improve bioclip baseline, run script, gitignore (#37)

.gitignore

@@ -1,5 +1,6 @@
 sample_result_submission/
 input_data/
+reference_data/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

baselines/BioCLIP_code_submission/model.py

@@ -4,8 +4,7 @@
 - predict: uses the model to perform predictions.
 - load: reloads the model.
 '''
-from open_clip import create_model
-from torchvision import transforms
+from open_clip import create_model_and_transforms
 import torch
 import pickle
 import os
@@ -18,20 +17,14 @@ def __init__(self):
     def load(self):
         self.device='cuda' if torch.cuda.is_available() else 'cpu'
 
-        model = create_model("hf-hub:imageomics/bioclip", output_dict=True, require_pretrained=True)
+        model, _, preprocess_val = create_model_and_transforms("hf-hub:imageomics/bioclip", precision="amp", output_dict=True)
+        model.eval()
         self.model = model.to(self.device)
+        self.preprocess_img = preprocess_val
 
         with open(os.path.join(os.path.dirname(__file__), "clf.pkl"), "rb") as f:
             self.clf = pickle.load(f)
 
-        self.preprocess_img = transforms.Compose(
-            [
-                transforms.ToTensor(),
-                transforms.Resize((224, 224), interpolation=transforms.InterpolationMode.BICUBIC),
-            ]
-        )
-
-
     def predict(self, datapoint):
 
         with torch.no_grad():

baselines/BioCLIP_code_submission/requirements.txt

@@ -1,4 +1,4 @@
 open-clip-torch==2.24.0
 torch==2.3.0
 torchvision==0.18.0
-scikit-learn==1.3.2
+scikit-learn==1.4.2

run.sh

@@ -2,27 +2,36 @@
 : <<'END_COMMENT'
 1. 
 1.a. If run in docker:
-docker pull icreateadockerid/anomaly_challenge
-docker run -it -v [repo path]:/codabench icreateadockerid/anomaly_challenge:cpu /bin/bash
+
+docker pull [image_id]
+
+1.a.1. If use a GPU:
+docker run -it --gpus device=0 -v [repo path]:/codabench [image_id] /bin/bash
+1.a.2. If only use CPU:
+docker run -it -v [repo path]:/codabench [image_id] /bin/bash
+
 cd codabench
 
 1.b. If run with conda env:
+
 1.b.1. create running env with conda, venv, etc.
 conda create --name [name] python=3.10
 conda activate [name]
 1.b.2. install all necessary dependencies
 pip install pillow==10.3.0 tqdm==4.66.4 pandas==2.2.2 scikit-learn==1.4.2
 
 2. edit and run the script
+
 chmod +x run.sh
+
 ./run.sh
 END_COMMENT
 
 
 export task_type="folder; predict; evaluate" #folder; predict; evaluate
-export data_split="test"
+export data_split="dev"
 export baseline_model="bioclip"
-export task_folder="$data_split_$baseline_model"
+export task_folder="${data_split}_${baseline_model}"
 
 
 ## create folder structure
@@ -57,12 +66,12 @@ if [[ "$task_type" == *"predict"* ]]; then
         exit 1
     fi
 
-    python ingestion_program/ingestion.py $input_dir $output_dir $program_dir $submission_dir
+    python3 ingestion_program/ingestion.py $input_dir $output_dir $program_dir $submission_dir
 fi
 
 ## score the predictions
 if [[ "$task_type" == *"evaluate"* ]]; then
     export input_dir="sample_result_submission/$task_folder"
     export output_dir="sample_result_submission/$task_folder"
-    python scoring_program/score_combined.py $input_dir $output_dir
-fi
+    python3 scoring_program/score_combined.py $input_dir $output_dir
+fi