Techkriti is an annual techfest conducted by IIT Kanpur. As a part of the fest, competitions are conducted is various themes like Enterprenurial , Technology , Miscellaneous. This year (March 14th, 2021) ML Hackathon is conducted as a part of Technology division.

I got first position (Team Name : Aine) in ML Hackathon and this repo gives details of my approach to problem statement.

Check results here

Visit official website

Classify the images into following 6 categories

• buildings
• forest
• glacier
• mountain
• sea
• street

More details here

Dataset is uploaded in Kaggle

• Total images in training data : 14034
• Total Images in Validation data : 3000
• Total images in test data : 7301

• Height, Width = 150, 150
• Augmentations are added to images using tf.keras.preprocessing.image.ImageDataGenerator
  • random rotation
  • horizontal flip
  • width shift range
  • shear range
  • zoom range

• Metric : Accuracy
• GPU used : NVIDIA TESLA P100
• Framework : TensorFlow
• Pretrained models present in tf.keras.applications are used.
• Metric : Categorical Accuracy
• Loss : Categorical CrossEntropy
• Optimizer : Adam
• Callbacks : Model checkpoint , Reduce RON Plateau , Early Stopping

• All the experiments done are present in experiments folder
• Summary can be found here in csv file.
• Best Model (Experiment-9 , Accuracy : 94.6)

    pretrained = tf.keras.applications.DenseNet201(include_top=False,
                                                      weights='imagenet',
                                                      pooling="avg",
                                                      input_shape=[HEIGHT,WIDTH, 3])
    
    x = pretrained.output
    x = tf.keras.layers.Dropout(0.3) (x)
    x = tf.keras.layers.Dense(128) (x)
    x = tf.keras.layers.LeakyReLU(alpha=0.2) (x)
    x = tf.keras.layers.GaussianDropout(0.4) (x)
    outputs = tf.keras.layers.Dense(NUM_CLASSES,activation="softmax", dtype='float32')(x)
        
    model = tf.keras.Model(pretrained.input, outputs)

• Finally all the experiments and details are compiled into documentation.pdf
• prediction.csv contains final predictions on test data.