MosquitoAlert Challenge 2023


This notebook is for improving classification of mosquitoes detection of starter notebook.


*Note: Open this in colab as rendering in aicrowd is not appropriate for this.
As a dataset, this notebook uses cropped images of mosquito detection. The base model is derived from Keras-CV-attention since it contains numerous pretrained models that I used, for example, MaxViT, ConvNextV2, and EfficientNetV2. I only used convolutional models for submission because transformer models are overfitting for this dataset and have a long inference time on the CPU.
I used OpenVino to fit two seconds of inference time for larger models. Finally, I ran Model Soup on the best folds but saw no improvement in LB. Certain codes, such as CBAM and FocalLoss with labelsmoothing, have been purposefully omitted because neither LB nor CV provide a performance advantage. I did not use any external datasets because I misunderstood the rules. The CV split is 95%–5% of the data. Use the custom Lr function callback instead of WarmUpCosineLr for faster training.
 Colab Link

🦟 Mosquito Classification using Tensorflow Keras's Neural Networks and Transfer Learning

Thanks for visiting my notebook

This is forked from @Marionette 👺's kernel and I have adjusted it for TPU and added other augmentation methods, such as cut&mix. The base model is derived from keras-cv-attention since it contains numerous pretrained models that i used for example MaxViT, ConvNext, EfficientNetV2. For, Submission I only used Convolution models as transformer models are overfitting for this dataset.
In order to fit two seconds of inference time for larger models, I used OpenVino. Lastly, I ran Model Soup on several folds. Certain codes, such as CBAM and FocalLoss with labelsmoothing, have been purposefully omitted because there is no performance advantage in either LB or CV. I haven't used any external dataset for this. The CV split is 5% of train data. Use custom Lr function callback instead of WarmUpCosineLr for faster training.

Special thanks to Marionette 👺 for creating this amazing notebook and Keras CV Attention for the collection of various pretrained models.

📌 Feel free to fork or edit the notebook for your own convenience.

👋Thanks for Visting my Notebook

📌 Feel free to fork or edit the notebook for your own convenience.


Mosquitoes, small yet perilous insects, are responsible for transmitting diseases that pose a serious threat to humans and the environment. With over 3600 known species, a few of them have the ability to transmit various pathogens, leading to widespread illnesses such as Zika, Dengue, and Chikungunya. Controlling mosquito populations is vital to prevent disease outbreaks and protect communities worldwide.

In collaboration with the Mosquito Alert citizen science project, the Mosquito Identification Challenge—an opportunity to impact public health initiatives directly. Traditional mosquito surveillance methods are expensive and time-consuming, but community-based approaches empower citizens to report and collect mosquito specimens. By leveraging machine learning and deep learning techniques, we aim to automate the labour-intensive image validation process, making mosquito identification more efficient and accurate.

The dataset consists of six distinct classes, including two species and three genus classes, as well as a class for a species complex. Here is a summary of the classes and their descriptions: 1. Aedes aegypti - Species 2. Aedes albopictus - Species 3. Anopheles - Genus 4. Culex - Genus (Species classification is challenging, so it is given at the genus level) 5. Culiseta - Genus 6. Aedes japonicus/Aedes koreicus - Species complex (Difficult to differentiate between the two species)

❗ Note:

Make sure to run the cells from top to bottom with a TPU accelerator for training. There are some linux commands present in some cells so this is important to take into account. You can change the baseline to any model in Keras-CV-Attention pretrained models. Also, any suggestions, comments and recommendations to improve the notebook will be highly appreciated. Cheers!

In [ ]:
from google.colab import drive
Mounted at /content/drive

🏗️Import Necessary Libraries

In [ ]:
!pip install seaborn openvino-dev
!pip install tensorflow_addons wandb git+https://github.com/leondgarse/keras_cv_attention_models.git
!wandb login
In [ ]:
# Import Data Science Libraries
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split
import itertools
import random
import math
# Import visualization libraries
import matplotlib.pyplot as plt
import matplotlib.cm as cm
import cv2
import seaborn as sns

# Tensorflow Libraries
from tensorflow import keras
from tensorflow.keras import layers,models
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.layers import Dense, Dropout
from tensorflow.keras.callbacks import Callback, EarlyStopping,ModelCheckpoint
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras import Model
from tensorflow.keras.layers.experimental import preprocessing
from tensorflow.data.experimental import AUTOTUNE
from tensorflow.keras.callbacks import ReduceLROnPlateau
import tensorflow_addons as tfa
import tensorflow.keras.backend as K

# System libraries
from pathlib import Path
import os.path

# Metrics
from sklearn.metrics import classification_report, confusion_matrix
from wandb.keras import WandbCallback,WandbMetricsLogger

import wandb

SEED= 42
model_name = 'efficeintnetv2L-480' # @param {type:"string"}
USER_NAME = 'saidinesh' # @param {type:"string"}
wandb.init(entity=USER_NAME, project='mosquito-classification', job_type='train', name=model_name , sync_tensorboard=True)


