def gh_summ( #pylint: disable=unused-argument
train_steps: 'Integer' = 2019300,
project: str = 'YOUR_PROJECT_HERE',
github_token: str = 'YOUR_GITHUB_TOKEN_HERE',
working_dir: 'GCSPath' = 'gs://YOUR_GCS_DIR_HERE',
checkpoint_dir: 'GCSPath' = 'gs://aju-dev-demos-codelabs/kubecon/model_output_tbase.bak2019000/',
deploy_webapp: str = 'true',
data_dir: 'GCSPath' = 'gs://aju-dev-demos-codelabs/kubecon/t2t_data_gh_all/'
):
copydata = copydata_op(
data_dir=data_dir,
checkpoint_dir=checkpoint_dir,
model_dir='%s/%s/model_output' % (working_dir, dsl.RUN_ID_PLACEHOLDER),
action=COPY_ACTION,
)
train = train_op(
data_dir=data_dir,
model_dir=copydata.outputs['copy_output_path'],
action=TRAIN_ACTION, train_steps=train_steps,
deploy_webapp=deploy_webapp
)
serve = dsl.ContainerOp(
name='serve',
image='gcr.io/google-samples/ml-pipeline-kubeflow-tfserve:v5',
arguments=["--model_name", 'ghsumm-%s' % (dsl.RUN_ID_PLACEHOLDER,),
"--model_path", train.outputs['train_output_path']
]
)
# train.set_gpu_limit(1)
train.set_gpu_limit(1).apply(gcp.use_preemptible_nodepool()).set_retry(10)
with dsl.Condition(train.outputs['launch_server'] == 'true'):
webapp = dsl.ContainerOp(
name='webapp',
image='gcr.io/google-samples/ml-pipeline-webapp-launcher:v7ap',
arguments=["--model_name", 'ghsumm-%s' % (dsl.RUN_ID_PLACEHOLDER,),
"--github_token", github_token]
)
webapp.after(serve)
def gh_summ( #pylint: disable=unused-argument
train_steps=2019300,
project='YOUR_PROJECT_HERE',
github_token='YOUR_GITHUB_TOKEN_HERE',
working_dir='YOUR_GCS_DIR_HERE',
checkpoint_dir='gs://aju-dev-demos-codelabs/kubecon/model_output_tbase.bak2019000',
deploy_webapp='true',
data_dir='gs://aju-dev-demos-codelabs/kubecon/t2t_data_gh_all/'):
copydata = copydata_op(
working_dir=working_dir,
data_dir=data_dir,
checkpoint_dir=checkpoint_dir,
model_dir='%s/%s/model_output' % (working_dir, '{{workflow.name}}'),
action=COPY_ACTION).apply(gcp.use_gcp_secret('user-gcp-sa'))
log_dataset = metadata_log_op(log_type=DATASET,
workspace_name=WORKSPACE_NAME,
run_name='{{workflow.name}}',
data_uri=data_dir)
train = train_op(
working_dir=working_dir,
data_dir=data_dir,
checkpoint_dir=checkpoint_dir,
model_dir='%s/%s/model_output' % (working_dir, '{{workflow.name}}'),
action=TRAIN_ACTION,
train_steps=train_steps,
deploy_webapp=deploy_webapp).apply(gcp.use_gcp_secret('user-gcp-sa'))
log_model = metadata_log_op(log_type=MODEL,
workspace_name=WORKSPACE_NAME,
run_name='{{workflow.name}}',
model_uri='%s/%s/model_output' %
(working_dir, '{{workflow.name}}'))
serve = dsl.ContainerOp(
name='serve',
image='gcr.io/google-samples/ml-pipeline-kubeflow-tfserve',
arguments=[
"--model_name",
'ghsumm-%s' % ('{{workflow.name}}', ), "--model_path",
'%s/%s/model_output/export' % (working_dir, '{{workflow.name}}')
])
log_dataset.after(copydata)
train.after(copydata)
log_model.after(train)
serve.after(train)
train.set_gpu_limit(4).apply(gcp.use_preemptible_nodepool()).set_retry(5)
train.set_memory_limit('48G')
with dsl.Condition(train.output == 'true'):
webapp = dsl.ContainerOp(
name='webapp',
image='gcr.io/google-samples/ml-pipeline-webapp-launcher:v2ap',
