Mri tumor segmentation scenario

ci/Dockerfile.train

@@ -15,6 +15,12 @@ RUN python3.9 get-pip.py
 RUN pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu
 RUN pip3 --default-timeout=1000 install pyspark pandas opacus onnx onnx2pytorch scikit-learn scipy matplotlib 
 
+## ADDED: For language tasks
+RUN pip3 install transformers datasets peft
+
+## ADDED: For vision tasks
+RUN pip3 install --default-timeout=100 opencv-python pillow monai tqdm
+
 RUN apt-get install -y jq
 
 # Install contract ledger client

scenarios/llm-finetune/.gitignore

@@ -0,0 +1,3 @@
+/data/
+/model/
+/output/

scenarios/llm-finetune/README.md

@@ -0,0 +1,74 @@
+# LLM Fine-tuning with Differential Privacy
+
+This scenario demonstrates how a Large Language Model (LLM) can be fine-tuned for medical question answering using the join of multiple (potentially PII-sensitive) datasets. The Training Data Consumer (TDC) building the model gets into a contractual agreement with multiple Training Data Providers (TDPs), and the model is fine-tuned on the joined datasets in a data-blind manner within the CCR, maintaining privacy guarantees using differential privacy. For demonstration purposes, this scenario uses open-source models and datasets from HuggingFace.
+
+The end-to-end training pipeline consists of the following phases:
+
+1. Data pre-processing
+2. Data packaging, encryption and upload
+3. Model packaging, encryption and upload
+4. Encryption key import with key release policies
+5. Deployment and execution of CCR
+6. Model decryption
+
+## Build container images
+
+Build container images required for this sample as follows:
+
+```bash
+cd scenarios/llm-finetune
+./ci/build.sh
+```
+
+This script builds the following container images:
+
+- `preprocess-icmr, preprocess-cowin, preprocess-index`: Containers that pre-process the text datasets.
+- `ccr-model-save`: Container that saves the base model to be fine-tuned.
+
+## Data pre-processing and de-identification
+
+Acting as TDPs for each dataset, run the following scripts to de-identify the datasets:
+
+```bash
+cd scenarios/llm-finetune/deployment/docker
+./preprocess.sh
+```
+
+This script performs pre-processing of the text datasets before the training process.
+
+## Prepare model for training
+
+Next, acting as a TDC, load and save a sample model using the following script:
+
+```bash
+./save-model.sh
+```
+
+This script will save the base model within `scenarios/llm-finetune/model/`.
+
+## Deploy locally
+
+Assuming you have cleartext access to all the datasets, you can fine-tune the model as follows:
+
+```bash
+./train.sh
+```
+
+The script joins the datasets and fine-tunes the model using a pipeline configuration defined in [pipeline_config.json](./config/pipeline_config.json). The fine-tuning process uses:
+
+- LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning
+- Differential Privacy via Opacus
+- Weight quantization for memory efficiency
+
+If all goes well, you should see training progress output similar to:
+
+```
+Epoch: 1 | Step: 50 | Train loss: 2.342
+Epoch: 1 | Step: 100 | Train loss: 2.123
+...
+Epoch 1 completed. Average loss: 2.156
+```
+
+The fine-tuned model will be saved under the folder `/scenarios/llm-finetune/output`.
+
+## Deploy on CCR

scenarios/llm-finetune/ci/Dockerfile.chatdoctor

@@ -0,0 +1,10 @@
+FROM ubuntu:20.04
+
+ENV DEBIAN_FRONTEND="noninteractive"
+
+RUN apt-get upgrade && apt-get update \
+    && apt-get install -y python3 python3-pip
+
+RUN pip3 install datasets pandas pathlib
+
+COPY load_chatdoctor_dataset.py load_chatdoctor_dataset.py

