aidlab/lc_stylist_agent
7
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Enable data auto process for new data

Browse Source
This commit is contained in:
pangkaicheng
2025-12-10 17:27:56 +08:00
parent 0b1d948f77
commit 0e9546aa1a
12 changed files with 936 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
app/server/ChatbotAgent/agent_server.py
View File

@@ -51,7 +51,7 @@ class AgentRequestModel(BaseModel):
session_id: str
num_outfits: int
stylist_path: str
batch_source: str
batch_sources: List[str]
callback_url: str
gender: str
max_len: int = 9
@@ -112,8 +112,8 @@ class LCAgent(ls.LitAPI):
request_summary=request_summary,
occasions=occasions,
stylist_name=request.stylist_path,
batch_source=request.batch_source,
start_outfit=[],
batch_sources=request.batch_sources,
num_outfits=request.num_outfits,
user_id=request.user_id,
gender=request.gender,
@@ -162,7 +162,7 @@ class LCAgent(ls.LitAPI):
return str(parsed_result.summary), [occ.value for occ in parsed_result.occasions]
async def recommend_outfit(
self, request_summary: str, occasions: List[str], batch_source: str, stylist_name: str, start_outfit=[],
self, request_summary: str, occasions: List[str], stylist_name: str, start_outfit: List = [], batch_sources: List[str] = [],
num_outfits: int = 1, user_id: str = "test", gender: str = "male",
callback_url: str = None, max_len: int = 9, outfit_ids=None
):
@@ -186,9 +186,9 @@ class LCAgent(ls.LitAPI):
task = agent.run_styling_process(
request_summary=request_summary,
occasions=occasions,
batch_source=batch_source,
stylist_name=stylist_name,
start_outfit=start_outfit,
batch_sources=batch_sources,
user_id=user_id,
callback_url=callback_url,
gender=gender,
@@ -227,9 +227,9 @@ class LCAgent(ls.LitAPI):
new_task = agent.run_styling_process(
request_summary=request_summary,
occasions=occasions,
batch_source=batch_source,
stylist_name=stylist_name,
start_outfit=start_outfit,
batch_sources=batch_sources,
user_id=user_id,
callback_url=callback_url
)
@@ -295,9 +295,9 @@ if __name__ == "__main__":
task = agent.run_styling_process(
request_summary=request_summary,
occasions=occasions,
batch_source="2025_q4",
stylist_name=stylist_name,
start_outfit=[],
batch_sources=["2025_q4"],
user_id=test_content['test_case_id'],
callback_url="http://mock-callback.com/result",
gender="female",

73
app/server/ChatbotAgent/core/stylist_agent_server.py
View File

@@ -16,11 +16,16 @@ from app.server.utils.img_operation import merge_images_to_square
from app.server.utils.minio_client import minio_client, oss_upload_image
from app.server.utils.request_post import post_request
from app.config import settings
from app.taxonomy import CLOTHING_CATEGORY, ACCESSORY_CATEGORY
from app.taxonomy import CATEGORY, ALL_CATEGORY, IGNORE_CATEGORY
logger = logging.getLogger(__name__)
IGNORE_CATEGORY = set(IGNORE_CATEGORY)
CLOTHING_CATEGORY = set(CATEGORY['clothing'] + CATEGORY['shoes'] + CATEGORY['bags']) - IGNORE_CATEGORY
ACCESSORY_CATEGORY = set(CATEGORY['accessories']) - IGNORE_CATEGORY
class AsyncStylistAgent:
def __init__(self, local_db, max_len: int, gemini_model_name: str, outfit_id=str):
# self.outfit_items: List[Dict[str, str]] = []
@@ -145,7 +150,7 @@ class AsyncStylistAgent:
```
* `action`: Must always be `"recommend_item"` until the outfit is complete.
* `category`: Must be an unused category from the following list: {CLOTHING_CATEGORY} (strictly no repeats, per the Category Uniqueness Mandate).
* `category`: Must be an unused category from the following list: {list(CLOTHING_CATEGORY)} (strictly no repeats, per the Category Uniqueness Mandate).
* `description`: This must be an **extremely detailed and precise** description of the item. This description is used for **high-accuracy vector search** in the database and must include:
* **Color** (e.g., milk tea, pure white, dark gray)
* **Fit/Silhouette** (e.g., Oversize, loose, slim-fit)
@@ -193,7 +198,7 @@ class AsyncStylistAgent:
---
## STRICT RULES
1. **Batch Recommendation**: Do NOT recommend items one by one. You must output the **COMPLETE LIST** of accessories (e.g., jewelry, bag, watch, hat) in a single response using the 'recommended_accessories' list.
2. **Allowed Categories**: Select only from: {ACCESSORY_CATEGORY}.
2. **Allowed Categories**: Select only from: {list(ACCESSORY_CATEGORY)}.
3. **Harmony & Constraints**:
- The accessories must complement the [Current Outfit Base].