arguments=[
"--model_name",
'ghsumm-%s' % ('{{workflow.name}}', ), "--github_token",
github_token
])
webapp.after(serve)
def flipcoin():
flip = FlipCoinOp().apply(gcp.use_preemptible_nodepool()).set_gpu_limit(
1, 'nvidia').set_retry(5)
def baseline_repro_pipeline(
data_bucket: str = 'voxsrc-2020-voxceleb-v4',
test_list: str = 'vox1_full.txt',
# @note test_utterances_list is in the same format as train_list, but for
# the test data. Whereas test_list contains utterance pairs for
# evaluation
test_utterances_list: str = 'vox1_full_utterances.txt',
train_list: str = 'vox2_full.txt',
test_path: str = 'vox1_full.tar.gz',
train_path: str = 'vox2_full.tar.gz',
checkpoint_bucket: str = 'voxsrc-2020-checkpoints',
batch_size: int = 750,
max_epoch: int = 21,
n_speakers: int = 2,
test_interval: int = 3,
feature_extraction_threads: int = 16,
data_loader_threads: int = 7,
# @note This run ID contains "full" pre-extracted features for vox1 and vox2
reuse_run_with_id: str = "milo_webster-19rvuxfu",
gaussian_noise_std: float = .9,
):
# set prod_hw=True to enable production hardware (preemptible V100).
# Encountered odd issues when node resource constraints aren't known at
# "compile time" of kf pipeline file
prod_hw = True
run_id = '{{workflow.uid}}'
feature_extraction_task = feature_extraction_op(
data_bucket=data_bucket,
test_utterances_list=test_list,
train_list=train_list,
test_path=test_path,
train_path=train_path,
run_id=run_id,
num_threads=feature_extraction_threads,
reuse_run_with_id=reuse_run_with_id)
# default feature extractor to high-perf pool if not in pass-through mode
# if in pass-through mode, there's no reason to use a beefy node
if not reuse_run_with_id:
feature_extraction_task.set_cpu_request("9").set_cpu_limit("16")
train_task = train_op(
data_bucket=data_bucket,
test_list=test_list,
train_list=train_list,
test_path=feature_extraction_task.outputs['test_feats_tar_path'],
train_path=feature_extraction_task.outputs['train_feats_tar_path'],
batch_size=batch_size,
max_epoch=max_epoch,
checkpoint_bucket=checkpoint_bucket,
run_id=run_id,
n_speakers=n_speakers,
test_interval=test_interval,
gaussian_noise_std=gaussian_noise_std,
n_data_loader_thread=data_loader_threads,
)
train_task.add_pvolumes({'/dev/shm': ipc_shared_mem_volume})
train_task.after(feature_extraction_task)
# add Weights & Biases credentials
if "WANDB_API_KEY" in os.environ:
train_task.add_env_variable(
k8s_client.V1EnvVar(name='WANDB_API_KEY',
value=os.environ["WANDB_API_KEY"]))
else:
raise 'Error: No WandB API key set in environment'
# @note These resource requests autoscale an autoscalable node pool from
# 0->1 that matches the corresponding config. Autoscaled nodes will be
# deactivated on GCP after 10 minutes of inactivity
if prod_hw:
# require training to run on a preemptible node pool
train_task\
.apply(gcp.use_preemptible_nodepool(hard_constraint=True))\
.set_retry(5)
# require training to run on a node with a gpu of type 'train_gpu_type'
train_task\
.set_gpu_limit(1)\
.add_node_selector_constraint('cloud.google.com/gke-accelerator',
'nvidia-tesla-v100')