TPU strategy

In [ ]:
# Detect TPU, return appropriate distribution strategy

    tpu = tf.distribute.cluster_resolver.TPUClusterResolver()
    print('Running on TPU ', tpu.master())
except ValueError:
    tpu = None

if tpu:
    strategy = tf.distribute.experimental.TPUStrategy(tpu)
    strategy = tf.distribute.get_strategy()

print("REPLICAS: ", strategy.num_replicas_in_sync)
Running on TPU  grpc://
WARNING:absl:`tf.distribute.experimental.TPUStrategy` is deprecated, please use  the non experimental symbol `tf.distribute.TPUStrategy` instead.
In [ ]:
# Seed Everything to reproduce results for future use cases
def seed_everything(seed=42):
    # Seed value for TensorFlow

    # Seed value for NumPy

    # Seed value for Python's random library

    # Force TensorFlow to use single thread
    # Multiple threads are a potential source of non-reproducible results.
    session_conf = tf.compat.v1.ConfigProto(

    # Make sure that TensorFlow uses a deterministic operation wherever possible

    sess = tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(), config=session_conf)


🤙Create helper functions

In [ ]:
!wget https://raw.githubusercontent.com/mrdbourke/tensorflow-deep-learning/main/extras/helper_functions.py

# Import series of helper functions for our notebook
from helper_functions import create_tensorboard_callback, plot_loss_curves, unzip_data, compare_historys, walk_through_dir, pred_and_plot
--2023-10-21 08:53:47--  https://raw.githubusercontent.com/mrdbourke/tensorflow-deep-learning/main/extras/helper_functions.py
Resolving raw.githubusercontent.com (raw.githubusercontent.com)...,,, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)||:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 10246 (10K) [text/plain]
Saving to: ‘helper_functions.py’

helper_functions.py 100%[===================>]  10.01K  --.-KB/s    in 0s      

2023-10-21 08:53:47 (25.5 MB/s) - ‘helper_functions.py’ saved [10246/10246]

📥Load and Transform Data

In [ ]:
BATCH_SIZE = 8 # @param {type:"number"}
INPUT_SIZE = "480" #@param ["384","480","512"]
TARGET_SIZE = (int(INPUT_SIZE),int(INPUT_SIZE)) #(224, 224)
BATCH_SIZE = 8 * strategy.num_replicas_in_sync #64
INIT_LR = 0.00001 * (BATCH_SIZE/64)
DISABLE_CUTMIX = False # @param {type:"boolean"}
# ImageNet mean and std
MEAN = [0.485, 0.456, 0.406]
STD = [0.229, 0.224, 0.225]
In [ ]:
gcs_path = 'gs://kds-9a7c3559b6f8b4ef686fe4a6c2224bed3900561891c6ec018ef0fcc2'
# walk_through_dir(gcs_path)
training_filepath = tf.io.gfile.glob(gcs_path + '/train_crop.tfrecords')
validation_filepath = tf.io.gfile.glob(gcs_path + '/val_crop.tfrecords')
test_filepath = validation_filepath

📅Load Tfrecord as Bytes

In [ ]:
# Load TFRecord file from the folder as bytes
raw_training_dataset = tf.data.TFRecordDataset(training_filepath)
raw_validation_dataset = tf.data.TFRecordDataset(validation_filepath)
raw_test_dataset = tf.data.TFRecordDataset(test_filepath)
In [ ]:
# Create a dictionary describing the features
labeled_feature_description = {
    'label': tf.io.FixedLenFeature([], tf.int64),
    'image': tf.io.FixedLenFeature([], tf.string)
#Class names of Mosquitoes
CLASSES = ['aegypti', 'albopictus', 'anopheles', 'culex', 'culiseta', 'japonicus-koreicus']
In [ ]:
# Create a function to read and extract images from dataset
def _parse_labeled_image_function(example_proto):
    example = tf.io.parse_single_example(example_proto, labeled_feature_description)
    image = tf.io.decode_jpeg(example['image'])
    image = tf.cast(image, tf.float32) / 255.
    image = tf.image.resize(image, TARGET_SIZE)
    #label = tf.cast(example['label'], tf.int32)
    label = tf.one_hot(example['label'], depth=6)
    return image, label

# Parse labeled images, shuffle and batch
training_dataset = (

# Parse unlabeled images and batch
validation_dataset = (

# Parse unlabeled images and batch
test_dataset = (

🔭Visualizing images from the dataset

In [ ]:
# Display images in a 5x5 grid

image_batch, label_batch = next(iter(training_dataset))

def display_images(image_batch, label_batch):
    plt.figure(figsize = [20,12])
    for i in range(25):

display_images(image_batch, label_batch)