Upload 10 files

distillation.py

@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sat Jun 17 16:20:22 2023
+
+@author: fujidai
+"""
+
+
+from sentence_transformers import SentenceTransformer, LoggingHandler, models, evaluation, losses
+import torch
+from torch.utils.data import DataLoader
+from sentence_transformers.datasets import ParallelSentencesDataset
+from datetime import datetime
+
+import os
+import logging
+import sentence_transformers.util
+import csv
+import gzip
+from tqdm.autonotebook import tqdm
+import numpy as np
+import zipfile
+import io
+
+logging.basicConfig(format='%(asctime)s - %(message)s',
+                    datefmt='%Y-%m-%d %H:%M:%S',
+                    level=logging.INFO,
+                    handlers=[LoggingHandler()])
+logger = logging.getLogger(__name__)
+
+
+teacher_model_name = '/Users/fujidai/sinTED/paraphrase-mpnet-base-v2'   #Our monolingual teacher model, we want to convert to multiple languages
+#teacher_model_name = '/Users/fujidai/TED2020_data/tisikizyouryu/bert-large-nli-mean-tokens'   #Our monolingual teacher model, we want to convert to multiple languages
+
+student_model_name = '/Users/fujidai/dataseigen/09-MarginMSELoss-finetuning-7-5'       #Multilingual base model we use to imitate the teacher model
+
+max_seq_length = 128                #Student model max. lengths for inputs (number of word pieces)
+train_batch_size = 128               #Batch size for training
+inference_batch_size = 128           #Batch size at inference
+max_sentences_per_language = 500000 #Maximum number of  parallel sentences for training
+train_max_sentence_length = 250     #Maximum length (characters) for parallel training sentences
+
+num_epochs = 3                       #Train for x epochs
+num_warmup_steps = 10000             #Warumup steps
+
+num_evaluation_steps = 1000          #Evaluate performance after every xxxx steps
+dev_sentences = 1000                 #Number of parallel sentences to be used for development
+
+
+######## Start the extension of the teacher model to multiple languages ########
+logger.info("Load teacher model")
+teacher_model = SentenceTransformer(teacher_model_name,device='mps')
+
+
+logger.info("Create student model from scratch")
+
+word_embedding_model = models.Transformer(student_model_name, max_seq_length=max_seq_length)
+# Apply mean pooling to get one fixed sized sentence vector
+pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())#denseで次元数を７６８にする次元数をいじる
+student_model = SentenceTransformer(modules=[word_embedding_model, pooling_model],device='mps')
+
+print(teacher_model)
+print(student_model)
+
+
+from sentence_transformers.datasets import ParallelSentencesDataset
+
+train_data = ParallelSentencesDataset(student_model=student_model, teacher_model=teacher_model)
+train_data.load_data('/Users/fujidai/dataseigen/09-04_09-04.txt')#日本語英語をタブで繋げたやつ
+#train_data.load_data('/Users/fujidai/TED2020_data/wmt21/output-100.txt')#日本語英語をタブで繋げたやつ
+
+#train_data.load_data('/Users/fujidai/TED2020_data/data/tuikazumi/en-ja/TED2020.en-ja.en')
+train_dataloader = DataLoader(train_data, shuffle=True, batch_size=train_batch_size)
+train_loss = losses.MSELoss(model=student_model)
+
+print(train_data)
+
+
+#50000_all-MiniLM-L6-v2__paraphrase-distilroberta-base-v2_epoch-1
+
+# Train the model
+print('az')
+student_model.fit(train_objectives=[(train_dataloader, train_loss)],
+          epochs=num_epochs,
+          #device=device,
+          warmup_steps=num_warmup_steps,
+          evaluation_steps=num_evaluation_steps,
+          #output_path='best_paraphrase-mpnet-base-v2__xlm-roberta-base_epoch-3',
+          #save_best_model=True,
+          optimizer_params= {'lr': 2e-5, 'eps': 1e-6},
+          checkpoint_path='paraphrase-mpnet-base-v2_09-MarginMSELoss-finetuning-7-5_2',
+          checkpoint_save_steps=820
+          )
+
+student_model.save('paraphrase-mpnet-base-v2_09-MarginMSELoss-finetuning-7-5')
+
+
+
+
+
+
+
+
+
+
+
+#

