批量推理作业¶

from snowflake.ml.registry import Registry

registry = Registry(session=session, database_name=DATABASE, schema_name=REGISTRY_SCHEMA)
mv = registry.get_model('my_model').version('my_version')  # returns ModelVersion

from snowflake.ml.model.batch import OutputSpec

# how to run a batch job
job = mv.run_batch(
    compute_pool = "my_compute_pool",
    X = session.table("my_table"),
    output_spec = OutputSpec(stage_location="@my_db.my_schema.my_stage/path/"),
)

job.wait() # Optional: Blocking until the job finishes

from snowflake.ml.jobs import list_jobs, delete_job, get_job

# view logs to troubleshoot
job.get_logs()

# cancel a job
job.cancel()

# list to see all jobs
list_jobs().show()

# get the handle of a job
job = get_job("my_db.my_schema.job_name")

# delete a job that you no longer wish to run
delete_job(job)

# providing input from a query
X = session.sql("SELECT id, feature_1, feature_2 FROM feature_table WHERE feature_1 > 100"),

# reading from parquet files
X = session.read.option("pattern",".*file.*\\.parquet")
    .parquet("@DB.SCHEMA.STAGE/some/path")
    .select(col("id1").alias("id"), col("feature_1"), col("feature_2"))).filter(col("feature_1") > 100)

# Process a list of files
# The file paths have to be in the form of a full stage path as below
data = [
    ["@DB.SCHEMA.STAGE/dataset/files/file1"],
    ["@DB.SCHEMA.STAGE/dataset/files/file2"],
    ["@DB.SCHEMA.STAGE/dataset/files/file3"],
]
column_names = ["image"]
X = session.create_dataframe(data, schema=column_names)

from snowflake.ml.utils.stage_file import list_stage_files

# get all files under a path
X = list_stage_files(session, "@db.schema.my_stage/path")

# get all files under a path ending with ".jpg"
X = list_stage_files(session, "@db.schema.my_stage/path", pattern=".*\\.jpg")

# get all files under a path ending with ".jpg" and return the datafram with a column_name "IMAGES"
X = list_stage_files(session, "@db.schema.my_stage/path", pattern=".*\\.jpg", column_name="IMAGES")

mv.run_batch(
    X,
    input_spec=InputSpec(
 # we need to provide column_handling in the InputSpec to perform the necessary conversion
 # FULL_STAGE_PATH: fully qualified path (@db.schema.stage/path) to a file
 # RAW_BYTES: download and convert the file from the stage path to bytes
        column_handling={
            "path": {"input_format": InputFormat.FULL_STAGE_PATH, "convert_to": FileEncoding.RAW_BYTES}
        }
    ),
    ...
)

mv.run_batch(
    ...
    output_spec = OutputSpec(stage_location="@db.schema.stage/path/"),
)

session.read.option("pattern", ".*\\.parquet").parquet("@db.schema.stage/output_path/")

from snowflake.ml.model.batch import InputSpec, OutputSpec

mv.run_batch(
    X=input_df,
    compute_pool="my_compute_pool",
    input_spec=InputSpec(
        params={"temperature": 0.9, "max_tokens": 512}
    ),
    output_spec=OutputSpec(stage_location="@my_db.my_schema.my_stage/path/"),
)

from transformers import pipeline
from snowflake.ml.model import custom_model
from snowflake.ml.model import target_platform
from snowflake.ml.model.batch import InputSpec, OutputSpec, FileEncoding, InputFormat
from snowflake.ml.model.model_signature import core

# first we must define the schema, we'll expect audio file input as base64 string
signature = core.ModelSignature(
    inputs=[
        core.FeatureSpec(name="audio", dtype=core.DataType.STRING),
    ],
    outputs=[
        core.FeatureGroupSpec(
            name="outputs",
            specs=[
                core.FeatureSpec(name="text", dtype=core.DataType.STRING),
                core.FeatureGroupSpec(
                    name="chunks",
                    specs=[
                        core.FeatureSpec(
                            name="timestamp", dtype=core.DataType.DOUBLE, shape=(2,)
                        ),
                        core.FeatureSpec(name="text", dtype=core.DataType.STRING),
                    ],
                    shape=(-1,),
                ),
            ],
        ),
    ],
)

