Ukusebenza kwanoma iyiphi imodeli yokufunda yomshini kungase kubalulwe ngokuyinhloko ekuqinisekiseni kwedatha esetshenziselwa ukuqeqeshwa kanye nekhono lemodeli lokukhiqiza kahle kusuka kulolu lwazi. Ngakho-ke, ukuqeqesha imodeli enokulahlekelwa okuphansi kokuqinisekisa kubalulekile. Ngokuqeqesha imodeli ngokuphumelelayo, siqinisekisa ukuthi imodeli ayifaneleki futhi ayifaki ngokweqile.
Ukulahlekelwa kokuqinisekisa kuhambisana nezinga lephutha kusampula yokubamba isethi yokuqeqeshwa engasetshenziswa esigabeni sokuqeqesha futhi kusisiza silinganise amandla emodeli okwenza okuvamile. Umgomo oyinhloko kunoma iyiphi imodeli yokufunda yomshini uwukuthola ukulahlekelwa kokuqinisekisa okuphansi kakhulu, okubonisa ukuthi imodeli yethu ifunda futhi yenza okuvamile kahle.
Indlela Yokulondoloza Amamodeli Ngokulahlekelwa Okungcono Kakhulu Kokuqinisekisa
Ukuqeqesha imodeli kuhilela ukuphindaphinda okuningana, okubuye kwaziwe ngokuthi ama-epoch, futhi ukulahlekelwa kokuqinisekisa kuyehluka enkathini ngayinye. I-Python inikeza imitapo yolwazi eminingana efana ne-Keras egcina lawa mamodeli ngesikhathi ngasinye. Singasebenzisa isici esibizwa nge-ModelCheckpoint ukuze silondoloze imodeli noma nini lapho ukulahlekelwa kokuqinisekisa kuba ngcono kunangaphambili.
from keras.callbacks import ModelCheckpoint # specify the path to save the model filepath="weights.best.hdf5" # initiate the ModelCheckpoint function checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min') # define the list of callbacks callbacks_list = [checkpoint] # fit the model model.fit(X, Y, validation_split=0.33, epochs=150, batch_size=10, callbacks=callbacks_list, verbose=0)
Ukuqonda Ikhodi: Isinyathelo Ngesinyathelo
Ake sidlule kumazwibela ekhodi isinyathelo ngesinyathelo ukuze siqonde ingxenye ngayinye:
1. Ngenisa umsebenzi we-ModelCheckpoint kusuka ku-Keras.
2. Chaza indlela yefayela lapho ufuna ukulondoloza khona imodeli usebenzisa ifomethi ye-.hdf5. Le fomethi yakhelwe ukugcina nokuhlela amanani amakhulu edatha.
3. Qalisa umsebenzi we-ModelCheckpoint. Lapha, siqapha 'val_loss' ngemodi ethi 'min' ebonisa ukuthi sihlose ukunciphisa leli nani. Nge-'save_best_only=True', imodeli yakamuva ehamba phambili ngokwenani eligadiwe ngeke ibhalwe phezu.
4. Indawo yokuhlola ibe isifakwa ohlwini lwabafonayo. Amanye amapharamitha wokuqeqesha afana ne-EarlyStopping nawo angafakwa kulolu hlu.
5. Imodeli ibe isiqeqeshwa kudatha isebenzisa imodeli.fit(). I-agumenti yama-callbacks ingena ku-callbacks_list.
Ikhodi ilondoloza imodeli njenge-'weights.best.hdf5' enkathini ngayinye lapho ukulahlekelwa kokuqinisekisa kusezingeni eliphansi.
Imitapo yolwazi ye-Python Eyinhloko Yokulondoloza Amamodeli
I-Python inikeza uhlelo olucebile lwemitapo yolwazi yokonga amamodeli okufunda komshini. Esetshenziswa kakhulu yile:
- I-Keras: Leveli ephezulu ye-neural networks API ekwazi ukusebenza phezu kwamanye ama-API ezingeni eliphansi njenge-TensorFlow ivumela i-prototyping elula nesheshayo yamamodeli okufunda ajulile. Umsebenzi we-ModelCheckpoint ku-Keras unikeza ukuguquguquka kokuqapha amapharamitha ahlukahlukene phakathi nenqubo yokuqeqesha, nokulondoloza imodeli noma izisindo ngezigaba ezihlukahlukene.
- I-TensorFlow: Uhlaka lokufunda lomshini ovulekile lwe-Python luvumela abathuthukisi ukuthi bakhe amamodeli e-ML ayinkimbinkimbi kalula. Ihlinzeka ngemojula ye-SavedModel okuyifomethi yokuhlanganisa yonke indawo yamamodeli e-TensorFlow.
- I-Scikit-learn: Lo mtapo wezincwadi odumile wePython wokufunda ngomshini uhlinzeka ngezinsiza zokulondoloza nokulayisha amamodeli. Imojula ye-Joblib ivamise ukusetshenziselwa ukukhiqiza izinto ze-Python nge-serializing ezinkulu eziyi-numpy - isimo esivamile ekugelezeni komsebenzi wokufunda komshini.
Ukukhetha umtapo wolwazi kuncike kakhulu ezidingweni zephrojekthi yakho, ukujwayelana kwakho nomtapo wolwazi, kanye nokuba yinkimbinkimbi kwemodeli yakho. Noma kunjalo, iPython inikeza izinsiza ezanele zokuqeqesha, ukuhlola, ukonga nokulayisha amamodeli kalula.
