Update README.md

Author: n0obcoder

README.md

@@ -1,2 +1,99 @@
-# NIH-Chest-X-Rays-Multi-Label-Image-Classification-In-Pytorch
+# NIH-Chest-X-rays-Multi-Label-Image-Classification-In-Pytorch
 Multi-Label Image Classification of the Chest X-Rays In Pytorch
+
+# Dataset
+[NIH Chest X-ray Dataset](https://www.kaggle.com/nih-chest-xrays/data#Data_Entry_2017.csv) is used for  Multi-Label Disease Classification of of the Chest X-Rays.
+There are a total of 15 classes (14 diseases, and one for 'No findings')
+Images can be classified as "No findings" or one or more disease classes:
+* Atelectasis
+* Consolidation
+* Infiltration
+* Pneumothorax
+* Edema
+* Emphysema
+* Fibrosis
+* Effusion
+* Pneumonia
+* Pleural_thickening
+* Cardiomegaly
+* Nodule Mass
+* Hernia
+
+There are 112,120 X-ray images of size 1024x1024 pixels, in which 86,524 images are for training and 25,596 are for testing.
+
+# Model 
+Pretrained Resnet50 model is used for Transfer Learning on this new image dataset.
+
+# Loss Function
+There is a choice of loss function
+* Focal Loss (default)
+* Binary Cross Entropy Loss or BCE Loss
+
+# Training
+* ### From Scratch 
+    Following are the layers which are set to trainable-
+    * layer2
+    * layer3
+    * layer4
+    * fc
+
+    Terminal Code: 
+    ```
+    python main.py
+    ```
+
+* ### Resume From a Saved Checkpoint
+    A Saved Checkpoint needs to be loaded which is nothing but a dictionary containing the 
+    * epochs (number of epochs the model has been trained till that time)
+    * model (architecture and the learnt weights of the model)
+    * lr_scheduler_state_dict (state_dict of the lr_scheduler)
+    * losses_dict (a dictionary containing the following loses)
+
+      * mean train epoch losses for all the epochs
+      * mean val epoch losses for all the epochs
+      * batch train loss for all the training batches
+      * batch train loss for all the val batches
+ 
+Different layers of the model are freezed/unfreezed in different stages, defined inside the *main.py* to fit the model well on the data. 
+ 
+Terminal Code: 
+```
+python main.py --resume --ckpt checkpoint_file.pth
+```
+
+Training the model will create a **models** directory and will save the checkpoints in there.
+
+# Testing
+A Saved Checkpoint needs to be loaded using the **--ckpt** argument and **--test** argument needs to be passed for activating the Test Mode
+
+Terminal Code: 
+```
+python main.py --test --ckpt checkpoint_file.pth
+```
+
+# Result 
+The model achieved the **ROC AUC Score** of **0.88420** after training in the following stages-
+
+#### STAGE 1 
+* Loss Function: FocalLoss
+* lr: 1e-5
+* Training Layers: layer2, layer3, layer4, fc 
+* Epochs: 2
+
+#### STAGE 2 
+* Loss Function: FocalLoss
+* lr: 3e-4
+* Training Layers: layer3, layer4, fc 
+* Epochs: 1
+
+#### STAGE 3 
+* Loss Function: FocalLoss
+* lr: 1e-3
+* Training Layers: layer4, fc 
+* Epochs: 3
+
+#### STAGE 4 
+* Loss Function: FocalLoss
+* lr: 1e-3
+* Training Layers: fc 
+* Epochs: 2