model= model, X_train=X_train) #Proposed initialization you want to initialize
initName = 'Proposed_try_1'
optName = 'RMSProp'
lr = 0.0001
opt = RMSprop(lr=lr, decay=1e-6)
model.compile(loss='categorical_crossentropy', optimizer=opt, metrics=['accuracy'])
hist = model.fit(X_train,Y_train, batch_size=batch_size, epochs=epochs, validation_data=(X_test,Y_test),shuffle=True); #hist will store everything
earlyStop = EarlyStopping(monitor='val_acc', min_delta=0, patience=10, verbose=0, mode='max')
# hist = model.fit(X_train, Y_train, batch_size=batch_size, epochs=epochs, validation_data=(X_test, Y_test), shuffle=True,callbacks=[earlyStop]); # With Callback
# ---------------------------------------------- Save History & Model
model.save(os.path.join(save_dir, optName + initName + "_lr_{:1}_epoch_{:1d}_".format(lr, epochs) + model_name))