Consider the following pipeline:
example_gen = tfx.components.ImportExampleGen(input_base=_dataset_folder)
statistics_gen = tfx.components.StatisticsGen(examples=example_gen.outputs['examples'])
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True)
transform = tfx.components.Transform(
examples=example_gen.outputs['examples'],
schema=schema_gen.outputs['schema'],
module_file=os.path.abspath('preprocessing_fn.py'))
_trainer_module_file = 'run_fn.py'
trainer = tfx.components.Trainer(
module_file=os.path.abspath(_trainer_module_file),
examples=transform.outputs['transformed_examples'],
transform_graph=transform.outputs['transform_graph'],
schema=schema_gen.outputs['schema'],
train_args=tfx.proto.TrainArgs(num_steps=10),
eval_args=tfx.proto.EvalArgs(num_steps=6),)
pusher = tfx.components.Pusher(
model=trainer.outputs['model'],
push_destination=tfx.proto.PushDestination(
filesystem=tfx.proto.PushDestination.Filesystem(
base_directory=_serving_model_dir)
)
)
components = [
example_gen,
statistics_gen,
schema_gen,
transform,
trainer,
pusher,
]
_pipeline_data_folder = './simple_pipeline_data'
pipeline = tfx.dsl.Pipeline(
pipeline_name='simple_pipeline',
pipeline_root=_pipeline_data_folder,
metadata_connection_config=tfx.orchestration.metadata.sqlite_metadata_connection_config(
f'{_pipeline_data_folder}/metadata.db'),
components=components)
tfx.orchestration.LocalDagRunner().run(pipeline)
Now, let's assume that once the pipeline is down, I would like to do something with the artifacts. I know I can query the ML Metadata like this:
import ml_metadata as mlmd
connection_config = pipeline.metadata_connection_config
store = mlmd.MetadataStore(connection_config)
print(store.get_artifact_types())
But this way, I have no idea which IDs belong to the current pipeline. Sure, I can assume that the largest IDs represent the current pipeline artifacts but that's not going to be a practical approach in production when multiple executions might try to work with the same metadata store concurrently.
So, the question is how can I figure out the artifact IDs that were just created by the current execution?
[UPDATE]
To clarify the problem, consider the following partial solution:
def get_latest_artifact(metadata_connection_config, pipeline_name: str, component_name: str, type_name: str):
with Metadata(metadata_connection_config) as metadata:
context = metadata.store.get_context_by_type_and_name('node', f'{pipeline_name}.{component_name}')
artifacts = metadata.store.get_artifacts_by_context(context.id)
artifact_type = metadata.store.get_artifact_type(type_name)
latest_artifact = max([a for a in artifacts if a.type_id == artifact_type.id],
key=lambda a: a.last_update_time_since_epoch)
artifact = types.Artifact(artifact_type)
artifact.set_mlmd_artifact(latest_artifact)
return artifact
sqlite_path = './pipeline_data/metadata.db'
metadata_connection_config = tfx.orchestration.metadata.sqlite_metadata_connection_config(sqlite_path)
examples_artifact = get_latest_artifact(metadata_connection_config, 'simple_pipeline',
'SchemaGen', 'Schema')
Using get_latest_artifact function, I can get the latest artifact of a specific type from a specific pipeline. This will work even if two pipelines (with different names) create new artifacts concurrently. But it will fail when I try to extract the artifact of the "just finished" pipeline if multiple instances of the same pipeline are making changes to the store concurrently. That's because the function takes in the pipeline name as an input argument (as opposed to some pipeline unique ID).
I'm looking for a solution that works no matter how many different (or the same) pipelines work with the same store concurrently. At this point, I'm not sure if this can be done with MlMD. And if it cannot be done at the moment, I consider that a missed feature, a very crucial one.
OK, this is the solution I found. When defining the pipeline's components, you should use .with_id() method and give the component a custom ID. That way you can find it later on.
