I-Keras iyilabhulali yenethiwekhi ye-neural evulekile ebhalwe nge-Python. Iyakwazi ukusebenza phezu kwe-TensorFlow, i-Microsoft Cognitive Toolkit, i-R, i-Theano, noma i-PlaidML. Enye yezinzuzo ezibalulekile ze-Keras ukuvumela abathuthukisi ukuthi bathumele amamodeli ngezikhathi ezithile, okungaba yinzuzo enkulu ekuhleleni amamodeli nasekuhloleni ukusebenza kwawo.
Inkinga
Lapho siqeqesha imodeli yokufunda yomshini, ngokuvamile siqapha ukulahleka noma ukusebenza komsebenzi we-metric kwedatha ethile yokuqinisekisa. Izinkathi ezihlukene zingaholela ekusebenzeni kwemodeli okuhlukile. Kwesinye isikhathi, imiphumela ehamba phambili itholakala esikhathini esingahambelani nesiphetho senqubo yokuqeqesha. Ezimweni ezinjalo, kungaba usizo uma singagcina/sithekelisa imodeli ye-keras ngezikhathi ezithile.
Isixazululo
Isixazululo sale nkinga siku-Keras Callbacks. I-Callback yinto (isibonelo sekilasi esebenzisa izindlela ezithile) edluliselwa kumodeli ocingweni ukuze ilingane futhi ebizwa imodeli ezindaweni ezihlukahlukene phakathi nokuqeqeshwa. Inokufinyelela kuyo yonke idatha etholakalayo mayelana nesimo semodeli nokusebenza kwayo.
Umsebenzi wokuphinda ushaye ngokwezifiso usivumela ukuthi sicacise izenzo ezigabeni ezihlukene zokuqeqeshwa, njengasekuqaleni noma ekupheleni kwenkathi, ngaphambi noma ngemva kwenqwaba eyodwa, njll. Esinye sezenzo ezinjalo kungaba ukulondoloza imodeli ngezikhathi ezithile.
Incazelo yesinyathelo ngesinyathelo yeKhodi
Okokuqala, sichaza ukubuyisela emuva ngokwezifiso ukusindisa imodeli kuma-epoch athile.
class CustomSaver(keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
if epoch == 9: # or save after some epoch, each k-th epoch etc.
self.model.save("model_{}.hd5".format(epoch))
Okulandelayo, sengeza le callback kunqubo yokufaka imodeli.
model = ... # create model model.compile(optimizer='...', loss='...') # compile model saver = CustomSaver() model.fit(..., callbacks=[saver]) # put your X_train, Y_train ...
Ekhodini engenhla, sidala isibonelo sekilasi le-CustomSaver, bese silidlulisela endleleni yokulingana yemodeli njengengxenye yohlu lokufona emuva.
Ngokulungisa isimo esithi “uma” ngaphakathi kwendlela yethu ye-'on_epoch_end', singenza ukuphinda ushayele kongiwe ngemva kwenkathi ngayinye ethi 'k', noma noma nini lapho umbandela othile ufinyelelwa.
Imitapo yolwazi yaseKeras kanye Nemisebenzi Yayo
Idizayini kaKeras ethambile, esebenziseka kalula yenza kube lula ukudalwa kwemodeli yokufunda okujulile nokulungiswa. Iza namathuluzi amaningana, njengezigaba Ezilandelanayo kanye Nezimodeli zamamodeli wokwakha, izendlalelo ezihlukahlukene zamanethiwekhi emizwa (Convolutional, Pooling, Dense, njll.), kanye nama-callbacks okuqeqeshwa kokuqapha.
Esinye sezici eziwusizo kakhulu zokufonelwa emuva ukumisa ukuqeqeshwa kusenesikhathi, noma ukulondoloza imodeli ehamba phambili ngokuya ngokusebenza kokuqinisekisa. Kuyisici esivame ukubuzwa ukuthi i-Keras isivele iqukethe ama-callbacks akhelwe ngaphakathi kubo, okwaziwa nge-ModelCheckpoint kanye ne-EarlyStopping.
Yilokho kuphela! Manje sebenzisa lolu lwazi esimweni sakho futhi ulondoloze imodeli ye-Keras ngesikhathi lapho inikeza ukusebenza okuhle kakhulu. Ukuqeqeshwa okujabulisayo okuyimodeli!
