Skip to content
This repository has been archived by the owner on Mar 30, 2021. It is now read-only.

Latest commit

 

History

History
90 lines (49 loc) · 3.48 KB

README.md

File metadata and controls

90 lines (49 loc) · 3.48 KB
Raw

StreetScroe2.0

Acknowledgement

The raw votes data is from data/votes/safe.csv from https://github.com/abhimanyudubey/dlcity

The model is from ECCV 2016 paper "Deep Learning the City: Quantifying Urban Perception At A Global Scale" https://arxiv.org/abs/1608.01769

adopted Code from https://github.com/abhimanyudubey/dlcity

the pre-trained weight VGG_ILSVRC_19_layers.caffemodel from Caffe Model Zoo' s ILSVRC-2014 model (VGG team) with 19 weight layers

Download caffemodel url

How to get data

  1. Convert Raw Data

    Use python yeg_src/convert_format.py to convert raw data to get training images list input and list of images with empty url.

  2. Get Image URLs and Download

    Fetch the urls for images and download

    • Option 1:

      1. Put Google Static Streeview API Key and Secret in:
        yeg_src/image_download/googleapikey.txt
        yeg_src/image_download/googleapisecret.txt

      2. Run cp yeg_data\images_list_emptyurl.csv yeg_data\images_list_osc.csv

      3. Use python yeg_src/image_download_osc/fetch_osc_api_url.py to get urls for images from Google Streetview

      4. Use yeg_src/image_downlaod_osc/fetchImagesfromURL_multithread.py to download from the url got from Google Streetview

      5. Use yeg_src/cleanup_invalid_images.py to clean up

    • Option 2:

      1. Run cp yeg_data\images_list_emptyurl.csv yeg_data\images_list.csv

      2. Use python yeg_src/image_download_osc/fetch_osc_api_url.py to get urls for images from OpenStreetCam

      3. Use python yeg_src/image_download/fetchImagesfromURL_multithread.py to download from the url got from OpenStreetCam

      4. Use yeg_src/cleanup_invalid_images.py to clean up

Caffe

Read more about mean image and how to install and train with caffe at "Training With Your Own Dataset on Caffe"

Installation instructions

Mean Image

use yeg_src/create_lmdb.py to get the lmdb for the training images.

The model requires us to subtract the image mean from each image, so we have to compute the mean. caffe/tools/compute_image_mean.cpp implements that. This file is at yeg_src/compute_image_mean.cpp

Run <caffe installation folder>/build/tools/compute_image_mean <lmdb folder> data/ilsvrc12/imagenet_mean.binaryproto

imagenet_mean.binaryproto is the mean image file we need.

Training

use models/trial_model_1/train.sh to train with caffe

make sure the paths are all correct in
models/trial_model_1/train.sh,
models/trial_model_1/prototxt/solver.prototxt, and
models/trial_model_1/prototxt/train_val_trial.prototxt

Stop and Continue Training

when you want to end the training process and save the current weights, you CAN'T simply ctrl+c to terminate it, becasue it won't save the weights.

Instead run kill -s SIGINT <caffe_pid>

This checks the ID of Caffe process (hint: read the log - PID is the number right before the filename) and manually sending a signal to Caffe itself

use the saved weights to continue training with caffe with caffe's --weights flag