Here's an example. Let's say that I want to find the schema generated as part of the recently executed pipeline.
schema_gen = tfx.components.SchemaGen(
statistics=statistics_gen.outputs['statistics'],
infer_feature_shape=True).with_id('some_unique_id')
Then, the same function I defined above can be used like this:
def get_latest_artifact(metadata_connection_config, pipeline_name: str, component_name: str, type_name: str):
with Metadata(metadata_connection_config) as metadata:
context = metadata.store.get_context_by_type_and_name('node', f'{pipeline_name}.{component_name}')
artifacts = metadata.store.get_artifacts_by_context(context.id)
artifact_type = metadata.store.get_artifact_type(type_name)
latest_artifact = max([a for a in artifacts if a.type_id == artifact_type.id],
key=lambda a: a.last_update_time_since_epoch)
artifact = types.Artifact(artifact_type)
artifact.set_mlmd_artifact(latest_artifact)
return artifact
sqlite_path = './pipeline_data/metadata.db'
metadata_connection_config = tfx.orchestration.metadata.sqlite_metadata_connection_config(sqlite_path)
examples_artifact = get_latest_artifact(metadata_connection_config, 'simple_pipeline',
'some_unique_id', 'Schema')
Once your TFX pipeline completes the run, you can query the ML metadata using below code.
connection_config = interactive_context.metadata_connection_config
store = mlmd.MetadataStore(connection_config)
# All TFX artifacts are stored in the base directory
base_dir = connection_config.sqlite.filename_uri.split('metadata.sqlite')[0]
Once metadata is fetched, you can use below helper functions to view data from MD store. The display_types() function queries the list of all its stored ArtifactTypes. Then display_artifacts() function list all artifacts for given artifact type and their uri. The display_properties() function will give execution properties for given artifact.
Please refer MLMD tutorial for detailed implementation of below functions.
def display_types(types):
# Helper function to render dataframes for the artifact and execution types
table = {'id': [], 'name': []}
for a_type in types:
table['id'].append(a_type.id)
table['name'].append(a_type.name)
return pd.DataFrame(data=table)
def display_artifacts(store, artifacts):
# Helper function to render dataframes for the input artifacts
table = {'artifact id': [], 'type': [], 'uri': []}
for a in artifacts:
table['artifact id'].append(a.id)
artifact_type = store.get_artifact_types_by_id([a.type_id])[0]
table['type'].append(artifact_type.name)
table['uri'].append(a.uri.replace(base_dir, './'))
return pd.DataFrame(data=table)
def display_properties(store, node):
# Helper function to render dataframes for artifact and execution properties
table = {'property': [], 'value': []}
for k, v in node.properties.items():
table['property'].append(k)
table['value'].append(
v.string_value if v.HasField('string_value') else v.int_value)
for k, v in node.custom_properties.items():
table['property'].append(k)
table['value'].append(
v.string_value if v.HasField('string_value') else v.int_value)
return pd.DataFrame(data=table)
Example code to get the latest pushed model execution properties.
# get all artifacts with ArtifactType as PusherModel
pushed_models = store.get_artifacts_by_type("PushedModel")
# get the latest pushed model
pushed_model = pushed_models[-1]
# get execution properties for latest pushed model
display_properties(store, pushed_model)
Related
For my thesis, I want to train a PPO model on a collaborative multi-agent situation. The 'teammates' are already trained Behavior Cloning models each representing different kind of behavior. These models will not be trained simultaneously to the PPO. I want the PPO encounter all these different models during training, meaning: I want to switch out the 'teammate' of the PPO every x amount of epochs. How do I do this?
