data4fun | 小数据不简单修改

!python -m pip install paddlepaddle-gpu==3.1.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/

# 检验是否安装成功
!python -c "import paddle;paddle.utils.run_check()"

!git clone https://github.com/PaddlePaddle/ERNIE.git

%cd ERNIE-develop
!python -m pip install -r requirements/gpu/requirements.txt 
!python -m pip install -e .

!python -m pip install fastdeploy-gpu -i https://www.paddlepaddle.org.cn/packages/stable/fastdeploy-gpu-80_90/ --extra-index-url https://mirrors.tuna.tsinghua.edu.cn/pypi/web/simple

!pip install --upgrade aistudio-sdk
# 使用aistudio cli下载模型（推荐）
!aistudio download --model PaddlePaddle/ERNIE-4.5-0.3B-Paddle --local_dir baidu/ERNIE-4.5-0.3B-Paddle

### data
train_dataset_type: "erniekit"
eval_dataset_type: "erniekit"
train_dataset_path: "./examples/data/finetune_train_dataset.jsonl"
train_dataset_prob: "1.0"
eval_dataset_path: "./examples/data/finetune_eval_dataset.jsonl"
eval_dataset_prob: "1.0"
max_seq_len: 8192
num_samples_each_epoch: 6000000

### model
model_name_or_path: baidu/ERNIE-4.5-0.3B-Paddle
fine_tuning: LoRA
fuse_rope: True
use_sparse_head_and_loss_fn: True

### finetuning
# base
stage: SFT
seed: 23
do_train: True
do_eval: True
distributed_dataloader: False
dataloader_num_workers: 1
batch_size: 1
num_train_epochs: 1
max_steps: 100
max_evaluate_steps: 10000
eval_steps: 10000
evaluation_strategy: steps
save_steps: 10000000
save_total_limit: 5
save_strategy: steps
logging_steps: 1
release_grads: True
gradient_accumulation_steps: 8
logging_dir: ./vdl_log
output_dir: ./output
disable_tqdm: True

# train
warmup_steps: 20
learning_rate: 1.0e-5
lr_scheduler_type: cosine
min_lr: 1.0e-6
layerwise_lr_decay_bound: 1.0

# optimizer
weight_decay: 0.1
adam_epsilon: 1.0e-8
adam_beta1: 0.9
adam_beta2: 0.95
offload_optim: True

# performance
tensor_parallel_degree: 1
pipeline_parallel_degree: 1
sharding_parallel_degree: 1
sharding: stage1
sequence_parallel: True
pipeline_parallel_config: enable_delay_scale_loss enable_release_grads disable_partial_send_recv
recompute: False
recompute_use_reentrant: True
compute_type: bf16
fp16_opt_level: O2
disable_ckpt_quant: True
amp_master_grad: True
amp_custom_white_list:
 - lookup_table
 - lookup_table_v2
 - flash_attn
 - matmul
 - matmul_v2
 - fused_gemm_epilogue
amp_custom_black_list:
 - reduce_sum
 - softmax_with_cross_entropy
 - c_softmax_with_cross_entropy
 - elementwise_div
 - sin
 - cos
unified_checkpoint: True
unified_checkpoint_config: async_save

!erniekit train examples/configs/ERNIE-4.5-0.3B/sft/run_sft_lora_8k.yaml
!erniekit export examples/configs/ERNIE-4.5-0.3B/run_export.yaml lora=True

!erniekit server examples/configs/ERNIE-4.5-0.3B/run_chat.yaml
!erniekit chat examples/configs/ERNIE-4.5-0.3B/run_chat.yaml

from sentence_transformers import SentenceTransformer
import random
import glob,os
from PIL import Image
import matplotlib.pyplot as plt
#图片目录
image_folder = 'xxx/xx'
# 使用CLIP模型构建特征
def generate_clip_embeddings(images_path, model):

 image_paths = glob.glob(os.path.join(images_path, '**/*.jpg'), recursive=True)

 embeddings = []
 for img_path in image_paths:
 image = Image.open(img_path)
 embedding = model.encode(image)
 embeddings.append(embedding)

 return embeddings, image_paths

model = SentenceTransformer('clip-ViT-B-32')
embeddings, image_paths = generate_clip_embeddings(image_folder, model)

import faiss
import numpy as np


# FAISS 构建索引
features = np.array(embeddings).astype(np.float32)
faiss.normalize_L2(features)
vector_dim = features.shape[1]

index = faiss.IndexFlatIP(vector_dim)
index.add(features)

# 图片查询：选一张图（或上传图）再计算 embedding 作为查询向量
# query_image = Image.open('xxxx/feiji.jpeg')
# query_vector = model.encode(query_image)
# query_vector = np.array(query_vector).reshape(1, -1).astype(np.float32)
# faiss.normalize_L2(query_vector)

# 文本查询
query_text = "一辆自行车"
query_vector = model.encode(query_text) # 使用同一个模型对文本编码
query_vector = np.array(query_vector).reshape(1, -1).astype(np.float32)
faiss.normalize_L2(query_vector)

# 搜索最相似的图片
top_k = 5
distances, ann = index.search(query_vector, k=top_k)

# 显示搜索结果
for i, idx in enumerate(ann[0]):
 print(f"Top {i+1} match: {image_paths[idx]} (Score: {distances[0][i]:.4f})")

from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage, HumanMessage

llm = ChatOpenAI(
 api_key=env_settings.OPENAI_API_KEY,
 model_name="gpt-3.5-turbo",
 base_url='https://burn.hair/v1'
)

ai_message = llm.invoke([
 SystemMessage(content="please speak in chinese"),
 HumanMessage(content="hello"),
])

print(ai_message.content)
#输出：你好！有什么可以帮助你的吗？

from langchain_core.messages import SystemMessage, HumanMessage
from langchain_deepseek import ChatDeepSeek

llm = ChatDeepSeek(
 api_key=env_settings.DS_API_KEY,
 model="deepseek-ai/DeepSeek-R1",
 temperature=0,
 max_tokens=None,
 timeout=None,
 max_retries=2,
 api_base='https://api.siliconflow.cn'
)
ai_message = llm.invoke([
 SystemMessage(content="please speak in chinese"),
 HumanMessage(content="hello"),
])

print(ai_message.content)
# 你好！😊 有什么可以帮您的吗？