scenarios/llm-finetune/ci/Dockerfile.medqa

@@ -0,0 +1,10 @@
+FROM ubuntu:20.04
+
+ENV DEBIAN_FRONTEND="noninteractive"
+
+RUN apt-get upgrade && apt-get update \
+    && apt-get install -y python3 python3-pip
+
+RUN pip3 install datasets pandas pathlib
+
+COPY load_medqa_dataset.py load_medqa_dataset.py

scenarios/llm-finetune/ci/Dockerfile.medquad

@@ -0,0 +1,10 @@
+FROM ubuntu:20.04
+
+ENV DEBIAN_FRONTEND="noninteractive"
+
+RUN apt-get upgrade && apt-get update \
+    && apt-get install -y python3 python3-pip
+
+RUN pip3 install datasets pandas pathlib
+
+COPY load_medquad_dataset.py load_medquad_dataset.py

scenarios/llm-finetune/ci/Dockerfile.modelsave

@@ -0,0 +1,21 @@
+FROM ubuntu:20.04
+
+ENV DEBIAN_FRONTEND="noninteractive"
+
+RUN apt-get update && apt-get -y upgrade \
+    && apt-get install -y curl \
+    && apt-get install -y python3.9 python3.9-dev python3.9-distutils
+
+## Install pip
+RUN curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
+RUN python3.9 get-pip.py
+
+# Install CPU-only version of PyTorch
+RUN pip3 install torch --index-url https://download.pytorch.org/whl/cpu
+# RUN pip3 install torch
+RUN pip3 install transformers peft bitsandbytes pathlib
+# RUN pip3 install transformers peft pathlib
+
+# ENV CUDA_VISIBLE_DEVICES=""
+
+COPY load_base_model.py load_base_model.py

scenarios/llm-finetune/ci/build.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+docker build -f ci/Dockerfile.medqa src -t preprocess-medqa:latest
+docker build -f ci/Dockerfile.chatdoctor src -t preprocess-chatdoctor:latest
+docker build -f ci/Dockerfile.medquad src -t preprocess-medquad:latest
+docker build -f ci/Dockerfile.modelsave src -t ccr-model-save:latest

scenarios/llm-finetune/ci/push-containers.sh

@@ -0,0 +1,6 @@
+containers=("preprocess-medqa:latest" "preprocess-chatdoctor:latest" "preprocess-medquad" "ccr-model-save:latest")
+for container in "${containers[@]}"
+do
+  docker tag $container $CONTAINER_REGISTRY"/"$container
+  docker push $CONTAINER_REGISTRY"/"$container
+done

scenarios/llm-finetune/config/chatdoctor_config.json

@@ -0,0 +1,5 @@
+{
+    "DATASET_NAME": "lavita/ChatDoctor-HealthCareMagic-100k",
+    "SAVE_DIR": "/mnt/output/chatdoctor/",
+    "FORMAT": "csv"
+}

scenarios/llm-finetune/config/medqa_config.json

@@ -0,0 +1,6 @@
+{
+    "DATASET_NAME": "Malikeh1375/medical-question-answering-datasets",
+    "DATASET_SPLIT": "all-processed",
+    "SAVE_DIR": "/mnt/output/medqa/",
+    "FORMAT": "csv"
+}

scenarios/llm-finetune/config/medquad_config.json

@@ -0,0 +1,5 @@
+{
+    "DATASET_NAME": "keivalya/MedQuad-MedicalQnADataset",
+    "SAVE_DIR": "/mnt/output/medquad/",
+    "FORMAT": "csv"
+}

scenarios/llm-finetune/config/model_repo_config.json

@@ -0,0 +1,10 @@
+{
+    "MODEL_NAME": "facebook/opt-350m",
+    "REPO_READ_TOKEN": "",
+    "SAVE_DIR": "mnt/model",
+    "Q_PREC": "load_in_4bit",
+    "REPO_WRITE_TOKEN": "",
+    "REPO_URL": "https://huggingface.co/your_username/your_model_repo",
+    "MODEL_DESCRIPTION": "A fine-tuned model for medical question answering.",
+    "MODEL_LICENSE": "apache-2.0"
+}

