Unfreeze model and allocate 3 blocks of Resnet50 by default to retrain

README.md

@@ -94,6 +94,19 @@ evaluation:
   iou_threshold: 0.5
 ```
 
+---
+
+## 📝 **Training Strategies for Faster R-CNN with ResNet-50 Backbone**
+
+| Trainable Backbone Layers | Epochs | Batch Size | Recommended Learning Rate | Optimizer | Scheduler         |
+|--------------------------|--------|-----------|--------------------------|-----------|------------------|
+| 0                        | 10     | 4         | 0.0100                   | SGD       | StepLR(3, 0.1)   |
+| 3                        | 10     | 8         | 0.0050                   | SGD       | StepLR(3, 0.1)   |
+| 5                        | 10     | 16        | 0.0001                   | AdamW     | CosineAnnealing  |
+| 3                        | 20     | 8         | 0.0050                   | SGD       | StepLR(5, 0.1)   |
+| 5                        | 20     | 16        | 0.0001                   | AdamW     | CosineAnnealing  |
+
+
 ---
 
 ## 📡 **Evaluating the Model**

config.yaml

@@ -2,11 +2,12 @@ model:
   backbone: resnet50
   num_classes: 2
   pretrained: true
-  learning_rate: 0.0001
-  epochs: 20
+  learning_rate:  0.0050 
+  epochs: 10
   batch_size: 8
   optimizer: adam
   image_sample_size: 10
+  trainable_backbone_layers: 3
 
 dataset:
   image_dir: datasets/images

src/train.py

@@ -15,11 +15,11 @@ with open("config.yaml", "r") as f:
     config = yaml.safe_load(f)
 
 # Load Pretrained Model
-model = torchvision.models.detection.fasterrcnn_resnet50_fpn(weights="DEFAULT")
+model = torchvision.models.detection.fasterrcnn_resnet50_fpn(weights="DEFAULT", trainable_backbone_layers=config["model"]["trainable_backbone_layers"])
 
-# Freeze Backbone
-for param in model.backbone.parameters():
-    param.requires_grad = False
+# # Freeze Backbone
+# for param in model.backbone.parameters():
+#     param.requires_grad = False
 
 # Modify Predictor
 num_classes = config["model"]["num_classes"]