MNIST — Train a Classifier in One Epoch

This project demonstrates how to train a strong MNIST digit classifier in a single epoch using PyTorch. The workflow lives in the notebook MNIST_Training_in_one_epoch.ipynb.

---

✨ Highlights

• One-epoch training: reach high accuracy with the right optimizer, LR schedule, and regularization.
• Simple, readable PyTorch: minimal boilerplate, easy to tweak.
• Reproducible: fixed seeds and deterministic settings included.

---

📐 Model Architecture

Network (PyTorch)

• Input: [N, 1, 28, 28] (MNIST grayscale)
• All convolutions: kernel_size=3, stride=1, padding=0
• Pooling: MaxPool2d(kernel_size=2, stride=2)
• Activations: ReLU
• Classifier: Linear → ReLU → Linear (10 logits)

Layer-by-layer (with shapes)

Stage	Layer	In → Out shape
Input	—	[N, 1, 28, 28]
Block 1	Conv2d(1→4, 3×3) + ReLU	[N, 1, 28, 28] → [N, 4, 26, 26]
	Conv2d(4→8, 3×3) + ReLU	[N, 4, 26, 26] → [N, 8, 24, 24]
	MaxPool2d(2,2)	[N, 8, 24, 24] → [N, 8, 12, 12]
Block 2	Conv2d(8→16, 3×3) + ReLU	[N, 8, 12, 12] → [N, 16, 10, 10]
	Conv2d(16→32, 3×3) + ReLU	[N, 16, 10, 10] → [N, 32, 8, 8]
	MaxPool2d(2,2)	[N, 32, 8, 8] → [N, 32, 4, 4]
Flatten	reshape(-1, 32*4*4)	[N, 32, 4, 4] → [N, 512]
Head	Linear(512→30) + ReLU	[N, 512] → [N, 30]
	Linear(30→10)	[N, 30] → **[N, 10]** (logits)

Note: No softmax in the model — use nn.CrossEntropyLoss (it applies log_softmax internally). Apply softmax only at inference if you need probabilities.

---

🔢 Parameter Count (exact)

• conv1 (1→4, 3×3): 4·1·3·3 + 4 = 40
• conv2 (4→8, 3×3): 8·4·3·3 + 8 = 296
• conv3 (8→16, 3×3): 16·8·3·3 + 16 = 1,168
• conv4 (16→32, 3×3): 32·16·3·3 + 32 = 4,640
• fc1 (512→30): 512·30 + 30 = 15,390
• out (30→10): 30·10 + 10 = 310

Total parameters: 21,844

(Optional) quick check in code

total = sum(p.numel() for p in model.parameters())
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(total, trainable)  # both should be 21844

---

📊 Metrics

(Add these after your run. Template below.)

• Train accuracy (1 epoch): XX.X%
• Validation accuracy (1 epoch): XX.X%
• Test accuracy (final): XX.X%
• Loss (train/val): X.XXX / X.XXX
• Confusion matrix: (insert image or table)

Snippets to log metrics

# Accuracy helper
def accuracy(logits, targets):
    return (logits.argmax(dim=1) == targets).float().mean().item()

# After each epoch or at end:
print(f"Train Acc: {train_acc:.2%} | Val Acc: {val_acc:.2%}")

# Confusion matrix (PyTorch)
import torch
from sklearn.metrics import confusion_matrix
import numpy as np

model.eval()
all_preds, all_tgts = [], []
with torch.no_grad():
    for xb, yb in test_loader:
        logits = model(xb.to(device))
        all_preds.append(logits.argmax(1).cpu())
        all_tgts.append(yb.cpu())
cm = confusion_matrix(torch.cat(all_tgts), torch.cat(all_preds))
print(cm)

---

⚙️ Recommended one-epoch settings (for good results fast)

• Optimizer: AdamW(lr=1e-3, weight_decay=0.01)
• Scheduler: OneCycleLR(max_lr=1e-3, epochs=1, steps_per_epoch=len(train_loader), pct_start=0.3)
• Batch size: 128–256
• Loss: CrossEntropyLoss(label_smoothing=0.1)
• Light aug (optional): small rotations/affine transforms