scenarios/llm-finetune/config/pipeline_config.json

@@ -0,0 +1,72 @@
+{
+    "pipeline": [
+      {
+        "name": "TextJoin",
+        "config": {
+          "datasets": [
+            {
+              "id": "19517ba8-bab8-11ed-afa1-0242ac120002",
+              "name": "medqa",
+              "file": "medical-question-answering-datasets.csv",
+              "select_variables": ["input", "output"],
+              "num_rows": 200,
+              "mount_path": "/mnt/remote/medqa/"
+            },
+            {
+              "id": "216d5cc6-bab8-11ed-afa1-0242ac120002",
+              "name": "chatdoctor",
+              "file": "ChatDoctor-HealthCareMagic-100k.csv",
+              "select_variables": ["input", "output"],
+              "num_rows": 200,
+              "mount_path": "/mnt/remote/chatdoctor/"
+            },
+            {
+              "id": "2830a144-bab8-11ed-afa1-0242ac120002",
+              "name": "medquad",
+              "file": "MedQuad-MedicalQnADataset.csv",
+              "select_variables": ["Question", "Answer"],
+              "num_rows": 200,
+              "mount_path": "/mnt/remote/medquad/"
+            }
+          ],
+          "joined_dataset": {
+            "output_folder": "/tmp/",
+            "output_file": "medqa_chatdoctor_medquad_joined.csv"
+          }
+        }
+      },
+      {
+        "name": "PrivateLLMFineTune",
+        "config": {
+          "device": "cpu",
+          "saved_model_dir": "/mnt/remote/model",
+          "model_name": "facebook/opt-350m",
+          "trained_model_output_path": "/mnt/remote/output",
+          "input_dataset_path": "/tmp/medqa_chatdoctor_medquad_joined.csv",
+          "sample_prompts": [
+            "What is the treatment for diabetes?",
+            "How can I manage my hypertension?",
+            "What are the symptoms of asthma?",
+            "What is the best diet for heart health?"
+          ],
+          "MODEL_REPO_CONFIG": "location/to/model_repo_config",
+          "MAX_TOKENS": 64,
+          "BATCH_SIZE": 4,
+          "NUM_EPOCHS": 1,
+          "Q_PRECISION": null,
+          "LEARNING_RATE": 5e-5,
+          "EPSILON": 7.5,
+          "DELTA": 1e-5,
+          "MAX_GRAD_NORM": 1.0,
+          "MAX_PHYSICAL_BATCH_SIZE": 4,
+          "LORA_RANK": 8,
+          "LORA_ALPHA": 32,
+          "LORA_TARGET_MODULES": null,
+          "LORA_DROPOUT": 0.05,
+          "LORA_BIAS": "none",
+          "LOG_FREQ_STEPS": 50
+        }
+      }
+    ]
+  }
+

