103 classifier 2d template restructure

2D_Classifier/train/Dockerfile → project/classifier_2d/Dockerfile

@@ -5,7 +5,7 @@ WORKDIR /train
 ENV DEBIAN_FRONTEND=noninteractive
 
 RUN apt-get update && \
-    apt-get install -y --no-install-recommends build-essential git rsync software-properties-common ffmpeg libsm6 libxext6 && \
+    apt-get install -y --no-install-recommends build-essential git rsync ffmpeg libsm6 libxext6 && \
     rm -rf /var/lib/apt/lists/*
 
 ENV PYTHONPATH="/mlflow/projects/code/:$PYTHONPATH"

2D_Classifier/train/README.md → project/classifier_2d/README.md

@@ -1,3 +1,5 @@
+# 2D Classifier THIS NEEDS UPDATING
+
 ## Example Workflow:
 
 #### 1. Adapt XNATDataImport.py for your data
@@ -58,3 +60,5 @@ This dockerfile sets up the Docker image that the MLOps run will utilise.
 
 In the example this is just a simple environment running python version 3.10.
 You will most likely need to adapt this for your project.
+
+

2D_Classifier/train/scripts/train.py → project/classifier_2d/mlops/train.py

@@ -1,21 +1,18 @@
-import sys
 import configparser
+import json
 import logging
-import os
 import multiprocessing
-import json
+import os
 
 import mlflow
 import pytorch_lightning as pl
 import torch
-from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
-from ray.air.integrations.mlflow import setup_mlflow
+from pytorch_lightning.callbacks import EarlyStopping, LearningRateMonitor, ModelCheckpoint
 from torch.cuda import is_available as cuda_available
 
-from src.DataModule import DataModule
-from src.Network import Network
-from src.DataModule import label_dict
-from src.XNATDataImport import XNATDataImport
+from src.data_import_xnat import DataImportXNAT
+from src.datamodule import DataModule, label_dict
+from src.network import Network
 
 logger = logging.getLogger(__name__)
 
@@ -36,7 +33,7 @@ def train(config):
         else multiprocessing.cpu_count()
     )
 
-    importer = XNATDataImport(
+    importer = DataImportXNAT(
         xnat_configuration = xnat_configuration,
         num_workers = num_workers
         )
@@ -48,14 +45,6 @@ def train(config):
     data = importer.xnat_image_download(raw_data)
 
     # Set up mflow experiment
-    setup_mlflow(
-        tracking_uri=mlflow.get_tracking_uri(),
-        experiment_id=mlflow.get_experiment_by_name(
-            config["project"]["name"]
-        ).experiment_id
-        if mlflow.get_experiment_by_name(config["project"]["name"])
-        else mlflow.create_experiment(config["project"]["name"]),
-    )
     with mlflow.start_run(nested=True):
         save_best_model = True
 
@@ -93,6 +82,10 @@ def train(config):
             label_smoothing = float(config['params']['label_smoothing']),
         )
 
+        # Callbacks
+        checkpoint_metric = config['params']['checkpoint_metric']
+        checkpoint_mode = "min" if checkpoint_metric == "val_loss" else "max"
+
         # Callbacks
         callbacks = []
         callbacks.append(LearningRateMonitor(logging_interval="step"))
@@ -104,6 +97,14 @@ def train(config):
         )
         callbacks.append(checkpoint_callback)
 
+        early_stopping_callback = EarlyStopping(
+            monitor=checkpoint_metric,
+            patience=10,
+            mode=checkpoint_mode,
+            verbose=True,
+        )
+        callbacks.append(early_stopping_callback)
+
         # configure trainer
         trainer = pl.Trainer(
             precision="32" if cuda_available() else "16",

2D_Classifier/train/scripts/tune.py → project/classifier_2d/mlops/tune.py

@@ -1,22 +1,18 @@
-
 import configparser
 import logging
 import multiprocessing
 import os
 
 import mlflow
+import optuna
 import pytorch_lightning as pl
-from ray.air.integrations.mlflow import setup_mlflow
-from pytorch_lightning.callbacks import LearningRateMonitor
-from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint, EarlyStopping
 from torch.cuda import is_available as cuda_available
 
-from project.DataModule import DataModule
-from project.DataModule import label_dict
-from project.Network import Network
-from project.XNATDataImport import XNATDataImport
+from src.data_import_xnat import DataImportXNAT
+from src.datamodule import DataModule, label_dict
+from src.network import Network
 
-import optuna
 logger = logging.getLogger(__name__)
 
 # Obtain hyperparameters for this trial
@@ -105,6 +101,14 @@ def objective(trial,data,config):
         )
         callbacks.append(checkpoint_callback)
 
