detect/yolo11-p2.yaml

[github-actions[bot]]

Discussion for detect/yolo11-p2.yaml

📊 Metadata

👤 author: @Y-T-G
🎯 task: detect
🔑 keywords: cnn, yolo11, small-object
📝 description: YOLO11 object detection model with an extra P2 scale for small object detecton.
⚡ flops:
- n: 11.4
- s: 29.7
- m: 88.9
- l: 113.6
- x: 245.1

🔢 parameters:
- n: 2,740,032
- s: 9,625,968
- m: 20,591,120
- l: 26,116,368
- x: 57,849,456

💻 min_version: 8.3.0

If you found this config helpful, consider upvoting this discussion! 🔼

[Y-T-G]

Transferring pretrained weights from Ultralytics models

import torch
import torch.nn as nn
from collections import OrderedDict

def transfer_weights_sequential(model1_state_dict, model2, strict_matching=True):
    """
    Transfer weights from model1 to model2 where model2 has additional layers inserted.
    
    Args:
        model1_state_dict: State dict from the source model (Model 1)
        model2: Target model (Model 2) with inserted layers
        strict_matching: If True, requires exact parameter shape matching
    
    Returns:
        model2 with transferred weights
    """
    # Get model2's state dict
    model2_state_dict = model2.state_dict()
    
    # Extract layer types and shapes for mapping
    def get_layer_info(state_dict):
        layer_info = []
        for key, tensor in state_dict.items():
            # Extract layer number and parameter type (weight/bias)
            parts = key.split('.')
            if len(parts) >= 2:
                layer_num = int(parts[0])
                param_type = parts[1]  # 'weight' or 'bias'
                layer_info.append((layer_num, param_type, tensor.shape, key))
        return sorted(layer_info)
    
    model1_info = get_layer_info(model1_state_dict)
    model2_info = get_layer_info(model2_state_dict)
    
    # Create mapping between compatible layers
    mapping = {}
    model1_idx = 0
    
    for model2_idx, (m2_layer, m2_param, m2_shape, m2_key) in enumerate(model2_info):
        if model1_idx < len(model1_info):
            m1_layer, m1_param, m1_shape, m1_key = model1_info[model1_idx]
            
            # Check if parameters match (same type and shape)
            if m1_param == m2_param and m1_shape == m2_shape:
                mapping[m2_key] = m1_key
                model1_idx += 1
            # If shapes don't match but we're not in strict mode, skip this model1 layer
            elif not strict_matching and m1_param == m2_param:
                print(f"Skipping incompatible layer: {m1_key} {m1_shape} -> {m2_key} {m2_shape}")
                model1_idx += 1
    
    # Apply the mapping
    new_state_dict = model2_state_dict.copy()
    transferred_count = 0
    
    for m2_key, m1_key in mapping.items():
        new_state_dict[m2_key] = model1_state_dict[m1_key]
        transferred_count += 1
    
    # Load the new state dict
    model2.load_state_dict(new_state_dict)
    
    print(f"Successfully transferred {transferred_count} parameters")
    print(f"Model 1 had {len(model1_info)} parameters")
    print(f"Model 2 has {len(model2_info)} parameters")
    
    return model2

from ultralytics import YOLO
model = YOLO("community/cfg/detect/yolo11n-p2.yaml")
pretrained = YOLO("yolo11n.pt")
model.model.model = transfer_weights_sequential(pretrained.model.model.state_dict(), model.model.model)

model.save("yolo11n-p2.pt")  # use this checkpoint for training