2024 구글 클라우드 제미나이(Gemini) 스터디잼 참가후기

# PROJECT_ID, REGION 환경변수 설정
PROJECT_ID=$(gcloud config get-value project)
REGION=lab region
echo "PROJECT_ID=${PROJECT_ID}"
echo "REGION=${REGION}"

# 구글 사용자 계정 환경변수 설정
USER=$(gcloud config get-value account 2> /dev/null)
echo "USER=${USER}"

# Cloud AI Companion API for Gemini 활성화
gcloud services enable cloudaicompanion.googleapis.com --project ${PROJECT_ID}

#  사용자 계정에 필요한 IAM 역할 부여
gcloud projects add-iam-policy-binding ${PROJECT_ID} --member user:${USER} --role=roles/cloudaicompanion.user
gcloud projects add-iam-policy-binding ${PROJECT_ID} --member user:${USER} --role=roles/serviceusage.serviceUsageViewer

# select the sum of sale_price by Date(created_at) and product_id casted to day from bigquery-public-data.thelook_ecommerce.order_id as t1 joined this with products table in the same dataset as t2
SELECT
    SUM(sale_price),
    DATE(created_at) AS created_at_day,
    CAST(product_id as INT64)
FROM `bigquery-public-data.thelook_ecommerce.order_items` AS t1
JOIN `bigquery-public-data.thelook_ecommerce.products` AS t2 ON t1.product_id = t2.id
GROUP BY created_at_day, product_id

-- Forecasting ML model 생성하기
CREATE MODEL bqml_tutorial.sales_forecasting_model
OPTIONS(
    MODEL_TYPE='ARIMA_PLUS',
    time_series_timestamp_col='date_col',
    time_series_data_col='total_sales',
    time_series_id_col='product_id') AS
SELECT sum(sale_price) as total_sales, DATE(created_at) as date_col, product_id
FROM `bigquery-public-data.thelook_ecommerce.order_items` AS t1
INNER JOIN `bigquery-public-data.thelook_ecommerce.products` AS t2 ON t1.product_id = t2.id
GROUP BY 2, 3;

-- Forecasting ML model 호출하기
SELECT *
FROM ML.FORECAST(
    MODEL `PROJECT_ID.DATASET_ID.MODEL_NAME`,
    STRUCT(
        7 AS horizon,
        0.95 AS confidence_level
    )
)

-- bigquery
CREATE OR REPLACE TABLE ecommerce.customer_stats AS
SELECT
  user_id,
  DATE_DIFF(CURRENT_DATE(), CAST(MAX(order_created_date) AS DATE), day) AS days_since_last_order, ---RECENCY
  COUNT(order_id) AS count_orders, --FREQUENCY
  AVG(sale_price) AS average_spend --MONETARY
  FROM (
      SELECT
        user_id,
        order_id,
        sale_price,
        created_at AS order_created_date
        FROM `bigquery-public-data.thelook_ecommerce.order_items`
        WHERE created_at BETWEEN '2022-01-01' AND '2023-01-01'
  )
GROUP BY user_id;

import bigframes.pandas as bpd
from bigframes.ml.cluster import KMeans
from bigframes.ml.model_selection import train_test_split
import matplotlib.pyplot as plt

dataset_name = "ecommerce"
model_name = "customer_segmentation_model"
table_name = "customer_stats"

# Load dataset
df = bpd.read_gbq(f"{project_id}.{dataset_name}.{table_name}")

# Train
df_train, df_test = train_test_split(df, test_size=0.2,  random_state=42)
model = KMeans(n_clusters=5)
model.fit(df_train)
model.to_gbq(f"{project_id}.{dataset_name}.{model_name}")

# Predict
predictions_df = model.predict(df)

# Visualize
plt.scatter(predictions_df['days_since_last_order'], predictions_df['average_spend'], c=predictions_df['CENTROID_ID'])
plt.xlabel("days_since_last_order")
plt.ylabel("average_spend")
plt.title("Attribute grouped by K-means Cluster")
plt.show()

# 클러스터의 중심을 기반으로 각 클러스터의 평균 지출, 주문 수, 최근 주문일수의 특성값을 보여줌.
query = f"""
SELECT
    CONCAT('cluster ', CAST(centroid_id as STRING)) as centroid,
    average_spend,
    count_orders,
    days_since_last_order
FROM (
    SELECT centroid_id, feature, ROUND(numerical_value, 2) as value
    FROM ML.CENTROIDS(MODEL `{dataset_name}.{model_name}`)
)
PIVOT (
    SUM(value)
    FOR feature IN ('average_spend',  'count_orders', 'days_since_last_order')
)
ORDER BY centroid_id
"""

client = bigquery.Client(project=project_id)
df_query = client.query(query).to_dataframe()
df_query.to_string(header=False, index=False)

cluster_info = []
for i, row in df_query.iterrows():
    cluster_info.append(f"{row["centroid"]}, average spend ${row["average_spend"]}, count of orders per person {row["count_orders"]}, days since last order {row["days_since_last_order"]}")

cluster_info = (str.join("\n", cluster_info))
print(cluster_info)
# cluster 1, average spend $48.32, count of orders per person 1.36, days since last order 384.37
# cluster 2, average spend $202.34, count of orders per person 1.3, days since last order 482.62
# cluster 3, average spend $45.68, count of orders per person 1.36, days since last order 585.4
# cluster 4, average spend $44.71, count of orders per person 1.36, days since last order 466.26
# cluster 5, average spend $58.08, count of orders per person 3.92, days since last order 427.36

# 마케팅 문구 생성
prompt = f"""
You're a creative brand strategist, given the following clusters, come up with \
creative brand persona, a catchy title, and next marketing action, \
explained step by step.

