From 1696e306c11c364446bbc80ba515c58c38df3dae Mon Sep 17 00:00:00 2001 From: Anjana Tiha Date: Fri, 29 Mar 2019 18:00:53 -0400 Subject: [PATCH] Update README.md --- README.md | 9 +++------ 1 file changed, 3 insertions(+), 6 deletions(-) diff --git a/README.md b/README.md index a4bda33..de528de 100644 --- a/README.md +++ b/README.md @@ -7,13 +7,10 @@ Application : Image Recognition, Image Classification, Medical Imaging, B ### Description -1. Detected Malaria from segmented cells from the thin blood smear slide images from Malaria screening research activity by National Institutes of Health (NIH) with Deep Learning (Convolutional Neural Network). +1. Detected Malaria from segmented cells from the thin blood smear slide images collected from Malaria screening research activity by National Institutes of Health (NIH) with Deep Learning (Convolutional Neural Network) specifically by training pretrained model Nashnet completely from scratch. 2. Before feeding data into model, preprocessed and augmented image dataset containing 27,558 images (337MB) by adding random flips, rotations and shears. -3. For training, used pretrained model Nashnet and trained completely from scratch. -4. After loading pretrainied model NasNetMobile, added global max pooling, global average pooling, flattened layer to output of trained model and concatenated them. -5. Added dropout and batch normalization layers for regularization. -6. Added final output layer with - a dense layer with softmax activation and compiled with optimizer Adam with learning rate 0.001, metric- accuracy and loss-categorical crossentropy. -7. Trained for 10 iterations and attained training accuracy 96.47% and loss(categorical crossentrpy) 0.1026 and validation accuracy of 95.46% and loss 0.1385. +3. After loading pretrainied model NasNetMobile, added global max pooling, global average pooling, flattened layer to output of trained model and concatenated them and also added dropout and batch normalization layers for regularization before adding final output layer - a dense layer with softmax activation and compiling with optimizer-Adam with learning rate-0.0001, metric-accuracy and loss-categorical crossentropy. +4. Trained for 10 iterations and attained training accuracy 96.47% and loss(categorical crossentrpy) 0.1026 and validation accuracy of 95.46% and loss 0.1385. #### Code