optim = Padam(learning_rate=learning_rate, p=op['p'], beta1=op['b1'], beta2=op['b2'])
elif optimizer == 'adam':
optim = tf.train.AdamOptimizer(learning_rate=learning_rate, beta1=op['b1'], beta2=op['b2'])
elif optimizer == 'adamw':
# adamw = tf.contrib.opt.extend_with_decoupled_weight_decay(tf.train.AdamOptimizer)
optim = tf.contrib.opt.AdamWOptimizer(weight_decay=op['weight_decay'], learning_rate=learning_rate, beta1=op['b1'], beta2=op['b2'])
elif optimizer == 'amsgrad':
optim = AMSGrad(learning_rate=learning_rate, beta1=op['b1'], beta2=op['b2'])
elif optimizer == 'sgd':
optim = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=op['m'])
model.compile(optimizer=optim, loss='categorical_crossentropy', metrics=['accuracy', 'top_k_categorical_accuracy'], global_step=tf.train.get_global_step())
csv_logger = CSVLogger(logfile, append=True, separator=';')
history[optimizer] = model.fit_generator(datagen_train.flow(trainX, trainY, batch_size = batch_size), epochs = epochs,
validation_data = datagen_test.flow(testX, testY, batch_size = batch_size), verbose=1, callbacks = [csv_logger])
scores = model.evaluate_generator(datagen_test.flow(testX, testY, batch_size = batch_size), verbose=1)
print("Final test loss and accuracy:", scores)
# filepath = 'model_'+optimizer+'_' + dataset + '.h5'
save_model(save_model_filepath, model)
# f.close()
#train plot
plt.figure(1)
for optimizer in optim_array:
op = optim_params[optimizer]
train_loss = history[optimizer].history['loss']
epoch_count = range(1, len(train_loss) + 1)
plt.plot(epoch_count, train_loss, color=op['color'], linestyle=op['linestyle'])
model.compile(loss = custom_loss,
optimizer='adam',
metrics=[char_acc,image_acc])
datagen = ImageDataGenerator(width_shift_range=0.14,
height_shift_range=0.08,
fill_mode='constant',
zoom_range = 0.1,
rotation_range = 10,
#rescale =1./255
)
mcp_save = ModelCheckpoint(cf.CKP_PATH, save_best_only=True, monitor='val_loss', mode='min',verbose=1)
def scheduler(epoch):
if epoch <4 :
return 0.001
elif epoch < 10:
return 0.001/5
elif epoch < 15:
return 0.0001
elif epoch <30:
return 0.0001/2
n = train_x.shape[0]
#model.load_weights("plate.h5")
lr_reduce = LearningRateScheduler(scheduler,verbose = 1)
model.fit_generator(datagen.flow(train_x, train_y,batch_size=cf.BATCH_SIZE),
epochs = 10,
steps_per_epoch=n//cf.BATCH_SIZE,
callbacks=[lr_reduce,mcp_save],
validation_data=(val_x, val_y))
