# --- MAIN EXPERIMENT SETTINGS
name: test-s3prefix
description: 'Test benchmark with S3Prefix datasets end-to-end'
role: SageMakerRoleBenchmark
region: us-east-1
bucket: null                # 'null' for Default Bucket
output_prefix: experiments
parallelism: 6              # 0 is sync execution, >=1 is async execution
repeat: 1
disable_profiler: True
order: interleave           # 'default | 'interleave' | 'random' - order to run the 'repeat'ed trials (not to create hot-spots)
clean: True                 # 'True' - delete all previous trials before start



# --- DATASET DEFINITIONS
datasets:
  Synth-100MB-a-1MB:
    type: s3prefix
    format: tfrecord/raw
    bucket: sagemaker-benchmark-us-east-1-242711262407
    prefix: 'datasets/synthetic'

  Synth-1GB-a-100MB:
    type: s3prefix
    format: tfrecord/raw
    bucket: sagemaker-benchmark-us-east-1-242711262407
    prefix: 'datasets/synthetic'


# --- FSx FILESYSTEM SETUP (is required only if 'fsx' input mode is used)
fsx:
  fsx_file_system_id: fs-XXXXXXXXXXXXXXXXX   # <--- specify your FSX filesystem ID
  fsx_file_system_directory_path: /fsx
  security_group_ids:
      - sg-XXXXXXXXXXXXXXXXX                 # <--- specify your Security Group IDs
  subnets:
      - subnet-XXXXXXXXXXXXXXXXX             # <--- specify your Subnet IDs


# --- TRIAL DEFINITIONS
# Default values for trials (if not overriden by specific trial)
base_trial:
  #script: script-nomodel.py
  source_dir: scripts
  framework: tf
  framework_version: '2.4.1'
  py_version: py37
  volume_size: 225
  instance_type: ml.g4dn.2xlarge
  instance_count: 1
  max_run: 600
  hyperparameters:
    prefetch_first: False
    prefetch_size: -1
    prefetch_to_gpu: False
    num_parallel_calls: -1    # (None or <int>) goes as arg to tf.map (and tf.batch) 
    num_parallel_reads: -1    # (None or <int>) goes as arg to TFRecordDataset
    private_pool: 16          # (None or <int>) private workers thread pool
    epochs: 1
    batch_size: 32
    shuffle: False
    backbone_model: resnet50
    input_dim: 300
    cache: None               # None, 'mem', or <path> for on-disk cache (e.g. 'tmp/cache')
    

# Trial definitions
trials: 

# --- NOMODEL tests ---
  nomodel-ffm:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: ffm

  nomodel-file:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: file

  nomodel-pipe:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: pipe

  nomodel-fsx:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: fsx

  nomodel-s3tf:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: s3tf

  nomodel-ffm-2:
    script: script-nomodel.py
    inputs:
      train:
        dataset: Synth-1GB-a-100MB
        input_mode: ffm

# --- MODEL tests ---
  model-ffm:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: ffm

  model-file:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: file

  model-pipe:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: pipe

  model-fsx:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: fsx

  model-s3tf:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-100MB-a-1MB
        input_mode: s3tf

  model-ffm-2:
    script: script-model.py
    inputs:
      train:
        dataset: Synth-1GB-a-100MB
        input_mode: ffm