Clusters:
{cluster_info}

For each Cluster:
* Title:
* Persona:
* Next marketing step:
"""
from vertexai.language_models._language_models import TextGenerationModel

model = TextGenerationModel.from_pretrained("text-bison")
response = model.predict(prompt, max_output_tokens=1024, temperature=0.4)
print(response.text)

extract_log_data = FunctionDeclaration(
    name="extract_log_data",
    description="Extract details from error messages in raw log data",
    parameters={
        "type": "object",
        "properties": {
            "locations": {
                "type": "array",
                "description": "Errors",
                "items": {
                    "description": "Details of the error",
                    "type": "object",
                    "properties": {
                        "error_message": {
                            "type": "string",
                            "description": "Full error message",
                        },
                        "error_code": {"type": "string", "description": "Error code"},
                        "error_type": {"type": "string", "description": "Error type"},
                    },
                },
            }
        },
    },
)

extraction_tool = Tool(
    function_declarations=[extract_log_data],
)

prompt = """
[15:43:28] ERROR: Could not process image upload: Unsupported file format. (Error Code: 308)
[15:44:10] INFO: Search index updated successfully.
[15:45:02] ERROR: Service dependency unavailable (payment gateway). Retrying... (Error Code: 5522)
[15:45:33] ERROR: Application crashed due to out-of-memory exception. (Error Code: 9001)
"""

response = model.generate_content(
    prompt,
    generation_config=GenerationConfig(temperature=0),
    tools=[extraction_tool],
)

response.candidates[0].content.parts[0].function_call.args

# 아래의 형태의 데이터를 얻을 수 있다.
# {
#     "locations": [
#         {
#             "error_message": "File not found: /path/to/file.txt",
#             "error_code": "404",
#             "error_type": "FileError"
#         },
#         {
#             "error_message": "Access denied for user 'admin'",
#             "error_code": "403",
#             "error_type": "PermissionError"
#         },
#         {
#             "error_message": "Unexpected token in JSON at position 15",
#             "error_code": "500",
#             "error_type": "SyntaxError"
#         }
#     ]
# }

def extract_log_data(locations: list):
    # 외부 API 로직 추가
    return location

args = response.candidates[0].content.parts[0].function_call.args

response = chat.send_message(
    Part.from_function_response(
        name="extract_log_data",
        response={
            "content": extract_log_data(args),
        },
    ),
)

# Download an image from Google Cloud Storage
! gsutil cp "gs://cloud-samples-data/generative-ai/image/320px-Felis_catus-cat_on_snow.jpg" ./image.jpg

# Load from local file and prepare content
image = Image.load_from_file("image.jpg")
prompt = "Describe this image?"
contents = [image, prompt]

responses = multimodal_model.generate_content(contents, stream=True)

print("-------Prompt--------")
print_multimodal_prompt(contents)

print("\n-------Response--------")
for response in responses:
    print(response.text, end="") #  This is a photo of a cat walking in the snow. The cat is brown and white, and the snow is white. The cat is looking at the camera.

# Load images from Cloud Storage URI
image1_url = "https://storage.googleapis.com/github-repo/img/gemini/intro/landmark1.jpg"
image2_url = "https://storage.googleapis.com/github-repo/img/gemini/intro/landmark2.jpg"
image3_url = "https://storage.googleapis.com/github-repo/img/gemini/intro/landmark3.jpg"
image1 = load_image_from_url(image1_url)
image2 = load_image_from_url(image2_url)
image3 = load_image_from_url(image3_url)

# Prepare prompts
prompt1 = """{"city": "London", "Landmark:", "Big Ben"}"""
prompt2 = """{"city": "Paris", "Landmark:", "Eiffel Tower"}"""

# Prepare contents
contents = [image1, prompt1, image2, prompt2, image3]

responses = multimodal_model.generate_content(contents, stream=True)

print("-------Prompt--------")
print_multimodal_prompt(contents)

print("\n-------Response--------")
for response in responses:
    print(response.text, end="")

file_path = "github-repo/img/gemini/multimodality_usecases_overview/pixel8.mp4"
video_uri = f"gs://{file_path}"
video_url = f"https://storage.googleapis.com/{file_path}"

IPython.display.Video(video_url, width=450)

prompt = """
Answer the following questions using the video only:
What is the profession of the main person?
What are the main features of the phone highlighted?
Which city was this recorded in?
Provide the answer JSON.
"""

video = Part.from_uri(video_uri, mime_type="video/mp4")
contents = [prompt, video]

responses = multimodal_model.generate_content(contents, stream=True)

for response in responses:
    print(response.text, end="")

# 응답
# {
#   "profession": "photographer",
#   "features": "Night Sight, Video Boost",
#   "city": "Tokyo"
# }

prompt = """Answer the following questions using the video only:
What is the profession of the main person?
What are the main features of the phone highlighted?
Which city was this recorded in?
What's the main mood of the video?
What's the main color of the video?
Is the person in the video a man or a woman?
What is the person doing in the video?
Provide the answer JSON."""

# 응답
# {
#   "person": {
#     "profession": "photographer",
#     "name": "Seeaka Shimada"
#   },
#   "phone": {
#     "features": [
#       "Video Boost",
#       "Night Sight"
#     ]
#   },
#   "city": "Tokyo",
#   "mood": "calm",
#   "color": "blue",
#   "gender": "female",
#   "actions": [
#     "walking",
#     "taking pictures",
#     "smiling"
#   ]
# }

# 잘 대답하지 못한 prompt는 다음과 같다.
# If there is a scene with a person riding a bicycle in the video, please let me know at what time it appears.
# If there is a scene with a car driving on the road in the video, please let me know at what time it appears.