kwslib 0.0.4

Python client library for KWS Platform API - dataset splits, feature npz download, model/artifacts/metrics push

KWS Library (kwslib)

Thư viện Python dùng API từ backend KWS_Server, dùng trên Google Colab, Jupyter Notebook hoặc script Python.

Single Source of Truth (Operations): Tất cả tài liệu vận hành/quy trình/template của dự án nằm tại Workspace/ (entry: Workspace/README.md). Repo này chỉ giữ tài liệu kỹ thuật cho component.

Ba công việc chính (KWS_Lib ↔ KWS_Server)

KWS_Lib là thư viện hỗ trợ train model trên Google Colab, Jupyter Notebook hoặc script; mọi dữ liệu và metadata đều qua KWS_Server (DB + MinIO).

Công việc	Mô tả	API / High-level
1. Tạo dataset_split (train, val, test)	Lấy danh sách file feature (MFCC/npz) → chia train/val/test (pandas, sklearn) → đẩy lên DB.	dataset_splits.get_mfcc_files() → chia → dataset_splits.create_split_from_list(); hoặc DatasetPipeline.get_data() → split_and_push().
2. Tải npz (feature extraction từ MinIO)	Tải file .npz của split đã tạo (server đọc từ MinIO, stream qua API).	dataset_splits.download(split_id, output_path) (ZIP npz); DatasetSplitFilesClient.download_all_npz(split_id, output_dir); list_files(split_id, file_type="npz") + download_file(...).
3. Đẩy thông tin model (config, artifacts, metrics)	Đăng ký run, upload artifact (file model lên MinIO + DB), POST metrics.	ModelManager.register_run(), push_artifact(), push_metrics(); hoặc experiments.create_run(), artifacts.upload(), metrics.create(payload=...).

Chuẩn metrics khi POST: bắt buộc Accuracy, Precision, Recall, F1-Score, Confusion Matrix (dùng build_metrics_payload hoặc metrics_from_sklearn).

Features

• API Coverage: Bọc toàn bộ API backend (datasets, dataset_splits, models, experiments, metrics, mlflow)
• Data splitting: get_mfcc_files, create_split_from_list, script create_dataset_split.py + pandas
• Chuẩn metrics: build_metrics_payload, metrics_from_sklearn → payload đúng chuẩn để POST /api/v1/metrics
• MinIO qua API: Tải .wav / .npz qua API (stream), không cần kết nối trực tiếp MinIO
• Telegram: Optional thông báo khi job xong

Installation (PyPI)

pip install kwslib

Optional: TensorFlow (cho training CNN):

pip install kwslib[tensorflow]

From source:

git clone <repository>
cd KWS_Lib
pip install -e .

Quick Start

Smoke check (no login)

Verify KWS_Server is reachable:

python -m kwslib.smoke

Override base URL:

KWS_SERVER_URL=http://127.0.0.1:8000 python -m kwslib.smoke

Basic Usage

from kwslib import KWSClient

# Initialize client
client = KWSClient(base_url="http://localhost:8000")

# Login
client.login(username="admin", password="password")

# List datasets
datasets = client.datasets.list()
print(f"Found {datasets['total']} datasets")

# Get dataset details
dataset = client.datasets.get(dataset_id=1)
print(f"Dataset: {dataset['name']}")

Download Dataset Split Files for Training

from kwslib import KWSClient, DatasetSplitFilesClient

# Initialize API client
api = KWSClient(base_url="http://localhost:8000")
api.login(username="admin", password="password")

# Initialize files client (uses API, no direct MinIO connection)
files_client = DatasetSplitFilesClient(api)

# List all files in split
files_info = files_client.list_files(split_id=1, file_type="npz")
print(f"Found {files_info['total_files']} files")

# Download all .npz files
files_client.download_all_npz(
    split_id=1,
    output_dir="features"
)

# Download all .wav files
files_client.download_all_wav(
    split_id=1,
    output_dir="audio"
)

# Or download as ZIP
files_client.download_all_files_zip(
    split_id=1,
    file_type="npz",
    output_path="features.zip"
)

# Get presigned URLs (for Google Colab)
urls = files_client.get_file_urls(split_id=1, file_type="npz")
for file_info in urls["files"]:
    print(f"{file_info['file_name']}: {file_info['url']}")

