Sida Loo Sameeyo Farsamooyinka Muuqaalka (Fine-tuning) Waxtar Leh - Hage Bilowga

In codsiyada casriga ah ee barashada mashiinka iyo sirdoonka macmalka ah, farsamooyinka muuqaalka (Fine-tuning) oo ah farsamo muhiim ah oo lagu hagaajinayo moodalka si uu ugu habboonaado hawl gaar ah, ayaa si ballaaran looga hadlaa oo loo isticmaalaa. Hagegan wuxuu ujeedkiisu yahay inuu caawiyo bilowga in uu fahmo fikradaha aasaasiga ah ee farsamooyinka muuqaalka, xaaladaha codsiga iyo tillaabooyinka gaarka ah ee la fulinayo. Haddii aad rabto inaad kor u qaaddo saxnaanta moodalka barashada mashiinka, ama aad rabto inaad isticmaasho moodal hore loo tababaray mashruucaaga, barashada xirfadaha farsamooyinka muuqaalka waa mid aad muhiim u ah.

Waa maxay farsamooyinka muuqaalka?

Farsamooyinka muuqaalka waxay ka dhigan tahay in la tababaro moodal hore loo tababaray iyadoo la adeegsanayo xog cusub si loo hagaajiyo parameterrada moodalka si uu si fiican ugu habboonaado hawl gaar ah. Caadi ahaan, waxaanu isticmaali doonaa moodal hore loo tababaray oo ku saleysan xog ballaaran, ka dibna waxaanu isticmaalnaa xog yar oo gaar ah si aan u kordhino waxqabadka.

Faa'iidooyinka farsamooyinka muuqaalka:

Badbaadinta waqti iyo kheyraadka xisaabinta: Marka loo eego in laga bilaabo tababarka moodalka, farsamooyinka muuqaalka badanaa waxay u baahan yihiin kheyraadka xisaabinta iyo waqti ka yar.
Kordhinta waxqabadka moodalka: Iyadoo la adeegsanayo xog gaar ah, moodalka wuxuu heli karaa saxnaanta sare.
U habboonaanta hawlo kala duwan: Moodal aasaasi ah ayaa lagu hagaajin karaa si loogu habboonaado meelo ama hawlo kala duwan.

Xaaladaha Codsiga Farsamooyinka Muuqaalka

Hagaajinta Luqadda Dabiiciga ah (NLP): Isticmaalka moodallada luqadda hore loo tababaray (sida BERT, GPT) si loo sameeyo falanqaynta dareenka, nidaamyada su'aalaha iyo jawaabaha iwm.
Aragtida Kombiyuutarka: Hawlaha kala soocida sawirada, ogaanshaha walxaha iwm, isticmaalka shabakadaha neerfaha ee convolutional hore loo tababaray (sida ResNet, Inception) si loo sameeyo farsamooyinka muuqaalka.
Nidaamyada Talooyinka: Hagaajinta algorithms talo bixinta ee jira si loogu habboonaado kooxo isticmaale gaar ah ama noocyada alaabta.

Tallaabooyinka Gaarka ah ee Farsamooyinka Muuqaalka

1. Dooro moodal hore loo tababaray oo ku habboon

Doorashada moodal hore loo tababaray oo ku habboon ayaa ah tallaabada ugu horreysa ee farsamooyinka muuqaalka. Tusaale ahaan, hawlaha sawirada waxaad dooran kartaa ResNet, hawlaha qoraalka waxaad dooran kartaa BERT.

from transformers import BertTokenizer, BertForSequenceClassification
model_name = 'bert-base-uncased'
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForSequenceClassification.from_pretrained(model_name, num_labels=2)

2. Diyaarinta xogta

Farsamooyinka muuqaalka waxay u baahan yihiin xog gaar ah oo la calaamadeeyay. Xogtan waxay ka koobnaan doontaa tusaalooyinka gelinta hawsha la beegsanayo iyo calaamadaha u dhigma.

import pandas as pd
# Akhri xogta
 data = pd.read_csv('data.csv')
texts = data['text'].tolist()
labels = data['label'].tolist()

3. Diyaarinta xogta

Ka hor inta aan la sameynin farsamooyinka muuqaalka, caadi ahaan waxaa lagama maarmaan ah in la diyaariyo xogta qoraalka, oo ay ku jiraan kala soocidda, koodhinta iwm.

# Kala soocidda iyo koodhinta xogta
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")

4. Dejinta parameterrada tababarka

Dejinta parameterrada tababarka inta lagu jiro habka farsamooyinka muuqaalka, oo ay ku jiraan heerka barashada, cabirka xirmada, muddada tababarka iwm.

from transformers import Trainer, TrainingArguments

training_args = TrainingArguments(
    output_dir='./results',
    num_train_epochs=3,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=64,
    evaluation_strategy="epoch",
    logging_dir='./logs',
)

5. Abuur Trainer

Isticmaalka Trainer si loo tababaro loona qiimeeyo moodalka.

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
)

trainer.train()

6. Qiimeynta Moodalka

Ka dib marka la dhammeeyo farsamooyinka muuqaalka, waxaa lagama maarmaan ah in la qiimeeyo waxqabadka moodalka ee xogta xaqiijinta ama tijaabada, si loo helo saxnaanta, dib u soo celinta iwm.

metrics = trainer.evaluate()
print(metrics)

7. Kaydinta iyo Daabacaadda Moodalka

Ka dib marka la dhammeeyo farsamooyinka muuqaalka, waxaad kaydin kartaa moodalka si aad u isticmaasho mustaqbalka, adigoo dooranaya habka daabacaadda ee ku habboon.

model.save_pretrained('./fine-tuned-model')
tokenizer.save_pretrained('./fine-tuned-model')

Talooyin iyo Hababka Ugu Fiican

Dooro heerka barashada ku habboon: Waxaad isku dayi kartaa inaad isticmaasho jadwalka heerka barashada, si tartiib ah u yaree heerka barashada si aad u hesho natiijooyin wanaagsan.
La soco waxqabadka moodalka: Iyada oo la socota khasaaraha iyo saxnaanta waqtiga tababarka, si degdeg ah u hagaaji parameterrada.
Ka fogaanshaha overfitting: Isku day inaad isticmaasho istiraatiijiyadda joojinta hore (Early Stopping) si aad uga fogaato in moodalka uu ku dhaco xogta tababarka.
Kordhinta xogta: Marka tusaalooyinka ay yartahay, waxaad tixgelin kartaa inaad isticmaasho farsamooyinka kordhinta xogta si aad u kordhiso kala duwanaanshaha xogta.
Qiimeyn joogto ah: Inta lagu jiro farsamooyinka muuqaalka, qiimee waxqabadka moodalka si joogto ah, si aad u hubiso in moodalka uusan ka leexanayn yoolka.

Gunaanad

Farsamooyinka muuqaalka waa qayb aan laga maarmin oo ka mid ah hagaajinta moodalka barashada mashiinka, adigoo si dabacsan u dooranaya moodalka hore loo tababaray, parameterrada tababarka oo macquul ah iyo habka xogta oo waxtar leh, waxaad si weyn u kordhin kartaa waxqabadka moodalka hawl gaar ah. Iyada oo tignoolajiyada si joogto ah u horumareysa, farsamooyinka muuqaalka waxay noqon doonaan xirfad sii kordheysa, barashada xirfaddan waxay keeni doontaa qiimo weyn oo ku saabsan codsiyada AI.