Export lion patches to training dataset



In [1]:

    
INPUT_DIR='../../input/kaggle-sea-lion/02/'
OUTPUT_DIR='../../output/kaggle-sea-lion/99/'

IMAGE_DIMS_IN = (84,84,3)
IMAGE_DIMS_OUT = (56,56,3)
LABEL_DIMS = (6,)



In [2]:

    
%matplotlib inline
import cv2
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import h5py
import os

import modules.logging
from modules.logging import logger
import modules.lions as lions
from modules.utils import Timer
import modules.utils as utils









    



Using TensorFlow backend.

Prepare output dataset



In [3]:

    
utils.mkdirs(OUTPUT_DIR, recreate=True)
modules.logging.setup_file_logger(OUTPUT_DIR + 'out.log')
logger.info('Dir ' + OUTPUT_DIR + ' created')

x_ds = None
y_ds = None

logger.info('creating dataset')
dataset_path = OUTPUT_DIR + utils.dataset_name('lion-patches-0px', IMAGE_DIMS_OUT)
h5file_out = h5py.File(dataset_path, 'w')
x_ds_out, y_ds_out = utils.create_xy_dataset(h5file_out, IMAGE_DIMS_OUT, LABEL_DIMS, x_dtype='u1')

dataset_path = INPUT_DIR + utils.dataset_name('lion-patches-0px', IMAGE_DIMS_IN)
h5file_in = h5py.File(dataset_path, 'r')
x_ds_in = h5file_in['X']
y_ds_in = h5file_in['Y']

logger.info('done')









    



2017-06-09 00:36:46,657 INFO Dir ../../output/kaggle-sea-lion/99/ created
2017-06-09 00:36:46,660 INFO creating dataset
2017-06-09 00:36:46,677 INFO done

Open lion patches and crop to a new dimension

Process each sample



In [4]:

    
logger.info('Process all lion patches')
qtty = len(x_ds_in)
for i,x in enumerate(x_ds_in):
    y = y_ds_in[i]
    a = int((IMAGE_DIMS_IN[0]-IMAGE_DIMS_OUT[0])/2)
    b = int((IMAGE_DIMS_IN[1]-IMAGE_DIMS_OUT[1])/2)
    aa = a + IMAGE_DIMS_OUT[0]
    bb = b + IMAGE_DIMS_OUT[1]
    x_cropped = x[a:aa,b:bb]
    utils.add_sample_to_dataset(x_ds_out, y_ds_out, x_cropped, y)
    utils.print_progress(status='cropping', current_value=i, target_value=qtty)

h5file_in.close()
h5file_out.close()









    



2017-06-09 00:36:46,688 INFO Process all lion patches






    



2262/2263 [=========================>] 99% cropping

Validate dataset



In [5]:

    
dataset_path = OUTPUT_DIR + utils.dataset_name('lion-patches-0px', IMAGE_DIMS_OUT)
with h5py.File(dataset_path, 'r') as h5file:
    print(len(h5file['X']))
    print(len(h5file['Y']))
    utils.show_images(h5file['X'][0:100].astype('uint8'), image_labels=utils.onehot_to_label(h5file['Y'][0:100]), is_bgr=True, cols=10, size=2)
    logger.info('output dataset length ' + str(len(h5file['Y'][()])))
    logger.info('class distribution')
    logger.info(str(utils.class_distribution(h5file['Y'][()])))









    



2263





    



2017-06-09 00:36:53,355 INFO showing 100 images






    



2263






    



2017-06-09 00:36:53,358 INFO > [started] generating image patches. rows=11; cols=10...
/notebooks/datascience-snippets/kaggle-sea-lion/modules/utils.py:725: FutureWarning: comparison to `None` will result in an elementwise object comparison in the future.
  if(image_labels!=None):






    












    



2017-06-09 00:37:01,658 INFO > [done]    generating image patches. rows=11; cols=10 (8300.230 ms)
2017-06-09 00:37:01,663 INFO output dataset length 2263
2017-06-09 00:37:01,669 INFO class distribution
2017-06-09 00:37:01,673 INFO [  78   45  675  150  281 1034]



In [ ]:



In [ ]:



In [ ]: