Plant Pathology competition is a annual competition conducted by FGVC8 to detect foilar diseases in Apple Plants.

Check out the competition here

• Host : Kaggle
• Dataset : 23,000 high-quality RGB images of apple foliar diseases
• Evaluation Metric : Mean F1-score
• Framework used : TensorFlow
• Accelerators : GPU, TPU

This is a multi-label task i.e., an image can have more than one label.

• Number of classes : 6 (scab, healthy, frog eye leaf spot, rust, complex, powdery mildew)
• Number of unique labels : 12 (above 6 + scab, frog eye leaf spot, scab frog eye leaf spot complex, frog eye leaf spot, complex, rust frog eye leaf spot, rust complex, powdery mildew, complex)

I approached the task in 3 ways

• Considering number of unique labels as 12 (assuming no other combination of diseases) and proceded as Multi Class classification
• Considering number of unique labels as 6 (Multi label classification).
• Considering number of unique labels as 5 (if no label in 5 gave good probability then it is healthy).
• Last approach gave best resluts.

• Dataset Preparation

  • TPU as Accelerator
    • As the dataset is present in GCP, we can't access the images directly, We first need to decode the image using image = tf.image.decode_jpeg(tf.io.read_file(filepath), channels=3)
    • After loading the images, random augmentations are applied using tf.image.random. class.
    • Finally to prepare the datset for training, loaded images are converted to Tensorflow tf.data.Dataset format and adding cache, prefetch for faster loading of data.
    • More details are in TPU notebook.
  • GPU as Accelerator
    • ImageDataGenerator is used to apply several augmentations to images.
    • Images are loaded as datset using flow_from_dataframe function.
    • More details in GPU notebook.

• Training

  • In both cases (TPU,GPU) pretrained models present in tf.keras.applications are used.
  • Loss : Binary Cross Entropy.
  • Metrics tracked throughput training are F1-score , Accuracy.
  • Callbacks used
    • ModelCheckpoint - to store the best model.
    • ReduceRONPlateau - to change learning rate if there is no improvement observed.
    • EarlyStopping - to stop training i case of no improvement.
  • TPU as Accelerator
    • As we can't use any tensorflow dataset loaders, kfold split of sklearn is used to split the data into training and validation.
  • GPU as Accelerator
    • using split option in ImageDataGenerator datset is splitted into training and validation.

• Submission

  • Finally the best models are loaded and used to predict new labels.
  • More details in Inference Notebook.
  • Sample submission csv file can be found here.