Cleaner argument handling & nlp/common/ folder (#16)

By moving the arguments into their own dataclass (available in Python 3.7), we can group certain types of arguments, such as ModelArguments and SageMakerArguments. This lets us consolidate the sagemaker scripts into a single file, and makes the arguments simpler to pass around in functions.

Moves several files to common/. Users will need to set PYTHONPATH=/path/to/deep-learning-models/nlp. Also fixes PYTHONPATH to /opt/ml/... in the SageMaker container, so those jobs should run.

Also adds support to log hyperparameters in TensorBoard.

models/nlp/albert/README.md

@@ -20,7 +20,9 @@ Language models help AWS customers to improve search results, text classificatio
 3. Create an Amazon Elastic Container Registry (ECR) repository. Then build a Docker image from `docker/ngc_sagemaker.Dockerfile` and push it to ECR.
 
 ```bash
-export IMAGE=${ACCOUNT_ID}.dkr.ecr.us-east-1.amazonaws.com/${REPO}:ngc_tf21_sagemaker
+export ACCOUNT_ID=
+export REPO=
+export IMAGE=${ACCOUNT_ID}.dkr.ecr.us-east-1.amazonaws.com/${REPO}:ngc_tf210_sagemaker
 docker build -t ${IMAGE} -f docker/ngc_sagemaker.Dockerfile .
 $(aws ecr get-login --no-include-email)
 docker push ${IMAGE}
@@ -39,8 +41,13 @@ export SAGEMAKER_SECURITY_GROUP_IDS=sg-123,sg-456
 5. Launch the SageMaker job.
 
 ```bash
-python sagemaker_pretraining.py \
+# Add the main folder to your PYTHONPATH
+export PYTHONPATH=$PYTHONPATH:/path/to/deep-learning-models/models/nlp
+
+python launch_sagemaker.py \
     --source_dir=. \
+    --entry_point=run_pretraining.py \
+    --sm_job_name=albert-pretrain \
     --instance_type=ml.p3dn.24xlarge \
     --instance_count=1 \
     --load_from=scratch \
@@ -52,9 +59,35 @@ python sagemaker_pretraining.py \
     --total_steps=125000 \
     --learning_rate=0.00176 \
     --optimizer=lamb \
+    --log_frequency=10 \
     --name=myfirstjob
 ```
 
+6. Launch a SageMaker finetuning job.
+
+```bash
+python launch_sagemaker.py \
+    --source_dir=. \
+    --entry_point=run_squad.py \
+    --sm_job_name=albert-squad \
+    --instance_type=ml.p3dn.24xlarge \
+    --instance_count=1 \
+    --load_from=scratch \
+    --model_type=albert \
+    --model_size=base \
+    --batch_size=6 \
+    --total_steps=8144 \
+    --warmup_steps=814 \
+    --learning_rate=3e-5 \
+    --task_name=squadv2
+```
+
+7. Enter the Docker container to debug and edit code.
+
+```bash
+docker run -it -v=/fsx:/fsx --gpus=all --shm-size=1g --ulimit memlock=-1 --ulimit stack=67108864 --rm ${IMAGE} /bin/bash
+```
+
 <!-- ### Training results
 
 These will be posted shortly. -->

models/nlp/albert/launch_sagemaker.py

@@ -0,0 +1,54 @@
+import argparse
+import dataclasses
+
+from transformers import HfArgumentParser
+
+from common.arguments import (
+    DataTrainingArguments,
+    LoggingArguments,
+    ModelArguments,
+    SageMakerArguments,
+    TrainingArguments,
+)
+from common.sagemaker_utils import launch_sagemaker_job
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser = HfArgumentParser(
+        (
+            ModelArguments,
+            DataTrainingArguments,
+            TrainingArguments,
+            LoggingArguments,
+            SageMakerArguments,
+        )
+    )
+    model_args, data_args, train_args, log_args, sm_args = parser.parse_args_into_dataclasses()
+
+    hyperparameters = dict()
+    for args in [model_args, data_args, train_args, log_args]:
+        for key, value in dataclasses.asdict(args).items():
+            if value is not None:
+                hyperparameters[key] = value
+    hyperparameters["fsx_prefix"] = "/opt/ml/input/data/training"
+
+    instance_abbr = {
+        "ml.p3dn.24xlarge": "p3dn",
+        "ml.p3.16xlarge": "p316",
+        "ml.g4dn.12xlarge": "g4dn",
+    }[sm_args.instance_type]
+    job_name = f"{sm_args.sm_job_name}-{sm_args.instance_count}x{instance_abbr}"
+
+    launch_sagemaker_job(
+        hyperparameters=hyperparameters,
+        job_name=job_name,
+        source_dir=sm_args.source_dir,
+        entry_point=sm_args.entry_point,
+        instance_type=sm_args.instance_type,
+        instance_count=sm_args.instance_count,
+        role=sm_args.role,
+        image_name=sm_args.image_name,
+        fsx_id=sm_args.fsx_id,
+        subnet_ids=sm_args.subnet_ids,
+        security_group_ids=sm_args.security_group_ids,
+    )