import joblib;
from huggingface_hub import hf_hub_download;
from peft import PeftModel, PeftConfig;
from transformers import AutoTokenizer, TextClassificationPipeline, AutoModelForSequenceClassification;
from huggingface_hub import HfApi, login
with open('./api_key/HGF_TOKEN.txt', 'r') as hgf:
    login(token=hgf.read())
api = HfApi()
repo_id = "x2bee/plateer_classifier_v0.1"
data_id = "x2bee/plateer_category_data"

# Load Config, Tokenizer, Label_Encoder
config = PeftConfig.from_pretrained(repo_id, subfolder="last-checkpoint")
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="last-checkpoint")
label_encoder_file = hf_hub_download(repo_id=data_id, repo_type="dataset", filename="label_encoder.joblib")
label_encoder = joblib.load(label_encoder_file)

# Load base_model
base_model = AutoModelForSequenceClassification.from_pretrained("Qwen/Qwen2.5-1.5B", num_labels=17)
base_model.resize_token_embeddings(len(tokenizer))

# Load Model
model = PeftModel.from_pretrained(base_model, repo_id, subfolder="last-checkpoint")

import torch
class TextClassificationPipeline(TextClassificationPipeline):
    def __call__(self, inputs, top_k=5, **kwargs):
        inputs = self.tokenizer(inputs, return_tensors="pt", truncation=True, padding=True, **kwargs)
        inputs = {k: v.to(self.model.device) for k, v in inputs.items()}
        
        with torch.no_grad():
            outputs = self.model(**inputs)
        
        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
        scores, indices = torch.topk(probs, top_k, dim=-1)
        
        results = []
        for batch_idx in range(indices.shape[0]):
            batch_results = []
            for score, idx in zip(scores[batch_idx], indices[batch_idx]):
                temp_list = []
                label = self.model.config.id2label[idx.item()]
                label = int(label.split("_")[1])
                temp_list.append(label)
                predicted_class = label_encoder.inverse_transform(temp_list)[0]
                            
                batch_results.append({
                    "label": label,
                    "label_decode": predicted_class,
                    "score": score.item(),
                })
            results.append(batch_results)
        
        return results

classifier_model = TextClassificationPipeline(tokenizer=tokenizer, model=model)

def plateer_classifier(text, top_k=3):
    result = classifier_model(text, top_k=top_k)
    return result

Run

python

user_input = "머리띠"
result = plateer_classifier(user_input)[0]
print(result)

bash

{'label': 6, 'label_decode': '뷰티/케어', 'score': 0.42996299266815186}
{'label': 15, 'label_decode': '패션/의류/잡화', 'score': 0.1485249102115631}
{'label': 8, 'label_decode': '스포츠', 'score': 0.1281907707452774}

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 8
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 4
gradient_accumulation_steps: 4
total_train_batch_size: 128
total_eval_batch_size: 32
optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 10000
num_epochs: 1
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
0.5023	0.0292	5000	0.5044	0.8572
0.4629	0.0585	10000	0.4571	0.8688
0.4254	0.0878	15000	0.4201	0.8770
0.4025	0.1171	20000	0.4016	0.8823
0.3635	0.3220	55000	0.3623	0.8905
0.3192	0.6441	110000	0.3242	0.8997

Framework versions

PEFT 0.13.2
Transformers 4.46.3
Pytorch 2.2.1
Datasets 3.1.0
Tokenizers 0.20.3