# defining the custom model, we decode the input from base64 to bytes and
# use whisper to perform the transcription
class CustomTranscriber(custom_model.CustomModel):
    def __init__(self, context: custom_model.ModelContext) -> None:
        super().__init__(context)
        self.model = self.context.model_ref("my_model")

    @custom_model.inference_api
    def predict(self, df: pd.DataFrame) -> pd.DataFrame:
        import base64
        audio_b64_list = df["audio"].tolist()
        audio_bytes_list = [base64.b64decode(audio_b64) for audio_b64 in audio_b64_list]
        temp_res = [self.model(audio_bytes) for audio_bytes in audio_bytes_list]
        return pd.DataFrame({"outputs": temp_res})

# creating an instance of our transcriber for logging
transcriber = CustomTranscriber(
    custom_model.ModelContext(
        models={
            "my_model": pipeline(
                task="automatic-speech-recognition", model="openai/whisper-small"
            )
        }
    )
)

# log the model
mv = reg.log_model(
    transcriber,
    model_name="custom_transcriber",
    version_name="v1",
    signatures={"predict": signature},
)

# input dataframe
data = [
    ["@DB.SCHEMA.STAGE/dataset/audio/audio1.mp3"],
    ["@DB.SCHEMA.STAGE/dataset/audio/audio2.mp3"],
    ["@DB.SCHEMA.STAGE/dataset/audio/audio3.mp3"],
]
column_names = ["audio"] # This column was defined in the signature above
input_df = session.create_dataframe(data, schema=column_names)


job = mv.run_batch(
    X=input_df,
    compute_pool="my_compute_pool",
    output_spec=OutputSpec(stage_location="@my_db.my_schema.my_stage/path/"),
    input_spec=InputSpec(
# we need to provide column_handling in the InputSpec to perform the necessary conversion
# FULL_STAGE_PATH: fully qualified path (db.schema.stage/path) to a file
# BASE_64: download and convert the file from the stage path to base64 string
        column_handling={
            "audio": {"input_format": InputFormat.FULL_STAGE_PATH, "convert_to": FileEncoding.BASE64}
        }
    )
)

from transformers import pipeline
from snowflake.ml.model import target_platform
from snowflake.ml.model.batch import InputSpec, OutputSpec, FileEncoding, InputFormat

# supported Hugging Face tasks will have their signatures auto-inferred
classifier = pipeline(task="image-classification", model="google/vit-base-patch16-224")

# log the model
mv = reg.log_model(
    classifier,
    model_name="image_classifier",
    version_name="v1",
    target_platforms=target_platform.SNOWPARK_CONTAINER_SERVICES_ONLY,
    pip_requirements=[
        "pillow" # dependency for image classification
    ],
)

# input dataframe
data = [
    ["@DB.SCHEMA.STAGE/dataset/image/image1.mp3"],
    ["@DB.SCHEMA.STAGE/dataset/image/image2.mp3"],
    ["@DB.SCHEMA.STAGE/dataset/image/image3.mp3"],
]
# this column was defined in the auto-inferred signature
# you can view the signature by calling 'mv.show_functions()'
column_names = ["images"]
input_df = session.create_dataframe(data, schema=column_names)

mv.run_batch(
    X=input_df,
    compute_pool="my_compute_pool",
    output_spec=OutputSpec(stage_location=f"@my_db.my_schema.my_stage/path/"),
    input_spec=InputSpec(
# we need to provide column_handling in the InputSpec to perform the necessary conversion
# FULL_STAGE_PATH: fully qualified path (db.schema.stage/path) to a file
# RAW_BYTES: download and convert the file to bytes (matching the predefined signature)
        column_handling={
            "IMAGES": {"input_format": InputFormat.FULL_STAGE_PATH, "convert_to": FileEncoding.RAW_BYTES}
        }
    )
)