MLOps-Vishal_Kishore-B23CS1078

Deep Learning Assignments - Vishal Kishore

This repository contains machine learning assignments and projects for MLOps (B23CS1078).

Assignment 1: Deep Learning Model Training on MNIST and FashionMNIST

🔗 Quick Links

Resource	Link
📓 Colab Notebook	Open in Colab
📄 Report	View PDF
💾 Best Model	Download
📊 Results CSV	View

📋 Experiment Configuration

Parameter	Value
Datasets	MNIST, FashionMNIST
Models	ResNet-18, ResNet-34, ResNet-50 (pretrained=False)
Data Split	70%-10%-20% (Train-Val-Test)
Mixed Precision	USE_AMP = True
Early Stopping	5 epochs
Seeds	35, 78, 13

📊 Q1(a): Classification Results

MNIST Dataset - 20 Epochs

Batch Size	Optimizer	LR	ResNet-18 (No Sched.)	ResNet-18 (Cosine)	ResNet-50 (Cosine)
16	SGD	0.001	99.20 ± 0.10%	99.32 ± 0.04%	99.07 ± 0.07%
16	SGD	0.0001	98.96 ± 0.09%	98.90 ± 0.18%	98.50 ± 0.04%
16	Adam	0.001	99.23 ± 0.04%	99.40 ± 0.05%	99.26 ± 0.09%
16	Adam	0.0001	99.12 ± 0.14%	99.22 ± 0.19%	99.19 ± 0.10%
32	SGD	0.001	98.63 ± 0.10%	99.18 ± 0.09%	99.01 ± 0.03%
32	SGD	0.0001	99.12 ± 0.11%	98.57 ± 0.11%	97.99 ± 0.08%
32	Adam	0.001	99.30 ± 0.11%	99.45 ± 0.07%	99.27 ± 0.11%
32	Adam	0.0001	99.12 ± 0.14%	99.22 ± 0.19%	98.96 ± 0.10%

MNIST Dataset - 10 Epochs

Batch Size	Optimizer	LR	ResNet-18 (No Sched.)	ResNet-18 (Cosine)	ResNet-50 (No Sched.)	ResNet-50 (Cosine)
16	SGD	0.001	99.14 ± 0.05%	99.20 ± 0.06%	98.81 ± 0.10%	98.99 ± 0.05%
16	SGD	0.0001	98.84 ± 0.04%	98.76 ± 0.06%	98.32 ± 0.10%	98.12 ± 0.11%
16	Adam	0.001	99.23 ± 0.04%	99.45 ± 0.06%	98.90 ± 0.10%	99.30 ± 0.10%
16	Adam	0.0001	99.10 ± 0.04%	99.30 ± 0.07%	98.53 ± 0.09%	98.89 ± 0.15%
32	SGD	0.001	99.13 ± 0.09%	99.14 ± 0.10%	98.74 ± 0.11%	98.92 ± 0.03%
32	SGD	0.0001	98.55 ± 0.13%	98.44 ± 0.04%	97.68 ± 0.09%	97.48 ± 0.11%
32	Adam	0.001	99.09 ± 0.04%	99.41 ± 0.03%	98.77 ± 0.05%	99.27 ± 0.11%
32	Adam	0.0001	99.01 ± 0.09%	99.25 ± 0.19%	98.27 ± 0.12%	98.64 ± 0.05%

FashionMNIST Dataset - 20 Epochs

Batch Size	Optimizer	LR	ResNet-18 (Cosine)	ResNet-50 (Cosine)
16	SGD	0.001	91.73 ± 0.06%	90.40 ± 0.50%
16	SGD	0.0001	90.12 ± 0.37%	88.66 ± 0.36%
16	Adam	0.001	92.04 ± 0.07%	91.90 ± 0.11%
16	Adam	0.0001	91.42 ± 0.09%	88.66 ± 0.36%
32	SGD	0.001	91.28 ± 0.22%	89.73 ± 0.16%
32	SGD	0.0001	89.44 ± 0.04%	87.40 ± 0.36%
32	Adam	0.001	92.12 ± 0.09%	91.69 ± 0.26%
32	Adam	0.0001	91.18 ± 0.21%	89.87 ± 0.19%

FashionMNIST Dataset - 10 Epochs

Batch Size	Optimizer	LR	ResNet-18 (No Sched.)	ResNet-18 (Cosine)	ResNet-50 (No Sched.)	ResNet-50 (Cosine)
16	SGD	0.001	90.89 ± 0.29%	91.61 ± 0.14%	89.50 ± 0.31%	90.04 ± 0.64%
16	SGD	0.0001	89.70 ± 0.04%	88.71 ± 0.15%	87.04 ± 0.23%	86.70 ± 0.49%
16	Adam	0.001	91.34 ± 0.13%	92.28 ± 0.09%	89.24 ± 1.59%	91.90 ± 0.15%
16	Adam	0.0001	90.80 ± 0.27%	91.24 ± 0.13%	87.82 ± 0.49%	88.37 ± 0.69%
32	SGD	0.001	90.47 ± 0.22%	91.05 ± 0.36%	88.48 ± 0.53%	89.02 ± 0.36%
32	SGD	0.0001	88.93 ± 0.29%	88.80 ± 0.37%	85.75 ± 0.16%	85.41 ± 0.30%
32	Adam	0.001	91.03 ± 0.13%	92.22 ± 0.19%	88.81 ± 0.46%	91.47 ± 0.37%
32	Adam	0.0001	90.19 ± 0.20%	90.67 ± 0.40%	87.32 ± 0.37%	88.65 ± 0.46%

