Cache v0.4 update

[glenn-jocher]

Cache saving and loading migrated to numpy from torch for faster speeds. Profiling results with COCO:

            t0 = time.time()
            for i in range(3):
                np.save('test.npy', cache, allow_pickle=True)
            t1 = time.time()
            for i in range(3):
                x = np.load('test.npy', allow_pickle=True)
            t2 = time.time()
            print(t2 - t1, t1 - t0)

            t0 = time.time()
            for i in range(3):
                torch.save(cache, 'test.pt')
            t1 = time.time()
            for i in range(3):
                x = torch.load('test.pt')
            t2 = time.time()
            print(t2 - t1, t1 - t0)

Results indicate read and write speedup of about 2X over baseline.

2.449079990386963 3.9117438793182373
4.935853958129883 8.460411071777344

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Improved caching system for dataset labels in YOLOv5.

📊 Key Changes

• Switched caching from PyTorch to NumPy to save and load the labels cache.
• Cache version updated from 0.3 to 0.4.
• Added an assertion check to validate cache version and hash.
• Solved an issue where the .npy suffix had to be removed from the cache filename after saving.

🎯 Purpose & Impact

• 🚀 Ensures more robust loading of the dataset cache by checking for cache integrity.
• 🛠️ Fixes the cache path by cleaning up the filename, enhancing compatibility and order.
• ⚡ Improves the potential for easier debugging and maintenance with the assertion method on cache loading.
• 💾 The change in caching method may offer performance benefits and creates consistency in the cache file format.