Create README.md

mrcnn/README.md

@@ -0,0 +1,95 @@
+# Model Training
+
+## Table of Contents
+
+1. [Introduction](#introduction)  
+1. [Training](#training)  
+
+
+## Introduction
+
+
+## Setup
+
+Run data synthesizer module to generate the training  datasets in the following folder structure:
+
+~~~~~~~
+        project
+          |-- mrcnn
+               |-- scipts
+                    |-- config.py
+                    |-- model.py
+                    |-- train.py
+                    |-- pre_process.py
+                    |-- requirements.txt
+                    |-- utils.py
+                    |-- visualize.py
+                    |-- Inference_notebook.ipynb
+                    |-- utils.py
+
+               |--- mask_data
+                    |-- id_map.json
+                    |-- logs/
+                    |-- mask_rcnn_coco.h5
+                    |-- test_image
+                    |-- train_image
+~~~~~~~
+
+## Training
+
+
+### Pre-processing
+Data pre-processing using pre_process.py to generate .h5 file for masks.
+
+
+### Model and Training
+
+ * Modified Matterport's implementation of Mask-RCNN deep neural network for object instance segmentation.
+    Model - Added two new methods to train just specific masks.
+            mrcnn_mask: Just mask layers
+            mask_heads: Mask layers or rpn/fpn
+    Multiclass- Pre_processed images and mask files accordingly to prepare for multi classification.
+ * Increased maximum number of predicted objects since an image can contain 200 or more bottles/bags/boxes.
+ * Increased POST_NMS_ROIS_TRAINING to get more region proposals during training.
+ * Resized images and masks to 512x512.
+ * Used Default anchor size as we do not expect small objects: RPN_ANCHOR_SCALES = (16, 32, 64, 128, 256)
+ * Relied heavily on deep image augmentation due to small training set:
+      * Random horizontal or vertical flips
+      * Random 90 or -90 degrees rotation
+      * Random rotations in the range of (-20, 20) degrees
+      * Random scaling of image and mask scaling in the range (0.5, 2.0)
+ * Used Resnet101 architecture as a backbone encoder.
+ * Trained the model with Adam optimizer for 65 epochs:
+ * 5 epochs of heads with learning rate 1e-4 (To speed up the training process)
+ * 30 epochs with learning rate 1e-5
+ * 30 epochs with learning rate 1e-6
+ * changed mAP computation to be (0.5 - 0.8) 
+ * weighted mAP 
+ * weighted loss 
+ * LOSS_WEIGHTS = {
+        "rpn_class_loss": 20.,
+        "rpn_bbox_loss": 1.,
+        "mrcnn_class_loss": 10.,
+        "mrcnn_bbox_loss": 1.,
+        "mrcnn_mask_loss": 10.
+    }
+
+
+### Model Execution and Run-Time
+
+Run `python pre_process.py` to pre-process data 
+
+Run `python train.py` to train the model. Model weights are saved at ../data/logs/kaggle_bowl/mask_rcnn.h5.
+
+Run python inference_notebook.ipynb.py to evaluate model performance on test set 
+
+
+The following execution times are measured on Nvidia P100 GPUs provided by AWS Deep learning AMI
+
+```
+Each training epoch takes about 25 minutes.
+It takes about 18 hours to train the model from scratch.
+
+## Example model predictions
+
+[put graphs from notebook]