+        early_stopping_callback = EarlyStopping(
+            monitor="val_loss",
+            patience=10,
+            mode="min",
+            verbose=True,
+        )
+        callbacks.append(early_stopping_callback)
+
         # configure trainer
         trainer = pl.Trainer(
             precision="32" if cuda_available() else "16",
@@ -147,7 +151,7 @@ def tune(config):
         else multiprocessing.cpu_count()
     )
 
-    importer = XNATDataImport(
+    importer = DataImportXNAT(
         xnat_configuration = xnat_configuration,
         num_workers = num_workers
         )
@@ -158,20 +162,11 @@ def tune(config):
     # Download images from XNAT
     data = importer.xnat_image_download(raw_data)
 
-    # Set up mflow experiment
-    setup_mlflow(
-        tracking_uri=mlflow.get_tracking_uri(),
-        experiment_id=mlflow.get_experiment_by_name(
-            config["project"]["name"]
-        ).experiment_id
-        if mlflow.get_experiment_by_name(config["project"]["name"])
-        else mlflow.create_experiment(config["project"]["name"]),
-    )
 
     mlflow.pytorch.autolog(log_models=False)
 
     # Create optuna study (hyperparameter tuning framework)
-    study = optuna.create_study(study_name="scaphx-tune", direction="minimize")
+    study = optuna.create_study(study_name="project-tune", direction="minimize")
     study.optimize(lambda trial: objective(trial, data, config), n_trials=50)
 
     with open(('tune_log.txt'), 'w') as f:

2D_Classifier/train/src/XNATDataImport.py → project/classifier_2d/src/data_import_xnat.py

@@ -2,16 +2,16 @@
 from tqdm import tqdm
 from typing import List
 
-from utils.tools import DataBuilderXNAT
+from monai.data import Dataset
+from torch.utils.data import DataLoader
 from xnat.mixin import ImageScanData, SubjectData
 
 from src.transforms import load_xnat
-from monai.data import Dataset
-from torch.utils.data import DataLoader
+from src.utils.tools import DataBuilderXNAT
 
 logger = logging.getLogger(__name__)
 
-class XNATDataImport():
+class DataImportXNAT():
 
     def __init__(self, xnat_configuration: dict = None, num_workers: int = 4, test_batch: int = 0,
                  n_month_data_window=9999, run_type: str='train'):

2D_Classifier/train/src/DataModule.py → project/classifier_2d/src/datamodule.py

@@ -1,21 +1,19 @@
 import logging
 from collections import Counter
 from typing import List, Optional
+
 import mlflow
 import numpy as np
 import pytorch_lightning
 import torch
-
-from monai.data import CacheDataset, Dataset
-from monai.data import pad_list_data_collate
+from monai.data import CacheDataset, Dataset, pad_list_data_collate
 from monai.transforms import Compose
-
 from sklearn.model_selection import train_test_split
 from torch.cuda import is_available
 from torch.utils.data import DataLoader
 
-from src.transforms import normalise, train_augment, output 
-from src.transforms.SafeWrapper import SafeWrapperTransform
+from src.transforms import normalise, output, train_augment
+from src.transforms.safe_wrapper import SafeWrapperTransform
 
 logger = logging.getLogger(__name__)
 

2D_Classifier/train/src/Network.py → project/classifier_2d/src/network.py

@@ -1,25 +1,22 @@
 import logging
-import pytorch_lightning
+from abc import ABC
+
 import mlflow
+import numpy as np
+import pytorch_lightning
 import torch
-from abc import ABC
 from monai.data import decollate_batch
-from monai.transforms import (
-    AsDiscrete,
-    Compose,
-    Activations,
-)
-from sklearn.metrics import classification_report, confusion_matrix, ConfusionMatrixDisplay, recall_score
+from monai.transforms import Activations, AsDiscrete, Compose
+from sklearn.metrics import ClassificationReport, ConfusionMatrixDisplay, confusion_matrix, recall_score
 from timm import create_model
 from timm.data import Mixup
 from torch.nn import CrossEntropyLoss
 from torchmetrics import Accuracy, F1Score
-from src.DataModule import label_dict
 from torchmetrics.classification import MulticlassAUROC
-import numpy as np
 
-logger = logging.getLogger(__name__)
+from src.datamodule import label_dict
 
+logger = logging.getLogger(__name__)
 
 class Network(pytorch_lightning.LightningModule, ABC):
     """

2D_Classifier/train/src/transforms/__init__.py → project/classifier_2d/src/transforms/__init__.py

