Data Purchasing Challenge 2022

Apply the right Normalisation transformation for your data

Apply the right normalisation transformations to different datasets.


People usually get carried away with the main implementation and forget about the minimalistic pre-processing step. Many a time we copy the image transformation steps from Google. Yes, I know it happens, but there is a very simple way to know the normalization which is your dataset-specific. In this short notebook, you'll learn how to get the means and standard deviations of training, unlabelled, and validation datasets. 

About this notebook 📝

Image normalization is one of the crucial thing to do in computer vision tasks. This notebook is rather a very small one that will give an short summary on different datasets.

1) Login to AIcrowd 🤩

In [4]:
#@title Login to AIcrowd
!pip install -U aicrowd-cli > /dev/null
!aicrowd login 2> /dev/null
Please login here: https://api.aicrowd.com/auth/zo-wY53l5nP45gOH4OlcrGquu0Axok4c22AdzkgTWac
API Key valid
Gitlab access token valid
Saved details successfully!

2) Setup magically, run the below cell 😉

In [5]:
#@title Magic Box ⬛ { vertical-output: true, display-mode: "form" }
  import os
  if first_run and os.path.exists("/content/data-purchasing-challenge-2022-starter-kit/data/training"):
    first_run = False
  first_run = True

if first_run:
  %cd /content/
  !git clone http://gitlab.aicrowd.com/zew/data-purchasing-challenge-2022-starter-kit.git > /dev/null
  %cd data-purchasing-challenge-2022-starter-kit
  !aicrowd dataset list -c data-purchasing-challenge-2022
  !aicrowd dataset download -c data-purchasing-challenge-2022
  !mkdir -p data/
  !mv *.tar.gz data/ && cd data && echo "Extracting dataset" && ls *.tar.gz | xargs -n1 -I{} bash -c "tar -xvf {} > /dev/null"

def run_pre_training_phase():
  from run import ZEWDPCBaseRun
  run = ZEWDPCBaseRun()
  run.pre_training_phase = pre_training_phase
  run.pre_training_phase(self=run, training_dataset=training_dataset)
  # NOTE:It is critical that the checkpointing works in a self-contained way
  #      As, the evaluators might choose to run the different phases separately.

def run_purchase_phase():
  from run import ZEWDPCBaseRun
  run = ZEWDPCBaseRun()
  run.pre_training_phase = pre_training_phase
  run.purchase_phase = purchase_phase
  # Hacky way to make it work in notebook
  unlabelled_dataset.purchases = set()
  run.purchase_phase(self=run, unlabelled_dataset=unlabelled_dataset, training_dataset=training_dataset, budget=3000)
  del run

def run_prediction_phase():
  from run import ZEWDPCBaseRun
  run = ZEWDPCBaseRun()
  run.pre_training_phase = pre_training_phase
  run.purchase_phase = purchase_phase
  run.prediction_phase = prediction_phase
  run.prediction_phase(self=run, test_dataset=val_dataset)
  del run
fatal: destination path 'data-purchasing-challenge-2022-starter-kit' already exists and is not an empty directory.
                 Datasets for challenge #1024                                   
│ #  Title              Description                  Size │                  
│ 0 │ validation.tar.gz │ Validation dataset       │ 182 MiB │                  
│ 1 │ unlabelled.tar.gz │ Unlabelled image dataset │ 609 MiB │                  
│ 2 │ training.tar.gz   │ Training data            │ 304 MiB │                  
│ 3 │ debug.tar.gz      │ Debug dataset            │ 6.1 MiB │                  
validation.tar.gz: 100% 191M/191M [00:09<00:00, 19.6MB/s]
unlabelled.tar.gz: 100% 638M/638M [00:33<00:00, 19.3MB/s]
training.tar.gz: 100% 319M/319M [00:06<00:00, 51.0MB/s]
debug.tar.gz: 100% 6.43M/6.43M [00:00<00:00, 16.2MB/s]
Extracting dataset

3) Load datasets

In [6]:
from evaluator.dataset import ZEWDPCBaseDataset, ZEWDPCProtectedDataset

# Instantiate Training Dataset
training_dataset = ZEWDPCBaseDataset(
# Instantiate Unlabelled Dataset
unlabelled_dataset = ZEWDPCProtectedDataset(
    budget=3000,  # Configurable Parameter
# Instantiate Validation Dataset
val_dataset = ZEWDPCBaseDataset(
In [9]:
import numpy as np
import matplotlib.pyplot as plt

def show_dataset_summary(dataset):
    images = []
    for sample in dataset:

    images = np.array(images)
    r_dim = images[:,:,:,0].ravel()
    g_dim = images[:,:,:,1].ravel()
    b_dim = images[:,:,:,2].ravel()

    print("Dataset shape: ", images.shape)
    fig, ax = plt.subplots(nrows=1, ncols=3, figsize=(15, 5))
    ax[0].hist(r_dim, bins=50, density=True)
    ax[0].set_title("Red pixels distribution. \n Min={}, Max={}, Mean={:.2f}, Std={:.2f}".format(r_dim.min(), r_dim.max(), r_dim.mean(), r_dim.std()))
    ax[1].hist(g_dim, bins=50, density=True)
    ax[1].set_title("Green pixels distribution. \n Min={}, Max={}, Mean={:.2f}, Std={:.2f}".format(g_dim.min(), g_dim.max(), g_dim.mean(), g_dim.std()))
    ax[2].hist(b_dim, bins=50, density=True)
    ax[2].set_title("Blue pixels distribution. \n Min={}, Max={}, Mean={:.2f}, Std={:.2f}".format(b_dim.min(), b_dim.max(), b_dim.mean(), b_dim.std()))
In [10]:
print("Summary for Training dataset")

print("Summary for Unlabelled dataset")

print("Summary for Validation dataset")
Summary for Training dataset
Dataset shape:  (5000, 256, 256, 3)
Summary for Unlabelled dataset
Dataset shape:  (10000, 256, 256, 3)
Summary for Validation dataset
Dataset shape:  (3000, 256, 256, 3)

Distribution of channel pixels among the datasets are nearly same. That is good thing

Action points

Based on the above analysis, you can normalize your images for better performance with the below code snipped. For example the below normalization is given for training set.
NOTE: Don't get confused between the numbers shown in mean, std in graphs above and the numbers below. ToTensor normalize the image pixels by dividing each pixel with 255 and therefore you need to put the normalized means and standard deviations in torchvision.transforms.Normalize

transform = torchvision.transforms.Compose([
        mean=[0.566, 0.546, 0.490],
        std=[0.204, 0.193, 0.219],


12 months ago

How much did this help? I also wonder about other preprocessing steps you could do, like for example edge detection. Isn’t the model supposed to figure this out by itself during training?

You must login before you can post a comment.