Goal of this project was to exploit my ISS database picture, coming from my ISS-HDEV-wallpaper project. These pictures was taken from the ISS HDEV live. Most of them was posted on instagram. I decided to cluster images in order to identify which images could be cool to post or to find out which cluster are ugly ...
Unfortunately, HDEV stopped sending any data on July 18, 2019, it was declared, on August 22, 2019, to have reached its end of life... :'(
My new goal aim to create nice poster composed of different kind of cluster i found.
Project based on the cookiecutter data science project template. #cookiecutterdatascience
I use docker, see docker-compose.yaml file. Most of my routines are in Makefile file.
make docker_start
make docker_stop
I usually start a console inside my jupyter container (tensorflow jhub)
make docker_bash
And then, initialize environment
make requirements
I use visual studio code outside my container. To execute some code, i use console and type for example:
python -m iss.exec.bdd
# or
make populate_db
To use vscode debug, i use ptvsd
make debug src.exec.bdd
Configuration file of project is in config/config_<env>.yaml. See template in source.
Root directory contains a .env file with some environment variables
MODE=dev
PROJECT_DIR="/home/jovyan/work"
The MODE value will load config/config_MODE.yaml configuration. See iss/tools/config.py.
ISS images are stored online on personal server, i need to collect all of them (>14k images).
make sync_collections
i used data/raw/collections directory.
i have an history of location of ISS for some images, i store it in data/raw/history/history.txt
12.656456313474;-75.371420423828;20180513-154001;Caribbean Sea
-43.891574367708;-21.080797293704;20180513-160001;South Atlantic Ocean
-10.077472167643;-82.562993796116;20180513-172001;South Pacific Ocean
-51.783078834111;-3.9925568092913;20180513-174001;South Atlantic Ocean
27.255631526786;-134.89231579188;20180513-184001;North Pacific Ocean
See extract of config/config_dev.yml:
directory:
project_dir: ${PROJECT_DIR}
data_dir: ${PROJECT_DIR}/data
collections: ${PROJECT_DIR}/data/raw/collections
isr_dir: ${PROJECT_DIR}/data/isr
I use mysql database running in container to store in table :
- locations: history file
- embedding: clustering results
make populate_db
adminer is running to monitor mysql db
My clustering consist in using auto encoder to define a latent representation of my images. Latent representation are then use in a classical clustering.
I split into train, test and validation set
make sampling
See extract of config/config_dev.yml:
sampling:
autoencoder:
seed: 37672
proportions:
train: 0.7
test: 0.2
valid: 0.1
directory:
from: collections
base: ${PROJECT_DIR}/data/processed/models/autoencoder
train: ${PROJECT_DIR}/data/processed/models/autoencoder/train/k
test: ${PROJECT_DIR}/data/processed/models/autoencoder/test/k
valid: ${PROJECT_DIR}/data/processed/models/autoencoder/valid/k
Newbie here, i tried home made models:
- simple auto encoder:
iss/models/SimpleAutoEncoder.py - simple convolutional auto encoder:
iss/models/SimpleConvAutoEncoder.py<- model selected - Variational auto encoder:
iss/models/VariationalAutoEncoder.py - Variational convolutional auto encoder:
iss/models/VariationaConvlAutoEncoder.py
See extract of config/config_dev.yml:
models:
simple_conv:
save_directory: ${PROJECT_DIR}/models/simple_conv
model_name: model_dev
sampling: autoencoder
input_width: 48
input_height: 27
input_channel: 3
latent_width: 6
latent_height: 3
latent_channel: 16
learning_rate: 0.001
epochs: 2
batch_size: 128
verbose: 0
initial_epoch: 0
workers: 1
use_multiprocessing: false
steps_per_epoch: 4
validation_steps: 2
validation_freq:
activation: sigmoid
callbacks:
csv_logger:
directory: ${PROJECT_DIR}/models/simple_conv/log
append: true
checkpoint:
directory: ${PROJECT_DIR}/models/simple_conv/checkpoint
verbose: 1
period: 20
tensorboard:
log_dir: ${PROJECT_DIR}/models/simple_conv/tensorboard
limit_image: 5
floyd: True
I create simple training framework and launch it with:
make training
# or
python -m iss.exec.training --model-type=simple_conv
Actually, i use floydhub to train my models.
i added a floyd.yml file in root directory containing something like this:
env: tensorflow-1.12
task:
training:
input:
- destination: /iss-autoencoder
source: prise6/datasets/iss/1
machine: gpu
description: training autoencoder (simple_conv)
command: mv .env-floyd .env && make training
training_prod:
input:
- destination: /iss-autoencoder
source: prise6/datasets/iss/1
machine: gpu2
description: training autoencoder (simple_conv)
command: mv .env-floyd .env && make training
i use a special config file for floydhub so i provide a different .env file.
Training dashboard and dataset are public and available here
make floyd_training_prod
I tested google colab and train the final model with it, but result are really similar to the floydhub model.
Having fun with different approachs:
- Classical Clustering (PCA + kmeans + Hierarchical clustering):
iss/clustering/ClassicalClustering.py - Advanced Clustering:
iss/clustering/AdvancedClustering.py(no really used) - Not2Deep Clustering (see paper):
iss/clustering/N2DClustering.py<- selected - DbScan Clustering:
iss/clustering/DBScanClustering.py(no really used)
Clustering are trained onver a sample ~2.5k images. I create 50 clusters in order to find clusters with very similar images.
make exec_clustering
Parameters are in config/config_dev.yml:
clustering:
n2d:
version: 3
model:
type: 'simple_conv'
name: 'model_colab'
umap:
random_state: 98372
metric: euclidean
n_components: 2
n_neighbors: 5
min_dist: 0
kmeans:
n_clusters: 50
random_state: 883302
save_directory: ${PROJECT_DIR}/models/clustering/n2d
I save umap/t-sne embedding of latent space to plot it with bokeh:
(screenshot)
I populate my embedding mysql table with iss/exec/bdd.py script
Compute silhouette score on latent representation for every cluster to see quality.
Example of 0.2 silhouette score (cluster 1):
Another example of 0.2 silhouette score (cluster 39):
We see why it's low, but we detect the pattern why it's gathered.
Example of 0.8 silhouette score (cluster 10):
live is off, easy to cluster.
Example of negative silhouette score (cluster 35):
A bit messy.
See all mosaic cluster on imgur.
Let's try facets on this dataset ! Thanks to mysql db i can compare different clustering and visualize it with facets-dive.
make facets
Two html page are created in the directory reports/.
You can manipulate all your images:
Bin by cluster:
And zoom on it:
It's a bit messy because you cannot filter your data ... but the sprite trick make it fast!
Generate multiple posters based on a template. See poster examples on top.
make posters
original size : 1280x720 cut size : 48x27
inch = 2.54cm 150dpi 300dpi http://www.altelia.fr/actualites/calculateur-resolution-definition-format.htm https://fr.wikipedia.org/wiki/Point_par_pouce
2880px X 4320px donne : (en 150 dpi)
48.77cm x 73.15cm
bordure de 60px pour 1 cm en 150dpi
version2 :
- e63e25b
- df89064
- d4cb94
- c2ff00
- 2e58ed9 : ok
- 2b11acbe
- 2575f6
- 1b4cb13: ok