add label studio compliant taxonomy

Author: bw4sz

conf/classification_model/finetune.yaml

@@ -6,7 +6,7 @@ crop_image_dir: /blue/ewhite/b.weinstein/BOEM/classification/crops/
 under_sample_ratio: 0
 trainer:
   fast_dev_run: False
-  max_epochs: 1
+  max_epochs: 10
   lr: 0.00001
   batch_size: 16
   workers: 10

conf/config.yaml

@@ -34,7 +34,7 @@ predict:
   patch_size: 1000
   patch_overlap: 0
   min_score: 0.4
-  batch_size: 32
+  batch_size: 48
 
 pipeline:
   confidence_threshold: 0.9
@@ -54,10 +54,10 @@ detection_model:
   labels:
     - "Object"
   trainer: 
-    batch_size: 12
+    batch_size: 16
     train:
       fast_dev_run: False
-      epochs: 20
+      epochs: 10
       lr: 0.00001
     workers: 10
     validation:
@@ -86,7 +86,7 @@ active_learning:
   evaluation: 
   dask_client: 
   pool_limit: 500
-  gpus: 1
+  gpus: 2
 
 active_testing:
   image_dir: /blue/ewhite/b.weinstein/BOEM/sample_flight/JPG_2024_Jan27

src/active_learning.py

@@ -122,7 +122,7 @@ def human_review(predictions, min_score=0.1, confident_threshold=0.5):
 
     return confident_predictions, uncertain_predictions
 
-def generate_pool_predictions(image_dir, patch_size=512, patch_overlap=0.1, min_score=0.1, model=None, model_path=None, dask_client=None, batch_size=16, comet_logger=None, pool_limit=1000):
+def generate_pool_predictions(image_dir, patch_size=512, patch_overlap=0.1, min_score=0.1, model=None, model_path=None, dask_client=None, batch_size=16, comet_logger=None, pool_limit=1000, crop_model=None):
     """
     Generate predictions for the training pool.
     
@@ -136,6 +136,7 @@ def generate_pool_predictions(image_dir, patch_size=512, patch_overlap=0.1, min_
         dask_client (dask.distributed.Client, optional): A Dask client for parallel processing. Defaults to None.
         batch_size (int, optional): The batch size for prediction. Defaults to 16.
         comet_logger (CometLogger, optional): A CometLogger object. Defaults to None.
+        crop_model (bool, optional): A deepforest.model.CropModel object. Defaults to None.
         pool_limit (int, optional): The maximum number of images to consider. Defaults to 1000.
     
     Returns:
@@ -168,7 +169,7 @@ def update_sys_path():
         blocks = dask_pool.to_delayed().ravel()
         block_futures = []
         for block in blocks:
-            block_future = dask_client.submit(detection.predict, image_paths=block.compute(), patch_size=patch_size, patch_overlap=patch_overlap, model_path=model_path)
+            block_future = dask_client.submit(detection.predict, image_paths=block.compute(), patch_size=patch_size, patch_overlap=patch_overlap, model_path=model_path, crop_model=crop_model)
             block_futures.append(block_future)
         # Get results
         dask_results = []
@@ -177,7 +178,7 @@ def update_sys_path():
             dask_results.append(pd.concat(block_result))
         preannotations = pd.concat(dask_results)
     else:
-        preannotations = detection.predict(m=model, image_paths=pool, patch_size=patch_size, patch_overlap=patch_overlap, batch_size=batch_size)
+        preannotations = detection.predict(m=model, image_paths=pool, patch_size=patch_size, patch_overlap=patch_overlap, batch_size=batch_size, crop_model=crop_model)
         preannotations = pd.concat(preannotations)
 
     if comet_logger:

src/classification.py

@@ -128,8 +128,16 @@ def preprocess_images(model, annotations, root_dir, save_dir):
     # Remove any negative values
     annotations = annotations[(annotations['xmin'] >= 0) & (annotations['ymin'] >= 0) & (annotations['xmax'] >= 0) & (annotations['ymax'] >= 0)]
     boxes = annotations[['xmin', 'ymin', 'xmax', 'ymax']].values.tolist()
+
+    # Expand by 20 pixels on all sides
+    boxes = [[box[0]-20, box[1]-20, box[2]+20, box[3]+20] for box in boxes]
+    
+    # Make sure no negative values
+    boxes = [[max(0, box[0]), max(0, box[1]), max(0, box[2]), max(0, box[3])] for box in boxes]
+    
     images = annotations["image_path"].values
     labels = annotations["label"].values
+    
     model.write_crops(boxes=boxes, root_dir=root_dir, images=images, labels=labels, savedir=save_dir)
 
 def preprocess_and_train_classification(config, train_df=None, validation_df=None, comet_logger=None):

src/detection.py

@@ -31,7 +31,8 @@ def evaluate(model, test_csv, image_root_dir):
     """
     # create trainer
     devices = torch.cuda.device_count()
-    model.create_trainer(num_nodes=1, devices=devices)
+    strategy = "ddp" if devices > 1 else None
+    model.create_trainer(num_nodes=1, devices=devices, strategy=strategy)
     model.config["validation"]["csv_file"] = test_csv
     model.config["validation"]["root_dir"] = image_root_dir
     results = model.trainer.validate(model)
@@ -163,32 +164,28 @@ def train(model, train_annotations, test_annotations, train_image_dir, comet_log
                 model.config[key] = value
 