- Strictly follow the [Accessories Style Guide] regarding metals (gold/silver), numbers, and prohibited items.
@@ -295,40 +300,46 @@ class AsyncStylistAgent:
print(f"Raw response: {response_text}")
return None
def _get_next_item(self, item_description: str, category: str, occasions: List[str], batch_source: str = "2025_q4", gender: str = "female") -> Optional[Dict[str, str]]:
def _get_next_item(self, item_description: str, category: str, occasions: List[str], batch_sources: List[str] = [], gender: str = "female") -> Optional[Dict[str, str]]:
"""
1. 根据描述生成嵌入。
2. 查询本地数据库以找到最佳匹配项。
3. 模拟 Agent 审核匹配项(这里简化为总是通过)。
"""
# 1. 生成查询嵌入
query_embedding = self.local_db.get_clip_embedding(item_description, is_image=False)
# 2. 执行查询,并过滤类别
try:
# 1. 生成查询嵌入
query_embedding = self.local_db.get_clip_embedding(item_description, is_image=False)
results = self.local_db.get_matched_item(
query_embedding,
category,
occasions=occasions,
batch_sources=batch_sources,
gender=gender,
n_results=1
)
except ValueError as e:
print(f"检测到无效参数错误：{e}")
results = []
# 2. 执行查询,并过滤类别
results = self.local_db.get_matched_item(query_embedding, category, occasions=occasions, batch_source=batch_source, gender=gender, n_results=1)
if not results:
print(f"❌ 数据库中未找到符合 '{category}' 和描述的单品。")
return None
# 3. 模拟 Agent 审核(实际应用中,你需要将图片发回给 Agent进行审核)
best_meta = results[0] # 第一个 batch 的第一个 metadata
item_id = best_meta['item_id'].replace("_img", "")
return {
"item_id": item_id, # 从 metadata 字典中安全获取
"category": best_meta['category'],
"gpt_description": item_description,
'description': best_meta['description'],
# 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导
# 这里假设 item_id 就是文件名的一部分
"image_path": os.path.join(f"{item_id}.jpg")
}
except Exception as e:
print(f"An error occurred during item retrieval: {e}")
if not results:
print(f"数据库中未找到符合 '{category}' 和描述的单品。")
return None
# 3. 模拟 Agent 审核(实际应用中,你需要将图片发回给 Agent进行审核)
best_meta = results[0] # 第一个 batch 的第一个 metadata
item_id = best_meta['item_id'].replace("_img", "")
return {
"item_id": item_id, # 从 metadata 字典中安全获取
"category": best_meta['category'],
"gpt_description": item_description,
'description': best_meta['description'],
# 假设 'item_path' 存储在 metadata 中，或从 'item_id' 推导
# 这里假设 item_id 就是文件名的一部分
"image_path": os.path.join(f"{item_id}.jpg")
}
def _build_user_input(self, recommend_acc=False) -> str:
"""构建发送给 Gemini 的用户输入,包含已选单品信息。"""
if not self.outfit_items:
@@ -353,7 +364,7 @@ class AsyncStylistAgent:
response = post_request(url=callback_url, data=json.dumps(response_data), headers=self.headers)
logger.info(f"request data ：{response_data} | JAVA callback info -> status:{response.status_code} | message:{response.text}")
async def run_styling_process(self, request_summary, occasions, stylist_name, batch_source="2025_q4", start_outfit=[], user_id="test", callback_url="", gender: str = "male"):
async def run_styling_process(self, request_summary, occasions, stylist_name, start_outfit=[], batch_sources=[], user_id="test", callback_url="", gender: str = "male"):
self.outfit_items = start_outfit
"""主流程控制循环。"""
print(f"--- Starting Agent (Outfit ID: {self.outfit_id}) ---")
@@ -412,7 +423,7 @@ class AsyncStylistAgent:
continue
# 4b. 在本地 DB 中查询单品
new_item = self._get_next_item(description, category, occasions, batch_source, gender)
new_item = self._get_next_item(description, category, occasions, batch_sources, gender)
if not new_item or new_item['item_id'] in [x['item_id'] for x in self.outfit_items]:
self.post_operation(
response_data,
@@ -460,7 +471,7 @@ class AsyncStylistAgent:
continue
# 4b. 在本地 DB 中查询单品
new_item = self._get_next_item(description, category, occasions, batch_source, gender)
new_item = self._get_next_item(description, category, occasions, batch_sources, gender)