I am using the overcooked-ai environment. All code would be too much to include. But the generation of the trainer ends up returning this:
trainer = PPOTrainer(env="overcooked_multi_agent", config={
"multiagent": multi_agent_config,
"callbacks" : TrainingCallbacks,
"custom_eval_function" : get_rllib_eval_function(evaluation_params, environment_params['eval_mdp_params'], environment_params['env_params'],
environment_params["outer_shape"], 'ppo', 'ppo' if self_play else 'bc',
verbose=params['verbose']),
"env_config" : environment_params,
"eager" : False,
**training_params
}, logger_creator=custom_logger_creator
and the multi-agent-config is created like
def gen_policy(policy_type="ppo"):
# supported policy types thus far
assert policy_type in ["ppo", "bc"]
if policy_type == "ppo":
config = {
"model" : {
"custom_options" : model_params,
"custom_model" : "MyPPOModel"
}
}
return (None, env.ppo_observation_space, env.action_space, config)
elif policy_type == "bc":
bc_cls = bc_params['bc_policy_cls']
bc_config = bc_params['bc_config']
return (bc_cls, env.bc_observation_space, env.action_space, bc_config)
# Rllib compatible way of setting the directory we store agent checkpoints in
logdir_prefix = "{0}_{1}_{2}".format(params["experiment_name"], params['training_params']['seed'], timestr)
def custom_logger_creator(config):
"""Creates a Unified logger that stores results in <params['results_dir']>/<params["experiment_name"]>_<seed>_<timestamp>
"""
results_dir = params['results_dir']
if not os.path.exists(results_dir):
try:
os.makedirs(results_dir)
except Exception as e:
print("error creating custom logging dir. Falling back to default logdir {}".format(DEFAULT_RESULTS_DIR))
results_dir = DEFAULT_RESULTS_DIR
logdir = tempfile.mkdtemp(
prefix=logdir_prefix, dir=results_dir)
logger = UnifiedLogger(config, logdir, loggers=None)
return logger
# Create rllib compatible multi-agent config based on params
multi_agent_config = {}
all_policies = ['ppo']
# Whether both agents should be learned
self_play = iterable_equal(multi_agent_params['bc_schedule'], OvercookedMultiAgent.self_play_bc_schedule)
if not self_play:
all_policies.append('bc')
multi_agent_config['policies'] = { policy : gen_policy(policy) for policy in all_policies }
def select_policy(agent_id):
if agent_id.startswith('ppo'):
return 'ppo'
if agent_id.startswith('bc'):
return 'bc'
multi_agent_config['policy_mapping_fn'] = select_policy
multi_agent_config['policies_to_train'] = 'ppo'
Do I already change my configuration to include multiple BHC models (code above is built to handle 1), or can I somehow change the models in the trainingloop?
Any ideas would be super helpful!
I'm struggling to correctly set Vertex AI pipeline which does the following:
read data from API and store to GCS and as as input for batch prediction.
get an existing model (Video classification on Vertex AI)
create Batch prediction job with input from point 1.
As it will be seen, I don't have much experience with Vertex Pipelines/Kubeflow thus I'm asking for help/advice, hope it's just some beginner mistake.
this is the gist of the code I'm using as pipeline
from google_cloud_pipeline_components import aiplatform as gcc_aip
from kfp.v2 import dsl
from kfp.v2.dsl import component
from kfp.v2.dsl import (
Output,
Artifact,
Model,
)
PROJECT_ID = 'my-gcp-project'
BUCKET_NAME = "mybucket"
PIPELINE_ROOT = "{}/pipeline_root".format(BUCKET_NAME)
#component
def get_input_data() -> str:
# getting data from API, save to Cloud Storage
# return GS URI
gcs_batch_input_path = 'gs://somebucket/file'
return gcs_batch_input_path
#component(
base_image="python:3.9",
packages_to_install=['google-cloud-aiplatform==1.8.0']
)
def load_ml_model(project_id: str, model: Output[Artifact]):
"""Load existing Vertex model"""
import google.cloud.aiplatform as aip
model_id = '1234'
model = aip.Model(model_name=model_id, project=project_id, location='us-central1')
#dsl.pipeline(
name="batch-pipeline", pipeline_root=PIPELINE_ROOT,
)
def pipeline(gcp_project: str):
input_data = get_input_data()
ml_model = load_ml_model(gcp_project)
gcc_aip.ModelBatchPredictOp(
project=PROJECT_ID,
job_display_name=f'test-prediction',
model=ml_model.output,
gcs_source_uris=[input_data.output], # this doesn't work
# gcs_source_uris=['gs://mybucket/output/'], # hardcoded gs uri works
gcs_destination_output_uri_prefix=f'gs://{PIPELINE_ROOT}/prediction_output/'
)
if __name__ == '__main__':
from kfp.v2 import compiler
import google.cloud.aiplatform as aip
pipeline_export_filepath = 'test-pipeline.json'
compiler.Compiler().compile(pipeline_func=pipeline,
package_path=pipeline_export_filepath)
# pipeline_params = {
# 'gcp_project': PROJECT_ID,
# }
# job = aip.PipelineJob(
# display_name='test-pipeline',
# template_path=pipeline_export_filepath,
# pipeline_root=f'gs://{PIPELINE_ROOT}',
# project=PROJECT_ID,
# parameter_values=pipeline_params,
# )
# job.run()
When running the pipeline it throws this exception when running Batch prediction:
details = "List of found errors: 1.Field: batch_prediction_job.model; Message: Invalid Model resource name.
so I'm not sure what could be wrong. I tried to load model in the notebook (outside of component) and it correctly returns.
Second issue I'm having is referencing GCS URI as output from component to batch job input.
input_data = get_input_data2()
gcc_aip.ModelBatchPredictOp(
project=PROJECT_ID,
job_display_name=f'test-prediction',
model=ml_model.output,
gcs_source_uris=[input_data.output], # this doesn't work
# gcs_source_uris=['gs://mybucket/output/'], # hardcoded gs uri works
gcs_destination_output_uri_prefix=f'gs://{PIPELINE_ROOT}/prediction_output/'
)
During compilation, I get following exception TypeError: Object of type PipelineParam is not JSON serializable, though I think this could be issue of ModelBatchPredictOp component.
Again any help/advice appreciated, I'm dealing with this from yesterday, so maybe I missed something obvious.
libraries I'm using:
google-cloud-aiplatform==1.8.0
google-cloud-pipeline-components==0.2.0
kfp==1.8.10
kfp-pipeline-spec==0.1.13
kfp-server-api==1.7.1
UPDATE
After comments, some research and tuning, for referencing model this works:
#component
def load_ml_model(project_id: str, model: Output[Artifact]):
region = 'us-central1'
model_id = '1234'
model_uid = f'projects/{project_id}/locations/{region}/models/{model_id}'
model.uri = model_uid
model.metadata['resourceName'] = model_uid
and then I can use it as intended:
batch_predict_op = gcc_aip.ModelBatchPredictOp(
project=gcp_project,
job_display_name=f'batch-prediction-test',
model=ml_model.outputs['model'],
gcs_source_uris=[input_batch_gcs_path],
gcs_destination_output_uri_prefix=f'gs://{BUCKET_NAME}/prediction_output/test'
)
UPDATE 2
regarding GCS path, a workaround is to define path outside of the component and pass it as an input parameter, for example (abbreviated):
#dsl.pipeline(
name="my-pipeline",
pipeline_root=PIPELINE_ROOT,
)
def pipeline(
gcp_project: str,
region: str,
bucket: str
):
ts = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
gcs_prediction_input_path = f'gs://{BUCKET_NAME}/prediction_input/video_batch_prediction_input_{ts}.jsonl'
batch_input_data_op = get_input_data(gcs_prediction_input_path) # this loads input data to GCS path
batch_predict_op = gcc_aip.ModelBatchPredictOp(
project=gcp_project,
model=training_job_run_op.outputs["model"],
job_display_name='batch-prediction',
# gcs_source_uris=[batch_input_data_op.output],
gcs_source_uris=[gcs_prediction_input_path],
gcs_destination_output_uri_prefix=f'gs://{BUCKET_NAME}/prediction_output/',
).after(batch_input_data_op) # we need to add 'after' so it runs after input data is prepared since get_input_data doesn't returns anything
still not sure, why it doesn't work/compile when I return GCS path from get_input_data component
I'm glad you solved most of your main issues and found a workaround for model declaration.
For your input.output observation on gcs_source_uris, the reason behind it is because the way the function/class returns the value. If you dig inside the class/methods of google_cloud_pipeline_components you will find that it implements a structure that will allow you to use .outputs from the returned value of the function called.
If you go to the implementation of one of the components of the pipeline you will find that it returns an output array from convert_method_to_component function. So, in order to have that implemented in your custom class/function your function should return a value which can be called as an attribute. Below is a basic implementation of it.
class CustomClass():
def __init__(self):
self.return_val = {'path':'custompath','desc':'a desc'}
#property
def output(self):
return self.return_val
hello = CustomClass()
print(hello.output['path'])
If you want to dig more about it you can go to the following pages:
convert_method_to_component, which is the implementation of convert_method_to_component
Properties, basics of property in python.
I am trying to extract the content of the [Documentation] section as a string for comparision with other part in a Python script.