     devices = torch.cuda.device_count()
-    if comet_logger:
-        comet_logger.experiment.log_parameters(model.config)
-        comet_logger.experiment.log_table("train.csv", train_annotations)
-        comet_logger.experiment.log_table("test.csv", test_annotations)
-        model.create_trainer(logger=comet_logger, num_nodes=1, devices=devices)
-    else:
-        model.create_trainer(num_nodes=1, devices=devices)
-
-    with comet_logger.experiment.context_manager("train_images"):
-        non_empty_train_annotations = read_file(model.config["train"]["csv_file"], root_dir=train_image_dir)
-        # Sanity check for debug
-        n = 5 if non_empty_train_annotations.shape[0] > 5 else non_empty_train_annotations.shape[0]
-        for filename in non_empty_train_annotations.image_path.sample(n=n).unique():
-            sample_train_annotations_for_image = non_empty_train_annotations[non_empty_train_annotations.image_path == filename]
-            sample_train_annotations_for_image.root_dir = train_image_dir
-            visualize.plot_annotations(sample_train_annotations_for_image, savedir=tmpdir)
-            comet_logger.experiment.log_image(os.path.join(tmpdir, filename))
-    
-    with comet_logger.experiment.context_manager("test_images"):
-        non_empty_validation_annotations = read_file(model.config["validation"]["csv_file"], root_dir=train_image_dir)
-        n = 5 if non_empty_validation_annotations.shape[0] > 5 else non_empty_validation_annotations.shape[0]
-        for filename in non_empty_validation_annotations.image_path.sample(n=n).unique():
-            sample_validation_annotations_for_image = non_empty_validation_annotations[non_empty_validation_annotations.image_path == filename]
-            sample_validation_annotations_for_image.root_dir = train_image_dir
-            visualize.plot_annotations(sample_validation_annotations_for_image, savedir=tmpdir)
-            comet_logger.experiment.log_image(os.path.join(tmpdir, filename))
+    strategy = "ddp" if devices > 1 else None
+    comet_logger.experiment.log_parameters(model.config)
+    comet_logger.experiment.log_table("train.csv", train_annotations)
+    comet_logger.experiment.log_table("test.csv", test_annotations)
+    model.create_trainer(logger=comet_logger, num_nodes=1, accelerator="gpu", strategy="ddp", devices=2)
+
+    non_empty_train_annotations = read_file(model.config["train"]["csv_file"], root_dir=train_image_dir)
+    # Sanity check for debug
+    n = 5 if non_empty_train_annotations.shape[0] > 5 else non_empty_train_annotations.shape[0]
+    for filename in non_empty_train_annotations.image_path.sample(n=n).unique():
+        sample_train_annotations_for_image = non_empty_train_annotations[non_empty_train_annotations.image_path == filename]
+        sample_train_annotations_for_image.root_dir = train_image_dir
+        visualize.plot_annotations(sample_train_annotations_for_image, savedir=tmpdir)
+        comet_logger.experiment.log_image(os.path.join(tmpdir, filename),metadata={"name":filename,"context":'train_images'})
+
+    non_empty_validation_annotations = read_file(model.config["validation"]["csv_file"], root_dir=train_image_dir)
+    n = 5 if non_empty_validation_annotations.shape[0] > 5 else non_empty_validation_annotations.shape[0]
+    for filename in non_empty_validation_annotations.image_path.sample(n=n).unique():
+        sample_validation_annotations_for_image = non_empty_validation_annotations[non_empty_validation_annotations.image_path == filename]
+        sample_validation_annotations_for_image.root_dir = train_image_dir
+        visualize.plot_annotations(sample_validation_annotations_for_image, savedir=tmpdir)
+        comet_logger.experiment.log_image(os.path.join(tmpdir, filename),metadata={"name":filename,"context":'validation_images'})
 
     with comet_logger.experiment.context_manager("detection"):
         model.trainer.fit(model)

src/pipeline.py

@@ -162,7 +162,8 @@ def run(self):
             model_path=detection_checkpoint_path,
             dask_client=dask_client,
             batch_size=self.config.predict.batch_size,
-            comet_logger=self.comet_logger
+            comet_logger=self.comet_logger,
+            crop_model=trained_classification_model
         )
         self.comet_logger.experiment.log_table(tabular_data=training_pool_predictions, filename="training_pool_predictions.csv")
 

submit.sh

@@ -10,9 +10,10 @@
 #SBATCH --output=/home/b.weinstein/logs/BOEM%j.out   # Standard output and error log
 #SBATCH --error=/home/b.weinstein/logs/BOEM%j.err
 #SBATCH --partition=gpu
-#SBATCH --gpus=1
+#SBATCH --ntasks-per-node=2
+#SBATCH --gpus=2
 
 source activate BOEM
 
 cd ~/BOEM/
-srun python main.py check_annotations=True active_learning.pool_limit=10 active_testing.n_images=1 active_learning.n_images=1 ++classification_model.trainer.fast_dev_run=True ++detection_model.trainer.train.fast_dev_run=True debug=True
+srun python main.py check_annotations=True active_learning.pool_limit=10 active_testing.n_images=1 active_learning.n_images=1 debug=True