testing = pd.read_csv("test.csv",

header=None,

names=['name'],

dtype={'name': object})

training = pd.read_csv("training.csv",

header=None,

names=['name', 'fga'],

dtype={'name': object, 'fga': float})

patch_size=32,

rotations=4,

limit=None):

training = pd.read_csv("training.csv",

header=None,

names=['name', 'fga'],

dtype={'name': object, 'fga': float})

if limit:

training = training.head(limit)

from skimage.io import imsave

for i, row in training.iterrows():

row_patches = []

row_output = []

name = row['name']

fga = row['fga']

print("Transforming image %s" % name)

import sys

sys.stdout.flush()

for kind in ["DX", "TS"]:

img = imread("images/%s/%s-%s.png" % (kind, name, kind))

img_patches = transform_img(img, name, patch_size=patch_size)

for img_patch in img_patches:

for rot_i in range(rotations):

row_patches.append(np.rot90(img_patch, k=rot_i))

row_output.append(fga)

X_row = np.asarray(row_patches)

y_row = np.asarray(row_output)

with open(path.join(output_dir, "X_file_%s" % name), 'w') as f:

np.save(f, X_row)

with open(path.join(output_dir, "y_file_%s" % name), 'w') as f:

from skimage.transform import rotate

from skimage.exposure import is_low_contrast

rows, cols, _ = img.shape

max_dim = max(rows, cols)

# If more rows than cols, vertical

vertical = True if rows == max_dim else False

if rows < patch_size or cols < patch_size:

print("Image too small: %s-%s.png" % (name))

return []

# Make all images vertical for now (more rows than cols)

if not vertical:

img = img.transpose((1, 0, 2))

rows, cols, _ = img.shape

row_blocks = rows / patch_size

col_blocks = cols / patch_size

patches = []

for i in range(row_blocks):

for j in range(col_blocks):

patch = img[patch_size * i:patch_size * (i + 1),

patch_size * j:patch_size * (j + 1),

:]

if not is_low_contrast(patch):

patches.append(patch)