FakeRockert543/ckip-coreml

# 取得原始碼
git clone https://github.com/FakeRocket543/ckip-coreml.git
cd ckip-coreml

# 建立虛擬環境（需要 Python 3.13）
python3.13 -m venv .venv && source .venv/bin/activate

# 安裝依賴
pip install coremltools numpy huggingface_hub

# 從 HuggingFace 下載全部 .mlpackage
huggingface-cli download FakeRockert543/ckip-coreml --local-dir .

ckip-coreml/
├── ckip_ws_fp16.mlpackage      # 斷詞 fp16（推薦）
├── ckip_ws_fp32.mlpackage
├── ckip_ws_q8.mlpackage
├── ckip_pos_fp16.mlpackage     # 詞性 fp16
├── ckip_pos_fp32.mlpackage
├── ckip_pos_q8.mlpackage
├── ckip_ner_fp16.mlpackage     # 實體 fp16
├── ckip_ner_fp32.mlpackage
├── ckip_ner_q8.mlpackage
├── ckip_ws.mlpackage           # 原始版本 (=fp32)
├── ckip_pos.mlpackage
└── ckip_ner.mlpackage

# 方法一：從 MLX repo 下載 vocab.txt
huggingface-cli download FakeRockert543/ckip-mlx models/vocab.txt --local-dir .
mv models/vocab.txt vocab.txt && rm -rf models

# 方法二：從原始 BERT 下載
# wget https://huggingface.co/bert-base-chinese/resolve/main/vocab.txt

import coremltools as ct
import numpy as np

# 載入詞表
vocab = {}
with open("vocab.txt") as f:
    for i, line in enumerate(f):
        vocab[line.strip()] = i

# 載入模型（推薦 fp16）
model = ct.models.MLModel("ckip_ws_fp16.mlpackage")

# Tokenize（BERT 單字切分）
text = "台積電今天股價上漲三十元"
ids = [101] + [vocab.get(ch, 100) for ch in text] + [102]  # 101=[CLS], 102=[SEP], 100=[UNK]
input_ids = np.array([ids], dtype=np.int32)
attention_mask = np.ones_like(input_ids, dtype=np.int32)

# 推論
out = model.predict({"input_ids": input_ids, "attention_mask": attention_mask})
preds = np.argmax(out["logits"], axis=-1)[0]

# 解碼斷詞結果（B=0: 詞首, I=1: 詞中）
words, cur = [], ""
for i, ch in enumerate(text):
    p = preds[i + 1]  # +1 跳過 [CLS]
    if p == 0 and cur:
        words.append(cur)
        cur = ch
    else:
        cur += ch
if cur:
    words.append(cur)

print(words)
# ['台積電', '今天', '股價', '上漲', '三十', '元']

import json

pos_model = ct.models.MLModel("ckip_pos_fp16.mlpackage")
out = pos_model.predict({"input_ids": input_ids, "attention_mask": attention_mask})
preds = np.argmax(out["logits"], axis=-1)[0]

# POS id2label 對照表（從 MLX config.json 取得，或用以下常見標籤）
# 完整對照表見 GitHub repo 的 models/pos/config.json
for i, ch in enumerate(text):
    print(f"{ch} → label_id={preds[i + 1]}")

import CoreML

// 載入模型
let config = MLModelConfiguration()
config.computeUnits = .all  // 使用 ANE + GPU + CPU
let model = try MLModel(contentsOf: modelURL, configuration: config)

// 準備輸入
let inputIds = try MLMultiArray(shape: [1, seqLen as NSNumber], dataType: .int32)
let attentionMask = try MLMultiArray(shape: [1, seqLen as NSNumber], dataType: .int32)

// 填入 token IDs（[CLS] + 單字 IDs + [SEP]）
for (i, id) in tokenIds.enumerated() {
    inputIds[i] = NSNumber(value: id)
    attentionMask[i] = 1
}

// 推論
let input = try MLDictionaryFeatureProvider(dictionary: [
    "input_ids": inputIds,
    "attention_mask": attentionMask
])
let output = try model.prediction(from: input)
let logits = output.featureValue(for: "logits")!.multiArrayValue!

// 取 argmax 得到預測標籤