working up to training

nlp/aml-cli-v2/data-science/environment/requirements.txt

@@ -4,7 +4,7 @@ torch==1.11.0
 pytorch_lightning==1.6.4
 transformers==4.18.0
 datasets==2.0.0
-rouge_score==
+rouge_score==0.0.4
 
 # for metrics reporting/plotting
 mlflow==1.25.1

nlp/aml-cli-v2/data-science/src/summarization/data_prep.py

@@ -66,6 +66,17 @@ def main():
 
     logger.info(f"Running with arguments: {args}")
 
+    # get tokenizer ready
+    tokenizer = AutoTokenizer.from_pretrained(args.model_checkpoint)
+    logger.info(f"tokenizer: {tokenizer}")
+
+    prefix = args.source_prefix if args.source_prefix is not None else ""
+    if args.source_prefix is None and "t5" in args.model_checkpoint.lower():
+        logger.warning(
+            "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
+            "`--source_prefix 'summarize: ' `"
+        )
+
     # Load dataset
     raw_dataset = load_dataset(args.dataset_name, args.dataset_config)
 
@@ -86,16 +97,6 @@ def main():
         logger.info("sampled raw dataset:")
         logger.info(raw_dataset.num_rows)
 
-    tokenizer = AutoTokenizer.from_pretrained(args.model_checkpoint)
-    logger.info(f"tokenizer: {tokenizer}")
-
-    prefix = args.source_prefix if args.source_prefix is not None else ""
-    if args.source_prefix is None and "t5" in args.model_checkpoint.lower():
-        logger.warning(
-            "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
-            "`--source_prefix 'summarize: ' `"
-        )
-
     def preprocess_function(examples):
         # remove pairs where at least one record is None
 

nlp/aml-cli-v2/mlops/azureml/train/pipeline.yaml

@@ -1,17 +1,7 @@
 $schema: https://azuremlschemas.azureedge.net/latest/pipelineJob.schema.json
 type: pipeline
 experiment_name: text-summarization-pubmed-experiment
-description: A text summarization using pubmed dataset for fine-tuning
-
-# compute: azureml:gpu-cluster
-
-# task: text_classification
-# primary_metric: accuracy
-# log_verbosity: debug
-
-# limits:
-#   timeout_minutes: 60
-
+description: Train a text summarization model using PubMed dataset for fine-tuning
 
 # <inputs_and_outputs>
 inputs:
@@ -27,25 +17,25 @@ inputs:
     value: 0.00005
 
 outputs: 
-  generic_data_prepared:
-  trained_model:
-  predictions:
-  score_report:
-  deploy_flag:
+  prepared_data:
+  finetuned_model:
+  # predictions:
+  # score_report:
+  # deploy_flag:
 # </inputs_and_outputs>
 
 settings:
   default_datastore: azureml:workspaceblobstore
-  default_compute: azureml:cpu-cluster
+  default_compute: azureml:cpu-cluster-lg
   continue_on_step_failure: false
 
 jobs:
   prep_finetuning_dataset:
     name: prep_finetuning_dataset
-    display_name: "Prepare our domain-specific dataset for training"
+    display_name: "Prepare dataset for training"
     code: ../../../data-science/src
     command: >-
-      python data_prep.py 
+      python summarization/data_prep.py 
       --dataset_name ${{inputs.dataset_name}} 
       --dataset_config ${{inputs.dataset_config}} 
       --text_column ${{inputs.text_column}} 
@@ -70,14 +60,14 @@ jobs:
       pretrained_model_name: ${{parent.inputs.pretrained_model_name}}
       source_prefix: "summarize: "
     outputs:
-      encodings: ${{parent.outputs.large_dataset_preprocessed}}
+      encodings: ${{parent.outputs.prepared_data}}
 
   finetune_model:
     name: finetune_model
-    display_name: fine-tune a summarization model
+    display_name: Fine-tune summarization model
     code: ../../../data-science/src
     command: >-
-      python run_summarization.py 
+      python summarization/run.py 
       --preprocessed_datasets ${{inputs.preprocessed_datasets}} 
       --learning_rate ${{inputs.learning_rate}} 
       --per_device_train_batch_size ${{inputs.per_device_train_batch_size}} 
@@ -104,7 +94,7 @@ jobs:
       per_device_eval_batch_size: ${{parent.inputs.batch_size}}
       source_prefix: "summarize: "
     outputs:
-      trained_model: ${{parent.outputs.trained_model}}
+      trained_model: ${{parent.outputs.finetuned_model}}
     compute: azureml:gpu-cluster
     distribution:
       type: pytorch
@@ -114,21 +104,19 @@ jobs:
 
   evaluate_model:
     name: evaluate_model
-    display_name: evaluate-model
+    display_name: Evaluate model
     code: ../../../data-science/src
     command: >-
-      python run_summarization.py 
+      python summarization/run.py 
       --preprocessed_datasets ${{inputs.preprocessed_datasets}} 
       --max_samples ${{inputs.max_samples}} 
       --output_dir outputs 
       --model_path ${{inputs.model_path}} 
-      --trained_model_path ${{outputs.trained_model_path}} 
+      --trained_model_path ""
       --do_eval
     environment: azureml:nlp_summarization_train@latest
     inputs:
       preprocessed_datasets: ${{parent.jobs.prep_finetuning_dataset.outputs.encodings}}
-      registered_model_name: "t5-small-cnn-pubmed"
-      model_path: ${{parent.jobs.finetune_model.outputs.trained_model
+      model_path: ${{parent.jobs.finetune_model.outputs.trained_model}}
       max_samples: ${{parent.inputs.max_samples}}
-    outputs:
-       trained_model_path:
+      registered_model_name: "t5-small-cnn-pubmed"

nlp/aml-cli-v2/mlops/azureml/train/train-env.yaml

@@ -1,6 +1,6 @@
 
 $schema: https://azuremlschemas.azureedge.net/latest/environment.schema.json
 name: nlp_summarization_train
-version: dev1
+version: dev2
 build:
   path: ../../../data-science/environment/