CIFAR-10 ResNet50 from scratch
What it is
A ten-class image classifier using a custom CIFAR-style ResNet50-v1 architecture with 3-4-6-3 bottleneck stages, a 3×3 stride-1 stem, no initial max pool, global average pooling and a softmax head. The model is the author's own training and is released under MIT.
Training
Trained from scratch, custom CIFAR-style ResNet50-v1, random init.
Dataset and split
CIFAR-10 with 45,000 training, 5,000 validation and 10,000 test images. CIFAR-10 is by Alex Krizhevsky and Geoffrey Hinton, University of Toronto; dataset source and licence details are to be verified.
Preprocessing and input
Input is 32×32 RGB. Pixel values are converted to float32 and scaled from 0–255 to 0–1, matching the training notebook.
Result
Test accuracy: 85.12% — from saved notebook output, not re-run.
Loading
import keras
model = keras.saving.load_model("resnet50_cifar10_final.keras")
Limitations
This is coursework and was not re-run offline for this model card. Its evaluation is limited to the saved CIFAR-10 test result, and broader robustness was not assessed. This model is not for production use.
Repository: https://github.com/faiz-sheikh/Deep-Learning
- Downloads last month
- 24