if not new_item or new_item['item_id'] in [x['item_id'] for x in self.outfit_items]:
continue
else:

147
app/server/ChatbotAgent/core/vector_database.py
View File

@@ -8,13 +8,12 @@ from PIL import Image
from typing import List, Dict, Any
from transformers import CLIPProcessor, CLIPModel
from app.taxonomy import CATEGORY, OCCASION
from app.taxonomy import OCCASION, ALL_CATEGORY
class VectorDatabase():
def __init__(self, vector_db_dir: str, collection_name: str, embedding_model_name: str):
self.client = chromadb.PersistentClient(path=vector_db_dir)
self.collection = self.client.get_or_create_collection(
name=collection_name,
configuration={
@@ -23,17 +22,9 @@ class VectorDatabase():
}
}
)
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.model = CLIPModel.from_pretrained(embedding_model_name).to(self.device)
self.processor = CLIPProcessor.from_pretrained(embedding_model_name)
# self.cache_filtered_ids = self.load_filtered_ids([
# {"item_group_id": {"$ne": "Clothing"}},
# {"item_group_id": {"$ne": "Shoes"}},
# {"modality": "image"}
# ])
# self.total_count = len(self.cache_filtered_ids)
def get_clip_embedding(self, data: str | Image.Image, is_image: bool) -> List[float]:
"""生成图像或文本的 CLIP 嵌入，并进行 L2 归一化。"""
@@ -57,46 +48,32 @@ class VectorDatabase():
features = features / features.norm(p=2, dim=-1, keepdim=True)
return features.cpu().numpy().flatten().tolist()
def query_local_db(self, embedding: List[float], category: str, occasions: List[str] = [], n_results: int = 3) -> List[Dict[str, Any]]:
"""
基于嵌入向量在本地数据库中查询相似单品。
实际应执行 ChromaDB 查询，并根据 category 进行过滤(metadatas)。
"""
for occasion in occasions:
where_clauses = {
"$and": [
{"category": category},
{"modality": "image"},
{"batch_source": '2025_q4'}
]
}
if occasion not in OCCASION:
continue
else:
where_clauses['$and'].append({occasion: 1})
results = self.collection.query(
query_embeddings=[embedding],
n_results=n_results,
where=where_clauses,
include=['metadatas', 'distances']
)
return results
def get_matched_item(self, embedding: List[float], category: str, occasions: List[str] = [], batch_source: str = "2025_q4", gender: str = 'female', n_results: int = 1) -> List[Dict[str, Any]]:
def get_matched_item(self, embedding: List[float], category: str, occasions: List[str] = [], batch_sources: List[str] = [], gender: str = 'female', n_results: int = 1) -> List[Dict[str, Any]]:
if category not in ALL_CATEGORY:
raise ValueError(f"Recommended {category} is not valid.")
and_conditions = [
{"category": category},
{"modality": "image"},
{"$or": [
{"gender": gender},
{"gender": "unisex"},
]}
]
if batch_sources and len(batch_sources) > 0:
source_conditions = []
for source in batch_sources:
source_conditions.append({"batch_source": source})
# 将 Batch Source 的 OR 子句添加到主 AND 条件中
and_conditions.append({"$or": source_conditions})
results = self.collection.query(
query_embeddings=[embedding],
n_results=500,
where={
"$and": [
{"category": category},
{"modality": "image"},
{"gender": gender},
{"batch_source": batch_source}
]
},
include=['metadatas', 'distances']
where={"$and": and_conditions},
include=['metadatas', 'distances'],
)
if not results['ids'][0]:
return []
@@ -124,7 +101,7 @@ class VectorDatabase():
score_occ = score_occ / count if count else 0.0
final_score = 0.6 * score_vec + 0.3 * score_occ
final_score = 0.6 * score_vec + 0.4 * score_occ
final_scores.append(final_score)
scores_arr = np.array(final_scores)
@@ -139,79 +116,3 @@ class VectorDatabase():
sampled_index = np.random.choice(a=len(results['ids'][0]), p=probabilities, size=n_results, replace=False) # 不重复采样
sampled_items = [metadatas[i] for i in sampled_index]
return sampled_items
def load_filtered_ids(self, filter_item):
# print("\n--- 初始化阶段：加载所有符合条件的 ID ---")
start_time = time.time()
FILTER_CRITERIA = {