finetune.py

@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Fri Jun 30 08:47:31 2023
+
+@author: fujidai
+"""
+
+
+import torch
+from sentence_transformers import SentenceTransformer, InputExample, losses,models
+from sentence_transformers import SentenceTransformer,  SentencesDataset, LoggingHandler, losses
+from sentence_transformers.readers import InputExample
+from torch.utils.data import DataLoader
+from transformers import AutoTokenizer
+from sentence_transformers.SentenceTransformer import SentenceTransformer
+import torch
+import torch.nn.functional as F
+import numpy as np
+from sentence_transformers import SentenceTransformer, util
+
+
+word_embedding_model = models.Transformer('/Users/fujidai/sinTED/xlm-roberta-base', max_seq_length=510)# modelの指定をする
+pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())
+#dense_model = models.Dense(in_features=pooling_model.get_sentence_embedding_dimension(),out_features=16)
+model = SentenceTransformer(modules=[word_embedding_model, pooling_model],device='mps')
+print(model)
+
+
+with open('/Users/fujidai/dataseigen/up/pseudo-pseudo-english_english_100000_cos-sim-karanasi_09-04.txt', 'r') as f:#Negative en-ja cos_sim
+    raberu = f.read()
+raberu_lines = raberu.splitlines()#改行コードごとにリストに入れている
+data = []
+for i in range(len(raberu_lines)):
+    data.append(float(raberu_lines[i]))#Negative en-ja cos_simをdataに入れている
+
+with open('/Users/fujidai/dataseigen/up/pseudo-pseudo_en-ja-100000-karanasi_09-04.txt', 'r') as f:#TEDのenglish
+    left = f.read()
+left_lines = left.splitlines()
+
+with open('/Users/fujidai/dataseigen/up/pseudo-pseudo_ja-en-100000-karanasi_09-04.txt', 'r') as f:#TEDのjapanese
+    senter = f.read()
+senter_lines = senter.splitlines()
+
+with open('/Users/fujidai/dataseigen/up/pseudo-japanese-sentence-100000-karanasi_09-04.txt', 'r') as f:#pseudo japanese (TEDのenglishをgoogle翻訳に入れた疑似コーパス)
+    right = f.read()
+right_lines = right.splitlines()#改行コードごとにリストに入れている
+
+
+train_examples = []
+for i in range(len(left_lines)):
+    pair=[]
+    pair.append(left_lines[i])#left_lines側のi行目をtextsに追加している
+    pair.append(senter_lines[i])
+    pair.append(right_lines[i])#right_lines側のi行目をtextsに追加している
+    example = InputExample(texts=pair, label=1-data[i])#textsをラベル付きで追加している
+    #label=1-data[i]の１は positive cos_sim
+    train_examples.append(example)#学習として入れるものに入れている
+
+with open('/Users/fujidai/dataseigen/down/pseudo-english_english_100000_cos-sim-karanasi_09-04.txt', 'r') as f:##Negative ja-en cos_sim
+    raberu2 = f.read()
+raberu2_lines = raberu2.splitlines()#改行コードごとにリストに入れている
+data2 = []
+for i in range(len(raberu2_lines)):
+    data2.append(float(raberu2_lines[i]))#Negative ja-en cos_simをdata2に入れている
+
+with open('/Users/fujidai/dataseigen/down/pseudo-ja-en-100000-karanasi_09-04.txt', 'r') as f:#TEDのjapanese
+    left2 = f.read()
+left2_lines = left2.splitlines()
+
+with open('/Users/fujidai/dataseigen/down/pseudo-en-ja-100000-karanasi_09-04.txt', 'r') as f:#TEDのenglish
+    senter2 = f.read()
+senter2_lines = senter2.splitlines()
+
+with open('/Users/fujidai/dataseigen/down/pseudo-english-sentence-100000-karanasi_09-04.txt', 'r') as f:#pseudo english (TEDのjapaneseをgoogle翻訳に入れた疑似コーパス)
+    right2 = f.read()
+right2_lines = right2.splitlines()#改行コードごとにリストに入れている
+
+for i in range(len(left2_lines)):
+    pair=[]
+    pair.append(left2_lines[i])#left_lines側のi行目をtextsに追加している
+    pair.append(senter2_lines[i])
+    pair.append(right2_lines[i])#right_lines側のi行目をtextsに追加している
+    example = InputExample(texts=pair, label=1-data2[i])#textsをラベル付きで追加している
+    #label=1-data2[i]の１は positive cos_sim
+    train_examples.append(example)#学習として入れるものに入れている
+
+
+device = torch.device('mps')
+#print(device)
+
+import torch.nn.functional as F
+
+
+train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=8)
+train_loss = losses.MarginMSELoss(model=model,similarity_fct=F.cosine_similarity)
+
+
+#Tune the model
+model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=3, warmup_steps=1000,show_progress_bar=True,
+          #output_path='完成2best-6-30',
+          checkpoint_path='checkpoint_savename',checkpoint_save_steps=2300,#どのくらいのイテレーションごとに保存するか
+          save_best_model=True)#checkpoint_save_total_limit=5,
+model.save("savename")