With Telegram Notifications

from kwslib import KWSClient, TelegramNotifier

# Initialize
client = KWSClient(base_url="http://localhost:8000")
client.login(username="admin", password="password")

notifier = TelegramNotifier(
    bot_token="YOUR_BOT_TOKEN",
    chat_id="YOUR_CHAT_ID"
)

# Create experiment run (triggers background job)
run = client.experiments.create_run(
    experiment_id=1,
    model_id=1,
    dataset_split_id=1,
    config={"learning_rate": 0.001, "batch_size": 32},
    git_commit="manual-training",
)

# Wait for completion
job_id = run.get("job_id")
status = client.jobs.wait_for_completion(job_id)

# Send notification
if status["status"] == "completed":
    notifier.send(f"Training completed! Results: {status['result']}")
else:
    notifier.send(f"Training failed: {status.get('error')}")

API Modules

Authentication

• client.auth.login() - Login
• client.auth.logout() - Logout
• client.auth.get_me() - Get current user info

Datasets

• client.datasets.list() - List datasets
• client.datasets.get() - Get dataset
• client.datasets.create() - Create dataset
• client.datasets.update() - Update dataset
• client.datasets.delete() - Delete dataset
• client.datasets.list_versions() - List versions
• client.datasets.create_version() - Create version

Models

• client.models.list() - List models
• client.models.get() - Get model
• client.models.create() - Create model
• client.models.list_model_inits() - List model architectures

Experiments

• client.experiments.list() - List experiments
• client.experiments.create() - Create experiment
• client.experiments.create_run(experiment_id, model_id, dataset_split_id, config, git_commit) - Tạo run (background job)
• client.experiments.list_runs(experiment_id) - List runs của một experiment
• client.experiments.list_runs_global(experiment_id=..., model_id=...) - List tất cả runs (có lọc)
• client.experiments.get_run(experiment_id, run_id) - Chi tiết run

Dataset Splits

• client.dataset_splits.list(dataset_version_id=..., config_name=..., name=...) - List splits
• client.dataset_splits.get(split_id) - Lấy một split theo ID
• client.dataset_splits.create(dataset_version_id, name, config_name) - Tạo bản ghi split (chỉ metadata)
• client.dataset_splits.create_split_from_list(...) - Tạo split từ danh sách file (sau khi chia bằng pandas)
• client.dataset_splits.get_mfcc_files(dataset_version_id, ...) - Lấy danh sách file MFCC (DataFrame) để chia
• client.dataset_splits.download(split_id, output_path) - Tải split dạng ZIP (npz)
• client.dataset_splits.generate(split_id) - Trigger job generate split
• client.dataset_splits.list_files(split_id, file_type) - List file .wav / .npz trong split

Metrics (chuẩn: Accuracy, Precision, Recall, F1-Score, Confusion Matrix)

• client.metrics.list(model_id=..., dataset_split_id=...) - List metrics
• client.metrics.get(metric_id) - Chi tiết metric
• client.metrics.create(payload=payload) - POST metric (dùng payload từ build_metrics_payload hoặc metrics_from_sklearn)
• client.metrics.compare_metrics(model_ids=[1,2,3], split_id=...) - So sánh metrics giữa các model

Audio

• client.audio.list_keyword_samples() - List keyword audio
• client.audio.upload_keyword_sample() - Upload audio
• client.audio.get_keyword_sample_url() - Get presigned URL

Features

• client.features.get_keyword_features() - Get features
• client.features.extract_keyword_features() - Extract features

Jobs

• client.jobs.get() - Get job status
• client.jobs.list() - List jobs
• client.jobs.wait_for_completion() - Wait for job completion

Dataset Split Files Client

• files_client.list_files() - List all files in split
• files_client.download_wav() - Download a .wav file
• files_client.download_npz() - Download and load a .npz file
• files_client.download_all_wav() - Download all .wav files
• files_client.download_all_npz() - Download all .npz files
• files_client.download_all_files_zip() - Download all files as ZIP
• files_client.get_file_urls() - Get presigned URLs for all files

Telegram Notifier

• notifier.send() - Send message
• notifier.send_file() - Send file
• notifier.send_photo() - Send photo

