deivadharshini Goto Github PK

mathjs

An extensive math library for JavaScript and Node.js

open-solution-salt-detection

Open solution to the TGS Salt Identification Challenge https://www.kaggle.com/c/tgs-salt-identification-challenge

salt

salt-

salt-identification

salt-identification-challenge

tgs-salt-identification

tgs-salt-identification-challenge

In [ ]: %matplotlib inline import sys sys.path.append('../') import numpy as np import glob from PIL import Image import matplotlib.pyplot as plt import seaborn as sns from sklearn.externals import joblib from src.utils import plot_list, read_images from src.metrics import compute_ious, compute_eval_metric VALIDATION_RESULTS_FILEPATH = 'YOUR/validation_results.pkl' In [ ]: meta_valid, y_true, y_pred = joblib.load(VALIDATION_RESULTS_FILEPATH) In [ ]: meta_valid.head() In [ ]: ious = [compute_ious(gt, pred)[0][0] for gt, pred in zip(y_true, y_pred)] iouts = [compute_eval_metric(gt, pred) for gt, pred in zip(y_true, y_pred)] raw_imgs = read_images(meta_valid['file_path_image']) depths = meta_valid['z'].tolist() results = list(zip(ious, iouts, depths, raw_imgs, y_true, y_pred)) Score distributions In [ ]: sns.distplot(ious) In [ ]: sns.distplot(iouts) It seems that the model is either really good or really bad. For example: In [ ]: print(np.mean([score for score in iouts if score>0.1])) score by depth In [ ]: sns.regplot(depths, iouts, fit_reg=False) It seems that the model is better for lower depths. Not sure yet what to do with it. Predicted mask exploration In [ ]: def filter_results(results, iout_range): iout_min, iout_max = iout_range results_filtered = [] for tup in results: if iout_min<=tup[1]<=iout_max: results_filtered.append(tup) return results_filtered results_filtered = filter_results(results, iout_range=(0.0,0.5)) len(results_filtered) In [ ]: np.mean(list(zip(*results))[2]), np.mean(list(zip(*results_filtered))[2]) In [ ]: IMG_NR = 77 for iou, iout, z, img, gt, pred in results_filtered[:IMG_NR]: print('IOU {}, IOUT {}, depth {}'.format(iou, iout, z)) plot_list(images=[img],labels=[gt, pred])

youtube-like-predictor

YouTube Like Count Predictions using Machine Learning