geraldgong/kaggle_shopee

Shopee - Price Match Guarantee

https://www.kaggle.com/c/shopee-product-matching

2021-03-10

• Start with Image Embedding

• Ref:

  https://www.kaggle.com/cdeotte/rapids-cuml-tfidfvectorizer-and-knn/comments

  https://www.kaggle.com/cdeotte/part-2-rapids-tfidfvectorizer-cv-0-700#Compute-Baseline-CV-Score

2021-03-11

• Construct f1-score evaluator
• Try clustering the distance with DBSCAN
• Title embedding (NLP)

2021-03-15

• Opitmized memory usages

2021-03-16

• Try GPU and Implement with RAPIDS
• Tensorflow Hub for text embedding
  • Best score: 0.6986 → LB: 0.675

2021-03-17

• Ensemble text embedding models with different languages (failed)

• Evaluate best threshold for knn results

  • knn_image: 6.8
  • Ken_text: 0.8

• TfidfVectorizer

  • max_feature=10000, best score: 0.725 → LB: 0.678
  • max_feature=15000, best score: 0.735 → LB: 0.690
  • max_feature=20000, best score: 0.732
  • max_feature=25000, best score: 0.728

2021-03-22

• Submit with EfficientNetB0 + KNN (image) and TfidVectorizer + cosine similarity (title)
  • LB: 0.702
• Check the possiblity of augmentation of the training image (left-right flip)

2021-03-24

• DBSCAN on 50 nearest distance of image embeddings
  • LB: 0.720

2021-03-25

• Text preprocessing

• Try Indonesian and English pre-trained model embedding

• LaBSE

• XLM-Roberta

• English --> 18939 samples, Indonesia -->8715 samples, Malay --> 2398 samples, German --> 854 samples

• Polyglot language detection

2021-03-26

• re

• Remove emoji code, contents in (xxx) and [xxx]

• ? merge number with unit to a single character

• ? remove individual number

• ? Keep the dot between

2021-03-30

• Start try with a NLP model with multi-lingual
  • Only keep Noun in title
  • remove all the numbers, units
  • POS polyglot
• Idea: According to similar text to gather corresponding images

text_example = [
 'Vegetarian Kimchi korea sawi 1Kg vegie vegan vege'
 'Vegetarian Kimchi korea sawi 500 gram vegie vegan vege'
 'Fresh Kimchi sawi 500 gram dibuat oleh chef korea asli'
 'Kimchi original sawi 500 gram dibuat oleh chef korea asli enak'
 'kimchi fresh sawi 500 gram korea murah'
]

2021-04-01

• In order to clean the text, we applied stemming using the NLTK Snowball Stemmer, and removed stopwords/punctuation as well as transforming to lowercase.

• Avoid disconnection due to idleness

  Chrome: press F12, then run the following JavaScript snippet in the console

  function KeepClicking(){
  console.log("Clicking");45
  document.querySelector("colab-connect-button").click()
  }
  setInterval(KeepClicking,60000)

2021-04-08

• Finish preprocessing
• Start to evaluate LABSE model try to integrate with ArcFace
• check if the minimum counts of the target, >=2qw3edr2
• check how much of the matches==1 belong to target==2 -> 63%

2021-04-09

• ArcFace train image embedding model (EfficientNet B4)
• Indonesian pretrained model fine-tune with ArcMargin

Install RAPIDS on Colab

!git clone https://github.com/rapidsai/rapidsai-csp-utils.git
!bash rapidsai-csp-utils/colab/rapids-colab.sh stable

import sys, os

dist_package_index = sys.path.index('/usr/local/lib/python3.7/dist-packages')
sys.path = sys.path[:dist_package_index] + ['/usr/local/lib/python3.7/site-packages'] + sys.path[dist_package_index:]
sys.path
exec(open('rapidsai-csp-utils/colab/update_modules.py').read(), globals())

Kaggle Data Download

!pip install -q kaggle
from google.colab import files
files.upload()

!mkdir ~/.kaggle
!cp kaggle.json ~/.kaggle/
!chmod 600 ~/.kaggle/kaggle.json
!kaggle competitions download -c COMPETITION_NAME
!mkdir data
!unzip -q shopee-product-matching.zip -d ./data
!rm shopee-product-matching.zip

Install RAPIDS in conda-env

conda activate conda-env

conda install -c rapidsai-nightly -c nvidia -c numba -c conda-forge cudf python=3.8 cudatoolkit=11.0
conda install -c rapidsai -c nvidia -c conda-forge -c defaults blazingsql=0.18 cuml=0.18 python=3.8 cudatoolkit=11.0

Restrict Tensorflow usage

# RESTRICT TENSORFLOW TO 2GB OF GPU RAM
# SO THAT WE HAVE 14GB RAM FOR RAPIDS
LIMIT = 2.0
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
    try:
        tf.config.experimental.set_virtual_device_configuration(
            gpus[0],
            [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024*LIMIT)])
        logical_gpus = tf.config.experimental.list_logical_devices('GPU')
        #print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
    except RuntimeError as e:
        print(e)
print('We will restrict TensorFlow to max %iGB GPU RAM'%LIMIT)
print('then RAPIDS can use %iGB GPU RAM'%(16-LIMIT))

Evaluation

Image Embedding	Text Embedding	CV score	LB score
EfficientNetB0, NearestNeighbor: 0.65	Tifidf: 25000, Cosine Similarity: 0.7	0.7273	0.702
EfficientNetB0, NearestNeighbor: 0.65	Tifidf: 15000, Cosine Similarity: 0.7	0.7342	0.694
EfficientNetB0, NearestNeighbor: 0.65	Tifidf: 25000, Cosine Similarity: 0.6	0.7377	0.688
EfficientNetB0, DBSCAN: eps= 1	Tifidf: 25000, Cosine Similarity: 0.7	0.7488	0.718
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine Similarity: 0.72	0.7464	0.720
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine + DBSCAN: eps=0.1	0.7449	0.709
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine + DBSCAN: eps=0.12	0.7471	0.709
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 20000, Cosine Similarity: 0.72, preprocess	0.7496	0.719
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine Similarity: 0.6, preprocess	0.7519	0.696
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine Similarity: 0.72, preprocess, >=2 text matches if both image and text have only 1 match	0.7641	0.727
EfficientNetB0, DBSCAN: eps= 1.2	Tifidf: 25000, Cosine Similarity: 0.72, >=2 text matches if both image and text have only 1 match	0.7658	0.730