About

This this the code repository for my final project of MSc Data Science and Machine Learning degree at UCL. It's better to check out this repo with the thesis.

Introduction

In this project, we explore training strategies for deep architectures of Computer Vision. Especially, we focus on "ImageNet pre-training", and "multi-task learning" and "regularization" are also applied.

Requirements

• Python 3
• PyTorch >= 0.4.0
• Numpy

Data

In this project, we use several datasets for semantic segmentation and human part segmentation. See and modify their paths in config.json for your convenience.

• Pascal Voc 2012
• Semantic Boundary Dataset
• Look Into Person Dataset

Usage

python train_imagenet.py [--arch ARCH] [--epochs N]
                         [--dataset D] [--data_portion DP]
                         [--batch_size N] [--lr LR] [--momentum M]
                         [--momentum_bn M_BN] [--weight-decay W]
                         [--pretrained] [--dprob DPB]
                         [--n_classes C] [--optim O]
                         [--manual_seed MANUALSEED]

--arch, -a            model architecture: sunet64 | fcn32resnet18 |
                      fcn32resnet 34 | fcn32resnet50 | fcn32resnet101 |   
                      fcn32resnet152 | fcn32vgg
--dataset             dataset: sbd | pascal | pascalpart | lip
--epochs              number of epochs to train
--batch-size          mini-batch size (default: 10)
--lr                  initial learning rate
--momentum            momentum
--momentum_bn         momentum of batch normalization
--wd                  weight decay (default: 1e-4)
--pretrained          use pre-trained model
--manual_seed         manual seed 
--n_classes           number of classes of ground truth annotation
--dprob               dropout probability