oversampling ok

README.md

@@ -4,10 +4,6 @@ datasets:
 - frugal-ai-challenge/public-leaderboard-image
 ---
 
-# Notices
-
-- **Stratification is not used for data split** (splits are predefined in project)
-
 # Installation
 
 ✅ Checklist:
@@ -16,14 +12,11 @@ datasets:
 
 # 🚧 TODO
 
-- faire data augmentation en amont pour **no_smoke oversampling**
+- 👉 randomly blur & flip smoke images
 - réduire **n_epochs**
 - voir répartition partenaires, caméras, temporalité, annotations
 - métriques : **matrice de confusion** complète
 - décrire voir **erreurs de types et conséquences**
 - tester plusieurs pré-entraînements (est-ce que charger un modèle pré-entraîné ImageNet aide vraiment ?)
-- modèles :
-  - [ ] EfficientNet
-  - [ ] EfficientDet
-- tester si amélioration avec et sans égalisation
+- tester si amélioration inférence avec et sans égalisation
 - voir répartition physique des annotations sur l'image

config.yaml

@@ -14,13 +14,4 @@ models_common:
   models_root_dir: "models"
   classes: ["no_smoke", "smoke"]
   log_dir: "runs"
-  chkpts_dir : "chkpts"
-# model_EffNet:
-#   model_name: "EfficientNetB0"
-#   input_size: (224, 224)
-#   batch_size: 48
-#   n_epochs: 100
-#   loss: "binary_crossentropy"
-#   metrics: ["accuracy"]
-#   optimizer: AdamW(learning_rate=0.0002, weight_decay=0.05)
-#   trained_model_uri: "EfficientNetB0_app.keras"
+  chkpts_dir : "chkpts"

src/load_data.py

@@ -4,6 +4,7 @@ from datasets import load_dataset
 import logging
 import os
 import pandas as pd
+from PIL import Image, ImageOps, ImageEnhance, ImageFilter
 import shutil
 import subprocess
 import yaml
@@ -201,15 +202,78 @@ def format_data_keras(df):
     return df
 
 
-def oversample_class(df):
-    """Oversample an under-represented class"""
-    count_df = df.groupby(["split", "label"]).size().reset_index(name="count")
-    count_df = count_df.loc[count_df["split"] != "val"]
+def add_data_aug(aug_name, df_sample, df_aug):
+    """Add data augmentation for under-represented class"""
+    # Rename images and update URI
+    df_sample.loc[:, "name"] += f"_DA-{aug_name[:-4]}"
+    df_sample.rename(columns={'uri': 'input_uri'}, inplace=True)
+    df_sample.loc[:, "uri"] = df_sample["input_uri"].str.replace(".jpg", f"_DA-{aug_name[:-4]}.jpg")
 
+    # Transform and export image
+    for _, row in df_sample.iterrows():
+        img = Image.open(row['input_uri'])
 
+        if aug_name == "blur_img":
+            img_aug = img.filter(ImageFilter.GaussianBlur(radius=2))
+        elif aug_name == "flip_img":
+            img_aug = img.transpose(Image.FLIP_LEFT_RIGHT)
+        elif aug_name == "blur_flip_img":
+            img_aug = img.filter(ImageFilter.GaussianBlur(radius=2)).transpose(Image.FLIP_LEFT_RIGHT)
+        elif aug_name == "eq_img":
+            img_aug = ImageOps.equalize(img)
+        elif aug_name == "180_img":
+            img_aug = img.rotate(180)
+        else:
+            logging.warn("Wrong data augmentation name: passing")
+            return df_aug
+        
+        img_aug.save(row['uri'])
 
+    # Add to dataframe
+    df_sample.drop(columns='input_uri', inplace=True)
+    result = pd.concat([df_aug, df_sample], axis=0, ignore_index=True)
 
-    return df
+    logging.info(f"✅ Transformed {len(df_sample)} images with {aug_name}")
+
+    return result
+
+
+def oversample_class(df):
+    """Oversample an under-represented class"""
+    count_df = df.groupby(["split", "label"]).size().reset_index(name="count")
+    count_df = count_df.loc[count_df["split"] != "test"]
+
+    df_aug = df.copy()
+
+    for split in count_df.split.unique():
+        logging.info(f"⚙️ Processing {split} split...")
+        _ = count_df.loc[count_df["split"] == split].reset_index(drop=True)
+        # Minimum label
+        idxmin = _["count"].idxmin()
+        min_row = _.loc[idxmin, :]
+        min_label = min_row["label"]
+        min_count = min_row["count"]
+        # Maximum label
+        idxmax = _["count"].idxmax()
+        max_row = _.loc[idxmax, :]
+        max_count = max_row["count"]
+        # Needed labels
+        need = max_count - min_count
+        
+        logging.info(f"⚙️ Transforming {need} images...")
+
+        # Loop over augmentation techniques until need is covered
+        for aug_name in ["blur_img", "flip_img", "blur_flip_img", "eq_img", "180_img"]:
+            # Create a new DF sample from which images will be transformed
+            df_sample = df.loc[(df["split"] == split) & (df["label"] == min_label)].copy()
+
+            if need < len(df_sample):
+                df_sample = df_sample.iloc[:need]
+
+            df_aug = add_data_aug(aug_name, df_sample, df_aug)
+            need -= len(df_sample)
+
+    return df_aug
 
 
 if __name__ == "__main__":