"$and": filter_item
}
MAX_LIMIT = 100000
try:
# 获取所有符合条件的 ID
all_ids_results = self.collection.get(
where=FILTER_CRITERIA,
limit=MAX_LIMIT,
include=[]
)
all_matched_ids = all_ids_results['ids']
# print(f"🎉 成功加载 {len(all_matched_ids)} 个 ID 到缓存。")
print(time.time() - start_time)
return all_matched_ids
except Exception as e:
print(f"❌ 初始化失败：获取 ID 列表时发生错误: {e}")
return []
def random_get_accessories(self, ids):
# 2. 调用 ChromaDB：只查询这一个 ID 的详细信息
try:
final_results = self.collection.get(
ids=ids,
include=["metadatas"] # 你只需要元数据
)
# 提取结果
if final_results['ids']:
return final_results
else:
return None
except Exception as e:
print(f"❌ 获取最终记录时发生错误: {e}")
return None
if __name__ == '__main__':
stylist = {
'text': "gold necklace",
'count': 2,
'category': "Jewelry"
}
max_len = 5
local_db = VectorDatabase(vector_db_dir="/workspace/lc_stylist_agent/db", collection_name="lc_clothing_embedding", embedding_model_name="openai/clip-vit-base-patch32")
A = local_db.load_filtered_ids([
{"item_group_id": {"$ne": "Clothing"}},
{"item_group_id": {"$ne": "Shoes"}},
{"modality": "image"}
])
# print(db.random_get_accessories())
start_time = time.time()
X = local_db.random_get_accessories(['ELI699_img'])
print(X)
print(time.time() - start_time)
# query_embedding = local_db.get_clip_embedding(stylist['text'], is_image=False)
#
# results = local_db.query_local_db(query_embedding, stylist['category'], n_results=10)
# # 2. 从结果集中抽 stylist['count'] 个item
# stylist_item = random.choices(results['metadatas'][0], k=stylist['count'])
# stylist_item_ids = [item_id['item_id'] for item_id in stylist_item]
#
# # 3. 从随机库中抽取配饰，总数达到9件 ，需过滤掉已经抽中的item
# accessories_count = 9 - max_len - stylist['count']
#
# random_single_ids = random.choices(list(set(local_db.cache_filtered_ids) - set([f"{i}_img" for i in stylist_item_ids])), k=accessories_count)
# random_items = local_db.random_get_accessories(random_single_ids)['metadatas']
# all_items = stylist_item + random_items

62
app/taxonomy.py
View File

@@ -1,15 +1,4 @@
# 这个文件用来储存所有的category和occasion，这是标准文件。
CATEGORY = [
'shoes', 'bags', 'dresses', 'tops', 'pants', 'skirts', 'outerwear', 'swimwear', 'suits',
'watches', 'sunglasses', 'belts', 'hats', 'jewelry', 'neckties', 'scarves & shawls'
]
CLOTHING_CATEGORY = [
'shoes', 'bags', 'dresses', 'tops', 'pants', 'skirts', 'outerwear', 'swimwear'
]
ACCESSORY_CATEGORY = [
'watches', 'sunglasses', 'belts', 'hats', 'jewelry', 'neckties', 'scarves & shawls'
]
OCCASION = [
"Casual", "Formal", "Activewear", "Resort", "Evening", "Outdoor",
"Business / workwear", "Cocktail / Semi-Formal", "Black Tie / White Tie",
@@ -17,3 +6,54 @@ OCCASION = [
"Travel / Transit", "Athleisure", "Beach / Swim", "Ski / Snow / Mountain",
"Garden Party / Daytime Event"
]
CATEGORY = {
'clothing': [
'coats',
'jackets',
'blazers',
'puffer',
'cardigan',
'sweater',
'shirts',
't-shirts',
'pullover',
'polos',
'bodysuits',
'dresses',
'skirts',
'jeans',
'shorts',
'leggings',
'jumpsuits',
'swimwear',
],
'shoes': [
'sneakers',
'formal shoes',
'heels',
'flats',
'sandals',
'slides',
'boots',
],
'bags': [
'bags'
],
'accessories': [
'necklaces',
'bracelets',
'jewellery',
'eyewear',
'scarves',
'hats',
'gloves',
'belts',
'socks',
'watches'
'ties',
]
}
ALL_CATEGORY = sum(CATEGORY.values(), [])
IGNORE_CATEGORY = ['socks']

92
data_ingestion/README.md Normal file
View File

@@ -0,0 +1,92 @@
## Steps
1. Prepare products-all.json and image_data (folder) using javascript to download. These files should be saved in `./data/BATCH_SOURCE` which is a new folder. Give a new batch_source id to each new incoming data.
1. Run `process_item.py` to categorize category, gender and occasions for each data. Output to `./data/{BATCH_SOURCE}/metadata_extraction.json`. This should be running on H200 device.