scenarios/llm-finetune/contract/contract.json

@@ -0,0 +1,92 @@
+{
+    "id": "f4f72a88-bab1-11ed-afa1-0242ac120002",
+    "schemaVersion": "0.1",
+    "startTime": "2023-03-14T00:00:00.000Z",
+    "expiryTime": "2024-03-14T00:00:00.000Z",
+    "tdc": "",
+    "tdps": [],
+    "ccrp": "did:web:ccrprovider.github.io",
+    "datasets": [
+      {
+        "id": "19517ba8-bab8-11ed-afa1-0242ac120002",
+        "name": "medqa",
+        "url": "https://ccrcontainer.blob.core.windows.net/medqa/data.img",
+        "provider": "",
+        "key": {
+          "type": "azure",
+          "properties": {
+            "kid": "MEDQAFilesystemEncryptionKey",
+            "authority": {
+              "endpoint": "sharedneu.neu.attest.azure.net"
+            },
+            "endpoint": ""
+          }
+        }
+      },
+      {
+        "id": "216d5cc6-bab8-11ed-afa1-0242ac120002",
+        "name": "chatdoctor",
+        "url": "https://ccrcontainer.blob.core.windows.net/chatdoctor/data.img",
+        "provider": "",
+        "key": {
+          "type": "azure",
+          "properties": {
+            "kid": "CHATDOCTORFilesystemEncryptionKey",
+            "authority": {
+              "endpoint": "sharedneu.neu.attest.azure.net"
+            },
+            "endpoint": ""
+          }
+        }
+      },
+      {
+        "id": "2830a144-bab8-11ed-afa1-0242ac120002",
+        "name": "medquad",
+        "url": "https://ccrcontainer.blob.core.windows.net/medquad/data.img",
+        "provider": "",
+        "key": {
+          "type": "azure",
+          "properties": {
+            "kid": "MEDQUADFilesystemEncryptionKey",
+            "authority": {
+              "endpoint": "sharedneu.neu.attest.azure.net"
+            },
+            "endpoint": ""
+          }
+        }
+      }
+    ],
+    "purpose": "TRAINING",
+    "constraints": [
+      {
+        "privacy": [
+          {
+            "dataset": "19517ba8-bab8-11ed-afa1-0242ac120002",
+            "epsilon_threshold": "1.5",
+            "noise_multiplier": "2.0",
+            "delta": "0.01",
+            "epochs_per_report": "2"
+          },
+          {
+            "dataset": "216d5cc6-bab8-11ed-afa1-0242ac120002",
+            "epsilon_threshold": "1.5",
+            "noise_multiplier": "2.0",
+            "delta": "0.01",
+            "epochs_per_report": "2"
+          },
+          {
+            "dataset": "2830a144-bab8-11ed-afa1-0242ac120002",
+            "epsilon_threshold": "1.5",
+            "noise_multiplier": "2.0",
+            "delta": "0.01",
+            "epochs_per_report": "2"
+          }
+        ]
+      }
+    ],
+    "terms": {
+      "payment": {},
+      "revocation": {}
+    }
+  }
+

scenarios/llm-finetune/deployment/docker/docker-compose-modelsave.yml

@@ -0,0 +1,7 @@
+services:
+  model_save:
+    image: ${CONTAINER_REGISTRY:+$CONTAINER_REGISTRY/}ccr-model-save:latest
+    volumes:
+      - $MODEL_OUTPUT_PATH:/mnt/model
+      - $CONFIG_PATH:/mnt/config/model_repo_config.json
+    command: ["python3.9", "load_base_model.py"]

scenarios/llm-finetune/deployment/docker/docker-compose-preprocess.yml

@@ -0,0 +1,19 @@
+services:
+  medqa:
+    image: ${CONTAINER_REGISTRY:+$CONTAINER_REGISTRY/}preprocess-medqa:latest
+    volumes:
+      - $MEDQA_OUTPUT_PATH:/mnt/output/medqa
+      - $CONFIG_PATH:/mnt/config/
+    command: ["python3", "load_medqa_dataset.py"]
+  chatdoctor:
+    image: ${CONTAINER_REGISTRY:+$CONTAINER_REGISTRY/}preprocess-chatdoctor:latest
+    volumes:
+      - $CHATDOCTOR_OUTPUT_PATH:/mnt/output/chatdoctor
+      - $CONFIG_PATH:/mnt/config/
+    command: ["python3", "load_chatdoctor_dataset.py"]
+  medquad:
+    image: ${CONTAINER_REGISTRY:+$CONTAINER_REGISTRY/}preprocess-medquad:latest
+    volumes:
+      - $MEDQUAD_OUTPUT_PATH:/mnt/output/medquad
+      - $CONFIG_PATH:/mnt/config/
+    command: ["python3", "load_medquad_dataset.py"]