I was told to use Robot framework API https://robot-framework.readthedocs.io/en/stable/
to extract but I have no idea how.
However, I am required to work with version 3.1.2
Example:
*** Test Cases ***
ATC Verify that Sensor Battery can enable and disable manufacturing mode
[Documentation] E1: This is the description of the test 1
... E2: This is the description of the test 2
[Tags] E1 TRACE{Trace_of_E1}
... E2 TRACE{Trace_of_E2}
Extract the string as
E1: This is the description of the test 1
E2: This is the description of the test 2
Have a look at these examples. I did something similar to generate testplans descritio. I tried to adapt my code to your requirements and this could maybe work for you.
import os
import re
from robot.api.parsing import (
get_model, get_tokens, Documentation, EmptyLine, KeywordCall,
ModelVisitor, Token
)
class RobotParser(ModelVisitor):
def __init__(self):
# Create object with remarkup_text to store formated documentation
self.text = ''
def get_text(self):
return self.text
def visit_TestCase(self, node):
# The matched `TestCase` node is a block with `header` and
# `body` attributes. `header` is a statement with familiar
# `get_token` and `get_value` methods for getting certain
# tokens or their value.
for keyword in node.body:
# skip empty lines
if keyword.get_value(Token.DOCUMENTATION) == None:
continue
self.text += keyword.get_value(Token.ARGUMENT)
def visit_Documentation(self,node):
# The matched "Documentation" node with value
self.remarkup_text += node.value + self.new_line
def visit_File(self, node):
# Call `generic_visit` to visit also child nodes.
return self.generic_visit(node)
if __name__ == "__main__":
path = "../tests"
for filename in os.listdir(path):
if re.match(".*\.robot", filename):
model = get_model(os.path.join(path, filename))
robot_parser = RobotParser()
robot_parser.visit(model)
text=robot_parser._text()
The code marked as best answer didn't quite work for me and has a lot of redundancy but it inspired me enough to get into the parsing and write it in a much readable and efficient way that actually works as is. You just have to have your own way of generating & iterating through filesystem where you call the get_robot_metadata(filepath) function.
from robot.api.parsing import (get_model, ModelVisitor, Token)
class RobotParser(ModelVisitor):
def __init__(self):
self.testcases = {}
def visit_TestCase(self, node):
testcasename = (node.header.name)
self.testcases[testcasename] = {}
for section in node.body:
if section.get_value(Token.DOCUMENTATION) != None:
documentation = section.value
self.testcases[testcasename]['Documentation'] = documentation
elif section.get_value(Token.TAGS) != None:
tags = section.values
self.testcases[testcasename]['Tags'] = tags
def get_testcases(self):
return self.testcases
def get_robot_metadata(filepath):
if filepath.endswith('.robot'):
robot_parser = RobotParser()
model = get_model(filepath)
robot_parser.visit(model)
metadata = robot_parser.get_testcases()
return metadata
This function will be able to extract the [Documentation] section from the testcase:
def documentation_extractor(testcase):
documentation = []
for setting in testcase.settings:
if len(setting) > 2 and setting[1].lower() == "[documentation]":
for doc in setting[2:]:
if doc.startswith("#"):
# the start of a comment, so skip rest of the line
break
documentation.append(doc)
break
return "\n".join(documentation)
I'm having some issues with persisting documents with pymongo when using insert_many.
I'm handing over a list of dicts to insert_many and it works fine from inside the same script that does the inserting. Less so once the script has finished.
def row_to_doc(row):
rowdict = row.to_dict()
for key in rowdict:
val = rowdict[key]
if type(val) == float or type(val) == np.float64:
if np.isnan(val):
# If we want a SQL style document collection
rowdict[key] = None
# If we want a NoSQL style document collection
# del rowdict[key]
return rowdict
def dataframe_to_collection(df):
n = len(df)
doc_list = []
for k in range(n):
doc_list.append(row_to_doc(df.iloc[k]))
return doc_list
def get_mongodb_client(host="localhost", port=27017):
return MongoClient(host, port)
def create_collection(client):
db = client["material"]
return db["master-data"]
def add_docs_to_mongo(collection, doc_list):
collection.insert_many(doc_list)
def main():
client = get_mongodb_client()
csv_fname = "some_csv_fname.csv"
df = get_clean_csv(csv_fname)
doc_list = dataframe_to_collection(df)
collection = create_collection(client)
add_docs_to_mongo(collection, doc_list)
test_doc = collection.find_one({"MATERIAL": "000000000000000001"})
When I open up another python REPL and start looking through the client.material.master_data collection with collection.find_one({"MATERIAL": "000000000000000001"}) or collection.count_documents({}) I get None for the find_one and 0 for the count_documents.