3. Organize all data and then embed them into db locally using `run_ingestion.py`
## Raw Data Structure
```json
## products-all.json
{
"id": "BUL808",
"name": "SARAH ZHUANG - 'Click & Link' diamond 18k gold earrings",
"brand": "SARAH ZHUANG",
"category": "Fine Jewellery And Watches",
"subcategory": "General",
"price": 17500,
"currency": "HKD",
"description": "Sarah Zhuang's Click & Link earrings embrace the allure of geometry. Forged into elegant rectangles with one side encrusted with diamonds, this gold pair will certainly elevate your cocktail ensembles.",
"tags": [
"sarah zhuang",
"fine jewellery and watches",
"in-stock",
"new",
"sarah",
"zhuang",
"'click",
"link'",
"diamond"
],
"imageUrl": "https://media.lanecrawford.com/B/U/L/BUL808_in_xl.jpg",
"url": "https://www.lanecrawford.com.hk/product/sarah-zhuang/-click-link-diamond-18k-gold-earrings/_/BUL808/product.lc?utm_medium=embed&utm_source=ai-recommended&utm_campaign=2025-christmas_lc_ai-recommended",
"color": "YELLOW GOLD",
"groupName": "Fine Jewellery",
"deptName": "Women's Fine Jewellery",
"onlineBU": "Fine Jewellery",
"stockAvailability": true
}
```
## Example in `metadata_extraction.json`
```json
"EOJ367": {
"category": "shoes",
"gender": "female",
"applicable_occasions": [
"Casual",
"Outdoor",
"Travel / Transit"
],
"inappropriate_occasions": [
"Formal",
"Black Tie / White Tie",
"Bridal / Wedding",
"Business / workwear",
"Cocktail / Semi-Formal"
]
}
```
## Metadata in Vector Database
```json
{
'item_id': 'EOJ128',
'category': 'sunglasses',
'gender': 'unisex',
'modality': 'image',
'brand': 'CELINE',
'color': 'BROWN',
'description': "Immerse yourself in the depth of classic style with CELINE\'s Tortoiseshell Logo Sunglasses. Featuring a rich, tortoiseshell acetate frame and adorned with the iconic CELINE logo in gold, these sunglasses are a testament to timeless elegance and luxury. Perfect for those who appreciate a sophisticated aesthetic, they offer optimal UV protection while ensuring you remain at the forefront of fashion.",
'tags': 'celine,accessories,in-stock,new,maxi,triomphe,acetate,round',
'price': 4500,
'url': 'https://www.lanecrawford.com.hk/product/celine/maxi-triomphe-acetate-round-sunglasses/_/EOJ128/product.lc?utm_medium=embed&utm_source=ai-recommended&utm_campaign=2025-christmas_lc_ai-recommended',
'batch_source': '2025_q4',
'Outdoor': 0,
'Ski / Snow / Mountain': 0,
'Festival / Concert': 0,
'Activewear': 0,
'Casual': 1,
'Cocktail / Semi-Formal': -1,
'Formal': -1,
'Party / Clubbing': 0,
'Evening': 0,
'Travel / Transit': 0,
'Beach / Swim': 0,
'Garden Party / Daytime Event': 1,
'Black Tie / White Tie': -1,
'Resort': 1,
'Athleisure': 0,
'Business / workwear': -1,
'Bridal / Wedding': -1,
}
```

280
data_ingestion/process_item.py Normal file
View File

@@ -0,0 +1,280 @@
import torch
import os
from transformers import AutoProcessor, AutoModelForVision2Seq
from PIL import Image
import json
from tqdm import tqdm
from app.taxonomy import OCCASION, CATEGORY, ALL_CATEGORY
# data config
BATCH_SOURCE = '2025_q4'
RAW_DATA_PATH = f'./data/{BATCH_SOURCE}/products-all.json'
IMAGE_DIR = f'./data/{BATCH_SOURCE}/image_data'
# MLLM config
MODEL_NAME = "meta-llama/Llama-3.2-11B-Vision-Instruct"
DEVICE = "cuda:0" # 确保设备设置正确，与您的 Traceback 匹配
BATCH_SIZE = 50
OUTPUT_FILE = f'./data/{BATCH_SOURCE}/metadata_extraction.json'
# Load Model
processor = AutoProcessor.from_pretrained(MODEL_NAME)
if processor.tokenizer.padding_side != 'left':
processor.tokenizer.padding_side = 'left'
print(f"Set tokenizer padding_side to '{processor.tokenizer.padding_side}' for correct generation.")
model = AutoModelForVision2Seq.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16).to(DEVICE)
model.eval()
# Load Data
with open(RAW_DATA_PATH, 'r', encoding='utf-8') as file:
data = json.load(file)
EXAMPLE_1_INFO = """
Product Name: ARMARIUM - Loren Wool Blend Tube Skirt
Category: Clothing / Bottoms
Color: RED
Description: Cut from cardinal-red virgin wool, Armarium's Loren skirt wields tailoring's exactitude in a column of colour. The low-slung waist and clean tube line are punctuated by a razor back slit—stride from boardroom to candlelit bar with modern hauteur.
Tags: armarium, clothing, in-stock, new, loren, wool, blend, tube
"""
EXAMPLE_1_JSON = json.dumps({
"category": "skirts",
"gender": "female",
"applicable_occasions": [
"Business/workwear", "Evening", "Cocktail / Semi-Formal", "Party / Clubbing", "Formal"
],
"inappropriate_occasions": [
"Activewear", "Beach / Swim", "Athleisure", "Ski / Snow / Mountain", "Casual"
]
}, indent=4)
# 示例 2：胸针 (Pin)
EXAMPLE_2_INFO = """
Product Name: TATEOSSIAN - Mayfair 18K Yellow Gold Rhodium Plated Sterling Silver Peg Pin
Category: Accessories / Accessories
Color: MULTI
Description: Crafted from 18k yellow gold and rhodium-plated sterling silver, this unique pins has been artfully finished with Tateossian's signature diamond engraving pattern.