@@ -1,29 +1,31 @@
 import torch
 from monai.transforms import (
-    LoadImage,
-    SqueezeDimd,
-    EnsureChannelFirstd,
-    CropForegroundd,
-    Resized,
-    ScaleIntensityd,
     CastToTyped,
-    RandFlipd,
-    RandZoomd,
-    RandRotated,
+    CropForegroundd,
+    EnsureChannelFirstd,
+    EnsureTyped,
+    LoadImage,
+    NormalizeIntensityd,
+    RandAdjustContrastd,
     RandAffined,
+    RandCoarseDropoutd,
+    RandFlipd,
     RandGaussianNoised,
     RandGaussianSmoothd,
+    RandRotated,
     RandScaleIntensityd,
-    RandAdjustContrastd,
-    RandCoarseDropoutd,
+    RandZoomd,
     ResizeWithPadOrCropd,
-    ToTensord,
+    Resized,
+    ScaleIntensityd,
+    ScaleIntensityRangePercentilesd,
     SelectItemsd,
-    EnsureTyped,
     Spacingd,
+    SqueezeDimd,
+    ToTensord,
 )
 
-from src.transforms.LoadImageXNATd import LoadImageXNATd
+from src.transforms.load_image_xnatd import LoadImageXNATd
 
 def load_xnat(xnat_configuration: dict):
     """
@@ -47,9 +49,8 @@ def normalise(image_size):
         EnsureChannelFirstd(keys=['image']),
         CropForegroundd(keys=['image'], source_key='image'),
         Resized(keys=['image'], size_mode='longest', spatial_size=image_size+20),
-        #Maybe limit top intensity in case of big spikes?
-        ScaleIntensityd(keys=["image"], minv=0.0, maxv=255.0),
-        CastToTyped(keys=["image"], dtype=torch.uint8),
+        ScaleIntensityRangePercentilesd(keys=["image"], lower=0, upper=99, b_min=0.0, b_max=255.0, clip=True),
+        CastToTyped(keys=["image"], dtype=torch.float32),
     ]
 
 def train_augment(image_size):
@@ -59,20 +60,20 @@ def train_augment(image_size):
     """
     return [
         RandFlipd(keys=['image'], spatial_axis=0, prob=0.5),
-        RandZoomd(keys=['image'], prob=0.2, min_zoom=1.05,max_zoom=1.1),
-        RandRotated(keys=['image'], prob=0.2, range_x=0.4),
-        RandAffined(keys=['image'], prob=0.2, padding_mode='zeros'),
-        RandGaussianNoised(keys=['image'], prob=0.1, mean=0.0, std=0.1),
-        RandGaussianSmoothd(keys=['image'], prob=0.2, sigma_x=(0.5,1.0)),
-        RandScaleIntensityd(keys=['image'], prob=0.15, factors=(0.75,1.25)),
-        RandAdjustContrastd(keys=['image'], prob=0.1, gamma=(0.5,2), retain_stats=True, invert_image=True),
-        RandAdjustContrastd(keys=['image'], prob=0.3, gamma=(0.5,2), retain_stats=True, invert_image=False),
+        RandZoomd(keys=['image'], prob=0.4, min_zoom=1.05,max_zoom=1.1),
+        RandRotated(keys=['image'], prob=0.4, range_x=0.4),
+        RandAffined(keys=['image'], prob=0.3, padding_mode='zeros'),
+        RandGaussianNoised(keys=['image'], prob=0.3, mean=0.0, std=10.0),
+        RandGaussianSmoothd(keys=['image'], prob=0.35, sigma_x=(0.5,1.0), sigma_y=(0.5,1.0)),
+        RandScaleIntensityd(keys=['image'], prob=0.3, factors=(0.75,1.25)),
+        RandAdjustContrastd(keys=['image'], prob=0.2, gamma=(0.5,2), retain_stats=True, invert_image=True),
+        RandAdjustContrastd(keys=['image'], prob=0.4, gamma=(0.5,2), retain_stats=True, invert_image=False),
         ResizeWithPadOrCropd(
             keys=["image"],
             spatial_size=(image_size,image_size),
             mode='replicate'
         ),
-        RandCoarseDropoutd(keys=['image'], prob=0.5, fill_value=0, holes=8, max_holes=16, spatial_size=(10,10), max_spatial_size=(36,36)),
+        RandCoarseDropoutd(keys=['image'], prob=0.35, fill_value=0, holes=8, max_holes=16, spatial_size=(10,10), max_spatial_size=(15,15)),
     ]
 
 def output(image_size):
@@ -86,6 +87,9 @@ def output(image_size):
             mode='replicate'
         ),
         ScaleIntensityd(keys=["image"], minv=0.0, maxv=1),
+        # Normalize with grayscale-averaged ImageNet stats (mean=0.449, std=0.226)
+        # Required for pretrained ImageNet models
+        NormalizeIntensityd(keys=["image"], subtrahend=0.449, divisor=0.226),
         ToTensord(keys=['image', 'label']),
         SelectItemsd(keys=['subject_id', 'image', 'label']),
         EnsureTyped(keys=['image', 'label'], track_meta=False),