File size: 8,003 Bytes
ba3ecf1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 | import logging
import sys
from datetime import datetime
import torch
from datasets import load_dataset, DatasetDict
from sentence_transformers import SentenceTransformer, losses
from sentence_transformers.evaluation import TripletEvaluator, SequentialEvaluator
from sentence_transformers.trainer import SentenceTransformerTrainer
from sentence_transformers.training_args import (
SentenceTransformerTrainingArguments,
BatchSamplers,
MultiDatasetBatchSamplers,
)
# ─────────────────────────────────────────────────────────────
# Logging
# ─────────────────────────────────────────────────────────────
logging.basicConfig(
format="%(asctime)s - %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
level=logging.INFO,
handlers=[logging.FileHandler("logs.txt")],
)
class Tee:
def __init__(self, *files): self.files = files
def write(self, obj):
for f in self.files: f.write(obj); f.flush()
def flush(self):
for f in self.files: f.flush()
def isatty(self): return False
sys.stdout = Tee(sys.stdout, open("logs.txt", "a"))
# ─────────────────────────────────────────────────────────────
# Config
# ─────────────────────────────────────────────────────────────
train_batch_size = 32
model_name = "/home/skiredj.abderrahman/khalil/sbert_training/third_training/output/arabert_20260326_1254/final"
model_nickname = "arabert_arcd_squad"
timestamp = datetime.now().strftime("%Y%m%d_%H%M")
output_dir = f"output/{model_nickname}_{timestamp}"
device = "cuda" if torch.cuda.is_available() else "cpu"
matryoshka_dims = [768, 512, 256, 128, 64]
# ─────────────────────────────────────────────────────────────
# 1. Model
# ─────────────────────────────────────────────────────────────
model = SentenceTransformer(model_name, device=device)
model.set_pooling_include_prompt(include_prompt=False)
print(f"Model running on: {device}")
# ─────────────────────────────────────────────────────────────
# TEST MODE
# ─────────────────────────────────────────────────────────────
TEST_MODE = False
TEST_SAMPLES = 100
# ─────────────────────────────────────────────────────────────
# 2. Dataset
# ─────────────────────────────────────────────────────────────
logging.info("Loading datasets...")
def load_csv(path):
ds = load_dataset("csv", data_files=path)["train"]
ds = ds.filter(lambda x: all(x[col] is not None and str(x[col]).strip() != "" for col in ds.column_names))
if TEST_MODE:
ds = ds.select(range(min(TEST_SAMPLES, len(ds))))
return ds
train_data = load_csv("/home/skiredj.abderrahman/khalil/sbert_training/fourth_training/data/pre_train_arcd_SQuAD_triplets.csv")
val_data = load_csv("/home/skiredj.abderrahman/khalil/sbert_training/fourth_training/data/pre_test_arcd_SQuAD_triplets.csv")
train_dataset = DatasetDict({"arcd_squad": train_data})
eval_dataset = DatasetDict({"arcd_squad": val_data})
logging.info(train_dataset)
logging.info(eval_dataset)
# ─────────────────────────────────────────────────────────────
# 3. Loss
# ─────────────────────────────────────────────────────────────
loss = {
"arcd_squad": losses.MatryoshkaLoss(
model,
losses.MultipleNegativesRankingLoss(model),
matryoshka_dims=matryoshka_dims,
)
}
# ─────────────────────────────────────────────────────────────
# 4. Evaluator
# ─────────────────────────────────────────────────────────────
def make_triplet_evaluators(dataset, name_prefix, max_samples=3_000):
sample = dataset.shuffle(seed=42).select(range(min(max_samples, len(dataset))))
return [
TripletEvaluator(
anchors=sample["anchor"],
positives=sample["positive"],
negatives=sample["negative"],
name=f"{name_prefix}-{dim}",
truncate_dim=dim,
)
for dim in matryoshka_dims
]
dev_evaluator = SequentialEvaluator(
make_triplet_evaluators(val_data, "val-arcd-squad"),
main_score_function=lambda scores: scores[0],
)
logging.info("Pre-training evaluation:")
dev_evaluator(model)
# ─────────────────────────────────────────────────────────────
# 5. Training Arguments
# ─────────────────────────────────────────────────────────────
args = SentenceTransformerTrainingArguments(
output_dir=output_dir,
seed=42,
num_train_epochs=3,
per_device_train_batch_size=train_batch_size,
per_device_eval_batch_size=train_batch_size,
gradient_accumulation_steps=2,
bf16=True,
fp16=False,
learning_rate=1e-5,
lr_scheduler_type="linear",
warmup_ratio=0.1,
weight_decay=0.01,
batch_sampler=BatchSamplers.NO_DUPLICATES,
multi_dataset_batch_sampler=MultiDatasetBatchSamplers.PROPORTIONAL,
dataloader_num_workers=8,
eval_strategy="steps",
eval_steps=1000,
save_strategy="steps",
save_steps=1000,
save_total_limit=2,
report_to="tensorboard",
logging_steps=100,
logging_dir=f"{output_dir}/runs",
)
# ─────────────────────────────────────────────────────────────
# 6. Trainer
# ─────────────────────────────────────────────────────────────
trainer = SentenceTransformerTrainer(
model=model,
args=args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
loss=loss,
evaluator=dev_evaluator,
)
trainer.train()
print("Finished")
# ─────────────────────────────────────────────────────────────
# 7. Save
# ─────────────────────────────────────────────────────────────
final_output_dir = f"{output_dir}/final"
model.save(final_output_dir)
print(f"Model saved to {final_output_dir}")
|