Examples

Data splitting (pandas + create_split_from_list)

from kwslib import KWSClient
from create_dataset_split import get_split_data, push_splits
from sklearn.model_selection import train_test_split

api = KWSClient(base_url="http://localhost:8000")
api.login(username="admin", password="password")

# 1. Lấy danh sách file MFCC
df = get_split_data(api=api, dataset_version_id=48, feature_type_id=2)

# 2. Chia train/test (stratify theo label)
train_df, test_df = train_test_split(
    df, train_size=0.8, test_size=0.2, random_state=42, stratify=df["derivative_label"]
)

# 3. Đẩy lên DB (tạo splits + gán file)
created = push_splits(
    api=api,
    dataset_version_id=48,
    config_name="config_80_20",
    splits={"train": train_df, "test": test_df},
)
# created = {"train": 123, "test": 124}

Metrics chuẩn (Accuracy, Precision, Recall, F1-Score, Confusion Matrix)

from kwslib import KWSClient, build_metrics_payload, metrics_from_sklearn
import numpy as np

api = KWSClient(base_url="http://localhost:8000")
api.login(username="admin", password="password")

# Cách 1: Từ y_true, y_pred (sklearn)
y_true = np.array([0, 1, 1, 0])
y_pred = np.array([0, 1, 0, 0])
payload = metrics_from_sklearn(
    y_true, y_pred,
    model_id=1, dataset_split_id=1, experiment_run_id=1,
    average="weighted",
)
api.metrics.create(payload=payload)

# Cách 2: Từ dict metrics đã tính
metrics = {
    "accuracy": 0.92,
    "precision": 0.91,
    "recall": 0.90,
    "f1_score": 0.905,
    "confusion_matrix": [[50, 2], [3, 45]],  # 2D list int
}
payload = build_metrics_payload(
    model_id=1, dataset_split_id=1, experiment_run_id=1,
    metrics=metrics,
)
api.metrics.create(payload=payload)

# So sánh nhiều model trên một split
comparison = api.metrics.compare_metrics(model_ids=[1, 2, 3], split_id=1)

Complete training workflow (tóm tắt)

# 1. Tạo split (metadata) hoặc dùng create_split_from_list sau khi chia pandas
split = api.dataset_splits.create(dataset_version_id=1, name="train", config_name="config_70_15_15")
# Hoặc: push_splits(api, dataset_version_id, config_name, splits={"train": train_df, "val": val_df, "test": test_df})

# 2. Generate split (job) nếu cần
# job = api.dataset_splits.generate(split_id)

# 3. Tạo experiment run (background job)
run = api.experiments.create_run(
    experiment_id=1,
    model_id=1,
    dataset_split_id=split_id,
    config={"learning_rate": 0.001, "batch_size": 32},
    git_commit="manual-training",
)

# 4. Sau khi train, POST metrics (chuẩn: accuracy, precision, recall, f1_score, confusion_matrix)
# run_id = ID experiment run (lấy từ list_runs sau khi job create_run hoàn thành)
payload = metrics_from_sklearn(y_true, y_pred, model_id=1, dataset_split_id=split_id, experiment_run_id=run_id)
api.metrics.create(payload=payload)

Google Colab Usage

# In Google Colab, use presigned URLs for direct download
from kwslib import KWSClient, DatasetSplitFilesClient

api = KWSClient(base_url="https://your-api.com")
api.login(username="admin", password="password")

files_client = DatasetSplitFilesClient(api)

# Get presigned URLs
urls = files_client.get_file_urls(split_id=1, file_type="npz")

# Download in Colab
import urllib.request
for file_info in urls["files"]:
    urllib.request.urlretrieve(
        file_info["url"],
        f"/content/{file_info['file_name']}"
    )

Configuration

Environment Variables

You can set default values using environment variables:

export KWS_BASE_URL="http://localhost:8000"
export KWS_USERNAME="admin"
export KWS_PASSWORD="password"
export MINIO_ENDPOINT="localhost:9000"
export MINIO_ACCESS_KEY="minioadmin"
export MINIO_SECRET_KEY="minioadmin"

Publishing to PyPI

pip install build twine
python -m build
twine upload dist/*

Đảm bảo đã tăng version trong pyproject.toml và setup.py trước khi build.

License

MIT License

Contributing

Contributions are welcome! Please open an issue or submit a pull request.