Tags: tateossian, accessories, in-stock, new, mayfair, yellow, gold, rhodium
"""
EXAMPLE_2_JSON = json.dumps({
"category": "jewelry",
"gender": "female",
"applicable_occasions": [
"Formal", "Black Tie / White Tie", "Bridal / Wedding", "Business/workwear", "Cocktail / Semi-Formal"
],
"inappropriate_occasions": [
"Casual", "Activewear", "Beach / Swim", "Outdoor", "Athleisure", "Ski / Snow / Mountain"
]
}, indent=4)
# --- 2. 构造对话格式 Prompt ---
BOS_TOKEN = "<|begin_of_text|>"
EOS_TOKEN = "<|eot_id|>"
SYSTEM_HEADER = "<|start_header_id|>system<|end_header_id|>\n"
USER_HEADER = "<|start_header_id|>user<|end_header_id|>\n"
ASSISTANT_HEADER = "<|start_header_id|>assistant<|end_header_id|>\n"
IMAGE_TOKEN = "<|image|>"
def format_product_info(product):
tags_str = ", ".join(product.get('tags', []))
info = (
f"Product Name: {product.get('name', 'N/A')}\n"
f"Category: {product.get('category', 'N/A')} / {product.get('deptName', 'N/A')}\n"
f"Color: {product.get('color', 'N/A')}\n"
f"Description: {product.get('description', '')}\n"
f"Tags: {tags_str}",
f"groupName: {product.get('groupName', 'N/A')}\n"
f"onlineBU: {product.get('onlineBU', 'N/A')}\n"
)
return info
def generate_full_prompt(product_info, raw_category):
if raw_category == 'Fine Jewellery And Watches':
category = 'accessories'
else:
category = raw_category.lower()
subcategory_list = CATEGORY.get(category)
SYSTEM_PROMPT = f"""You are an expert fashion AI assistant. Your task is to analyze the provided product image and product details to:
1. determine the suitable occasions for wearing or using the item. You must choose occasions ONLY from the following strict list: {json.dumps(OCCASION, indent=4)}. Only relevant suitable or inappropriate occasions should be selected.
2. categorize it into suitable category in strict list: {json.dumps(subcategory_list)}.
3. categorize it into appropriate gender in ["female", "male", "unisex"]
Output Format:
Return ONLY a valid JSON object with four keys: "category", "gender", "applicable_occasions" and "inappropriate_occasions". Do not include any analysis or extra text outside of the final JSON object.
"""
# 组合对话序列
dialogue_prompt = (
# 1. System Instruction
f"{BOS_TOKEN}{SYSTEM_HEADER}{SYSTEM_PROMPT}{EOS_TOKEN}"
# 2. Example 1 (Few-Shot Round 1)
# 格式: <|start_header_id|>user<|end_header_id|>\n<|image|>\n{Text Instruction}<|eot_id|>
f"{USER_HEADER}\n{EXAMPLE_1_INFO}{EOS_TOKEN}"
f"{ASSISTANT_HEADER}{EXAMPLE_1_JSON}{EOS_TOKEN}"
# 3. Example 2 (Few-Shot Round 2)
f"{USER_HEADER}\n{EXAMPLE_2_INFO}{EOS_TOKEN}"
f"{ASSISTANT_HEADER}{EXAMPLE_2_JSON}{EOS_TOKEN}"
# 4. Target Item (Target Query)
f"{USER_HEADER}{IMAGE_TOKEN}\nInput Data:\n{product_info}{EOS_TOKEN}"
f"{ASSISTANT_HEADER}" # 最后的 Assistant Header 告诉模型从这里开始生成
)
return dialogue_prompt
# 2. 加载数据
products = data['products']
product_list = [
product for product in products
if product.get('category') in ['Clothing', 'Accessories', 'Shoes', 'Bags', 'Fine Jewellery And Watches']
and os.path.exists(os.path.join(IMAGE_DIR, f"{product.get('id')}.jpg"))
]
def validate_results():
if os.path.exists(OUTPUT_FILE):
with open(OUTPUT_FILE, 'r') as f:
final_results = json.load(f)
else:
final_results = {}
unfinished_ids = []
for product in product_list:
item_id = product.get('id')
if item_id not in final_results.keys():
unfinished_ids.append(product)
else:
processed_item = final_results[item_id]
category = processed_item.get("category")
gender = processed_item.get("gender")
if category not in ALL_CATEGORY:
unfinished_ids.append(product)
if gender not in ['female', 'male', 'unisex']:
unfinished_ids.append(product)
return unfinished_ids, final_results
attemps = 0
while attemps < 3:
attemps += 1
unfinished_products, final_results = validate_results()
completion_ratio = len(unfinished_products) / len(product_list)
if (completion_ratio > 0.95):
print("valid results surpass 95%. Finish Now.")