📊 Q1(b): SVM Classifier Results

MNIST - RBF Kernel

C	Gamma	Degree	Accuracy (%)	Training Time (ms)
1	scale	3	94.45	7993.38
5	scale	3	95.40	7411.84
10	scale	3	95.55	8001.53
1	0.01	3	93.25	7175.44
5	0.01	3	95.10	9403.81
10	0.01	3	95.45	6162.32

MNIST - Polynomial Kernel

C	Gamma	Degree	Accuracy (%)	Training Time (ms)
1	scale	2	94.15	6056.24
5	scale	2	95.05	5633.33
10	scale	2	95.20	5204.85
1	scale	3	93.40	8414.45
5	scale	3	94.00	6237.90
10	scale	3	94.10	5884.44

FashionMNIST - RBF Kernel

C	Gamma	Degree	Accuracy (%)	Training Time (ms)
1	scale	3	86.00	10334.46
5	scale	3	87.45	6820.15
10	scale	3	87.35	6757.70
1	0.01	3	86.00	7202.21
5	0.01	3	87.45	6617.26
10	0.01	3	87.35	6687.58

FashionMNIST - Polynomial Kernel

C	Gamma	Degree	Accuracy (%)	Training Time (ms)
1	scale	2	84.75	6769.21
5	scale	2	86.60	5886.71
10	scale	2	86.80	6729.06
1	scale	3	82.80	8146.13
5	scale	3	85.35	6939.22
10	scale	3	85.75	6051.58

📊 Q2: CPU vs GPU Performance Comparison

FashionMNIST Dataset - CPU vs GPU Analysis

Compute	Batch Size	Optimizer	LR	R-18 Acc	R-34 Acc	R-50 Acc	R-18 Time (ms)	R-34 Time (ms)	R-50 Time (ms)
CPU	16	SGD	0.001	91.74 ± 0.06%	91.65 ± 0.38%	90.39 ± 0.50%	7,274,210	1,879,214	2,619,201
CPU	16	Adam	0.001	92.01 ± 0.06%	92.08 ± 0.25%	91.90 ± 0.10%	7,679,113	1,867,889	2,699,217
GPU	16	SGD	0.001	91.73 ± 0.06%	91.65 ± 0.38%	90.40 ± 0.50%	545,392	923,336	1,663,622
GPU	16	Adam	0.001	92.04 ± 0.07%	92.08 ± 0.25%	91.90 ± 0.11%	533,329	1,025,271	1,663,071

FLOPs Analysis

Compute Device	Batch Size	ResNet-18 FLOPs	ResNet-34 FLOPs	ResNet-50 FLOPs
CPU	16	0.03G	0.07G	0.08G
GPU	16	0.03G	0.07G	0.08G

🏆 Best Model Summary

Dataset	Best Model	Configuration	Test Accuracy
MNIST	ResNet-18	Adam, LR=0.001, Batch=32, Cosine, 20 Epochs	99.45 ± 0.07%
FashionMNIST	ResNet-18	Adam, LR=0.001, Batch=16, Cosine, 10 Epochs	92.28 ± 0.09%

Assignment 2: ResNet-18 on CIFAR-10

This project trains a ResNet-18 model on the CIFAR-10 dataset using PyTorch. It monitors training progress, logs metrics to WandB, and visualizes gradient flow and weight updates.

Model Statistics

Architecture: ResNet-18 (customized for CIFAR-10)
Parameters: 11.174M
FLOPs: 557.889M

Dataset Split

Train: 40,000 images
Validation: 10,000 images
Test: 10,000 images

Training Results

Best Validation Accuracy: 92.15% (Epoch 24)

Training Progression

| Epoch | Train Loss | Train Acc | Val Loss | Val Acc | Learning Rate | |——-|————|———–|———-|———|—————| | 1 | 2.1214 | 24.38% | 1.7948 | 35.39% | 0.0996 | | 5 | 0.9609 | 66.13% | 0.9116 | 67.24% | 0.0905 | | 10 | 0.5221 | 82.01% | 0.5502 | 80.83% | 0.0655 | | 15 | 0.3397 | 88.39% | 0.4178 | 85.76% | 0.0345 | | 20 | 0.1604 | 94.50% | 0.3117 | 89.95% | 0.0095 | | 24 | 0.0759 | 97.51% | 0.2576 | 92.15% | 0.0004 | | 25 | 0.0682 | 97.92% | 0.2596 | 92.09% | 0.0000 |