Is there a step where I need to call some method to persist the data to disk? db.collection.save() in the mongo client API sounds like what I need but it's just another way of inserting documents from what I have read. Any help would be greatly appreciated.
The problem was that I was getting my collection via client.db_name.collection_name and it wasn't getting the same collection I was creating with my code. client.db_name["collection-name"] solved my issue. Weird.
I have a Bigquery task which only aims to append a daily temp table (Table-xxxx-xx-xx) to an existing table (PersistingTable).
I am not sure how to handle the output(self) method. Indeed, I can not just output PersistingTable as a luigi.contrib.bigquery.BigQueryTarget, since it already exists before the process started. Has anyone asked himself such a question?
I could not find an answer anywhere else so I will give my solution even though this is a very old question.
I created a new class that inherits from luigi.contrib.bigquery.BigQueryLoadTask
class BigQueryLoadIncremental(luigi.contrib.bigquery.BigQueryLoadTask):
'''
a subclass that checks whether a write-log on gcs exists to append data to the table
needs to define Two Outputs! [0] of type BigQueryTarget and [1] of type GCSTarget
Everything else is left unchanged
'''
def exists(self):
return luigi.contrib.gcs.GCSClient.exists(self.output()[1].path)
#property
def write_disposition(self):
"""
Set to WRITE_APPEND as this subclass only makes sense for this
"""
return luigi.contrib.bigquery.WriteDisposition.WRITE_APPEND
def run(self):
output = self.output()[0]
gcs_output = self.output()[1]
assert isinstance(output,
luigi.contrib.bigquery.BigQueryTarget), 'Output[0] must be a BigQueryTarget, not %s' % (
output)
assert isinstance(gcs_output,
luigi.contrib.gcs.GCSTarget), 'Output[1] must be a Cloud Storage Target, not %s' % (
gcs_output)
bq_client = output.client
source_uris = self.source_uris()
assert all(x.startswith('gs://') for x in source_uris)
job = {
'projectId': output.table.project_id,
'configuration': {
'load': {
'destinationTable': {
'projectId': output.table.project_id,
'datasetId': output.table.dataset_id,
'tableId': output.table.table_id,
},
'encoding': self.encoding,
'sourceFormat': self.source_format,
'writeDisposition': self.write_disposition,
'sourceUris': source_uris,
'maxBadRecords': self.max_bad_records,
'ignoreUnknownValues': self.ignore_unknown_values
}
}
}
if self.source_format == luigi.contrib.bigquery.SourceFormat.CSV:
job['configuration']['load']['fieldDelimiter'] = self.field_delimiter
job['configuration']['load']['skipLeadingRows'] = self.skip_leading_rows
job['configuration']['load']['allowJaggedRows'] = self.allow_jagged_rows
job['configuration']['load']['allowQuotedNewlines'] = self.allow_quoted_new_lines
if self.schema:
job['configuration']['load']['schema'] = {'fields': self.schema}
# test write to and removal of GCS pseudo output in order to make sure this does not fail.
gcs_output.fs.put_string(
'test write for task {} (this file should have been removed immediately)'.format(self.task_id),
gcs_output.path)
gcs_output.fs.remove(gcs_output.path)
bq_client.run_job(output.table.project_id, job, dataset=output.table.dataset)
gcs_output.fs.put_string(
'success! The following BigQuery Job went through without errors: {}'.format(self.task_id), gcs_output.path)
it uses a second output (which might violate luigis atomicity principle) on google cloud storage. Example usage:
class LeadsToBigQuery(BigQueryLoadIncremental):
date = luigi.DateParameter(default=datetime.date.today())
def output(self):
return luigi.contrib.bigquery.BigQueryTarget(project_id=...,
dataset_id=...,
table_id=...), \
create_gcs_target(...)