break
else:
print(f"Start {attemps} categorization process. Current ratio: {completion_ratio * 100}%")
try:
# 按照 BATCH_SIZE 进行切片迭代
for i in tqdm(range(0, len(unfinished_products), BATCH_SIZE)):
batch_samples = unfinished_products[i:i + BATCH_SIZE]
target_images = []
target_prompts = []
target_products_in_batch = []
# 准备当前批次的输入数据
for product in batch_samples:
product_id = product['id']
raw_category = product.get('category')
image_path = os.path.join(IMAGE_DIR, f"{product_id}.jpg")
try:
# 收集图片、Prompt 和产品数据
image = Image.open(image_path).convert("RGB")
product_info = format_product_info(product)
full_prompt = generate_full_prompt(product_info, raw_category)
target_images.append(image)
target_prompts.append(full_prompt)
target_products_in_batch.append(product)
except Exception as e:
# 跳过任何加载失败的单个样本
print(f"Skipping product {product_id} due to loading error: {e}")
continue
if not target_images:
continue # 如果整个批次都没有有效图片，跳过
# 4. 批量推理
print(f"\nProcessing batch {i//BATCH_SIZE + 1}/{int(len(unfinished_products)/BATCH_SIZE)+1} (Size: {len(target_images)})...")
# 处理器输入：使用嵌套列表 [[img1], [img2], ...]
inputs = processor(
images=[[img] for img in target_images],
text=target_prompts,
return_tensors="pt",
padding=True,
truncation=True
).to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=150,
do_sample=False
)
# 5. 批量解码和解析结果
input_lengths = inputs.input_ids.size(1)
for j in range(len(target_products_in_batch)):
product = target_products_in_batch[j]
product_id = product['id']
# 提取当前 item 的生成结果
# 注意: outputs 是 [batch_size, sequence_length]
newly_generated_tokens = outputs[j, input_lengths:]
generated_text = processor.decode(newly_generated_tokens, skip_special_tokens=True)
# 清理和解析
if generated_text.endswith(processor.tokenizer.eos_token):
generated_text = generated_text[:-len(processor.tokenizer.eos_token)]
try:
start_idx = generated_text.find('{')
end_idx = generated_text.rfind('}') + 1
if start_idx == -1 or end_idx == -1:
raise ValueError("JSON start or end delimiter not found.")
json_str = generated_text[start_idx:end_idx]
result_dict = json.loads(json_str)
final_results[product_id] = result_dict
except Exception as e:
print(f"ID {product_id}: FAILED to parse JSON. Raw Output: {generated_text.strip()}")
final_results[product_id] = {"error": str(e), "raw_output": generated_text.strip()}
# 显存清理（可选，但在长任务中推荐）
del inputs, outputs
torch.cuda.empty_cache()
with open(OUTPUT_FILE, 'w', encoding='utf-8') as f:
json.dump(final_results, f, indent=4, ensure_ascii=False)
# 6. 保存最终结果
print("\n\n=== ALL BATCHES COMPLETE ===")
# 保存最终结果到 JSON 文件
with open(OUTPUT_FILE, 'w', encoding='utf-8') as f:
json.dump(final_results, f, indent=4, ensure_ascii=False)
print(f"Results saved to {OUTPUT_FILE}")
except Exception as e:
print(f"\n--- Execution Error ---")
print(f"An unexpected error occurred: {e}")

178
data_ingestion/run_ingestion.py Normal file
View File

@@ -0,0 +1,178 @@
import chromadb
import os
import json
from copy import deepcopy
import torch
from tqdm import tqdm
from PIL import Image
from transformers import CLIPProcessor, CLIPModel
from app.taxonomy import CATEGORY, ALL_CATEGORY, OCCASION
BATCH_SOURCE = '2025_q4'
DATA_DIR = f'./data/{BATCH_SOURCE}'
IMAGE_DIR = f'./data/{BATCH_SOURCE}/image_data'
RAW_DATA_PATH = f'{DATA_DIR}/products-all.json'
CATEGORIZED_METADATA_PATH = f'{DATA_DIR}/metadata_extraction.json'
## Load data
with open(RAW_DATA_PATH, 'r', encoding='utf-8') as file:
raw_data = json.load(file)
with open(CATEGORIZED_METADATA_PATH, 'r', encoding='utf-8') as file:
categorized_data = json.load(file)
# Create Collection
client = chromadb.PersistentClient(path='./data/db')
collection = client.get_or_create_collection(
name="lc_clothing_embedding"
)
# if you wish to delete some item, uncomment following
# results = collection.delete(
# where={
# "batch_source": BATCH_SOURCE
# }
# )
# Load model
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
def format_product_info(product):
tags_str = ", ".join(product.get('tags', []))
info = (
f"Product Name: {product.get('name', 'N/A')}\n"
f"Brand: {product.get('brand', 'N/A')}\n"
f"Category: {product.get('category', 'N/A')} / {product.get('deptName', 'N/A')}\n"
f"Color: {product.get('color', 'N/A')}\n"
f"Description: {product.get('description', '')}\n"
f"Tags: {tags_str}"
f"GroupName: {product.get('groupName', 'N/A')}\n"
f"DetpName: {product.get('deptName', 'N/A')}\n"
f"OnlineBU: {product.get('onlineBU', 'N/A')}\n"
)
return info
# Combine all data together
new_category = {}
valid_count = 0
all_count = 0
for raw_item in tqdm(raw_data['products']):
item_id = raw_item.get('id')
if not item_id:
print(f"This item {raw_item} did not have a valid item_id")
continue
raw_category = raw_item.get("category")
if raw_category not in ['Clothing', 'Accessories', 'Shoes', 'Bags', 'Fine Jewellery And Watches']:
continue
image_path = os.path.join(IMAGE_DIR, f"{item_id}.jpg")
if not os.path.exists(image_path):
print(f"Image not found: {image_path}")
continue
# All above is raw data error, it's not our business.
all_count += 1
processed_item = categorized_data.get(item_id, {})
if not processed_item:
print(f"{item_id} has not been categorized. It does not exist in {CATEGORIZED_METADATA_PATH}")
continue
category = processed_item.get("category")
gender = processed_item.get("gender")
applicable_occasions = processed_item.get("applicable_occasions", [])
inappropriate_occasions = processed_item.get("inappropriate_occasions", [])
if category not in ALL_CATEGORY:
print(f"{item_id}'s category, {category}, does not valid.")
if category not in new_category:
new_category[category] = [item_id]
else:
new_category[category].append(item_id)
continue
if gender not in ['female', 'male', 'unisex']:
print(f"{item_id}'s gender is not valid in {['female', 'male', 'unisex']}")
continue
occasions = applicable_occasions + inappropriate_occasions
if not set(occasions).issubset(set(OCCASION)):
# print(f"{item_id}'s some occasions is not vaild. \n Invalid occasion is {set(occasions).difference(set(OCCASION))}")
applicable_occasions = [o for o in applicable_occasions if o in OCCASION]
inappropriate_occasions = [o for o in inappropriate_occasions if o in OCCASION]
description = raw_item.get('description', "")
if not description:
f"{item_id}'s description is lost."
continue
url = raw_item.get('url', '')
if not url:
f"{item_id}'s url is lost."
continue
valid_count += 1
# Prepare metadata for db
item_img_metadata = {
"item_id": item_id,
"category": category,
"description": description,
"gender": gender,
'brand': raw_item.get('brand', ''),
'color': raw_item.get('color', ''),
'price': raw_item.get('price', ''),
'tags': ",".join(raw_item.get('tags', [])),
'url': url,
"modality": "image",
"batch_source": BATCH_SOURCE
}
for occasion in OCCASION:
item_img_metadata[occasion] = 0
for occasion in applicable_occasions:
item_img_metadata[occasion] = 1
for occasion in inappropriate_occasions:
item_img_metadata[occasion] = -1
item_txt_metadata = deepcopy(item_img_metadata)
item_txt_metadata["modality"] = "text"
# Get image feature
image = Image.open(image_path).convert("RGB")
inputs = processor(images=image, return_tensors="pt").to(device)
with torch.no_grad():
img_features = model.get_image_features(**inputs)
img_features = img_features / img_features.norm(p=2, dim=-1, keepdim=True)
img_embedding = img_features.cpu().numpy().flatten().tolist()
# Get text feature
inputs = processor(text=[description], return_tensors="pt", padding=True, truncation=True).to(device)
with torch.no_grad():
txt_features = model.get_text_features(**inputs)
txt_features = txt_features / txt_features.norm(p=2, dim=-1, keepdim=True)
txt_embedding = txt_features.cpu().numpy().flatten().tolist()
product_info = format_product_info(raw_item)
# 插入到 ChromaDB
collection.add(
ids=[f'{item_id}_img', f'{item_id}_txt'],
documents=[product_info, product_info],
embeddings=[img_embedding, txt_embedding],
metadatas=[item_img_metadata, item_txt_metadata],
)
print(f"Final valid ratio is {valid_count / all_count * 100}%. Total number is {all_count}, Valid number is {valid_count}")
print(f'Found new category for consideration: {new_category}')

BIN
docs/Edi.docx Normal file
View File

Binary file not shown.

263
docs/LC Recommendation Workflow.drawio Normal file
View File

File diff suppressed because one or more lines are too long

BIN
docs/LC Recommendation Workflow.pdf Normal file
View File

Binary file not shown.

BIN
docs/LC Stylist Rules 总结.docx Normal file
View File

Binary file not shown.

BIN
docs/vera.docx Normal file
View File

Binary file not shown.