def plot_class_prediction(subm_name, image_id, c):
subm = pd.read_csv('subm/%s.csv.gz' % subm_name)
image = load_image(image_id)
xmax = grid_sizes.loc[image_id, 'xmax']
ymin = grid_sizes.loc[image_id, 'ymin']
multi_poly = shapely.wkt.loads(
subm.loc[(subm['ClassType'] == c + 1) & (subm['ImageId'] == image_id),
'MultipolygonWKT'].values[0])
plt.title("%s - class %d" % (subm_name, c))
plt.imshow(poly_to_mask(multi_poly, image.shape[1:], [xmax, ymin]),
cmap='hot')
plt.show()
def plot_predictions(subm_name, image_id):
subm = pd.read_csv('subm/%s.csv.gz' % subm_name)
image = load_image(image_id)
xmax = grid_sizes.loc[image_id, 'xmax']
ymin = grid_sizes.loc[image_id, 'ymin']
f, ax = plt.subplots(2, 5, sharex='col', sharey='row')
ax[0][0].imshow(np.rollaxis(image, 0, 3))
for c in xrange(9):
multi_poly = shapely.wkt.loads(subm.loc[(subm['ClassType'] == c + 1) &
(subm['ImageId'] == image_id),
'MultipolygonWKT'].values[0])
ax[(c + 1) // 5][(c + 1) % 5].imshow(poly_to_mask(
multi_poly, image.shape[1:], [xmax, ymin]),
cmap='hot')
plt.tight_layout()
plt.show()
def plot_prediction(image_id, pred_id, cls=0):
image = np.load('cache/images/%s.npy' % image_id)
pred = np.load('cache/preds/%s.npy' % image_id)
xymax = (grid_sizes.loc[image_id, 'xmax'], grid_sizes.loc[image_id,
'ymin'])
plt.figure()
ax1 = plt.subplot(131)
ax1.set_title('image_id:%s' % image_id)
ax1.imshow(image[1, :, :], cmap=plt.get_cmap('gray'))
ax2 = plt.subplot(132)
ax2.set_title('predict bldg pixels')
ax2.imshow(pred[cls], cmap=plt.get_cmap('hot'))
ax3 = plt.subplot(133)
ax3.set_title('predict bldg polygones')
ax3.imshow(poly_to_mask(mask_to_poly(pred[cls], xymax), image.shape[1:],
xymax),
cmap=plt.get_cmap('hot'))
plt.title("%s - class %d" % (pred_id, cls))
plt.show()
import shapely.wkt as wkt
import cv2
print "Preparing train image masks..."
# Prepare location
for image_id, image_cls_wkt in train_wkt.groupby('image_id'):
print " Processing %s..." % image_id
xmax = grid_sizes.loc[image_id, 'xmax']
ymin = grid_sizes.loc[image_id, 'ymin']
meta = load_pickle('cache/meta/%s.pickle' % image_id)
mask = np.zeros((n_classes, meta['shape'][1], meta['shape'][2]),
dtype=np.float32)
for tp in image_cls_wkt.itertuples():
poly = wkt.loads(tp.multi_poly_wkt)
h, w = meta['shape'][1:]
mask[tp.cls - 1] = np.clip(
cv2.resize(poly_to_mask(poly, (w * 4, h * 4),
[xmax, ymin]).astype(np.float32), (w, h),
interpolation=cv2.INTER_AREA), 0, 1)
np.save('cache/masks/%s.npy' % image_id, mask)
print "Done."
for image_id in sorted(subm['ImageId'].unique()):
print "%s..." % image_id
xymax = (grid_sizes.loc[image_id, 'xmax'], grid_sizes.loc[image_id, 'ymin'])
subm.loc[subm['ImageId'] == image_id, 'MultipolygonWKT'] = 'MULTIPOLYGON EMPTY'
for cls in classes:
polys = [shapely.wkt.loads(s.loc[(s['ImageId'] == image_id) & (s['ClassType'] == cls), 'MultipolygonWKT'].iloc[0]) for s in subms]
if pre_buffer_size is not None:
polys = [p.buffer(pre_buffer_size) for p in polys]
if pixelize:
mask = sum(poly_to_mask(p, (18000, 18000), xymax) for p in polys) > len(polys) * 0.5
if 'mask_postprocess' in preset:
mask = preset['mask_postprocess'](mask)
res = mask_to_poly(mask, xymax, min_area=1.0, threshold=1.0)
else:
try:
poly_parts = []
for i in xrange(len(polys)):
for j in xrange(i+1, len(polys)):
poly_parts.append(polys[i].intersection(polys[j]))
res = unary_union(poly_parts)
except:
xymax = (grid_sizes.loc[image_id, 'xmax'], grid_sizes.loc[image_id,
'ymin'])
subm.loc[subm['ImageId'] == image_id,
'MultipolygonWKT'] = 'MULTIPOLYGON EMPTY'
for cls in classes:
polys = [
shapely.wkt.loads(
s.loc[(s['ImageId'] == image_id) & (s['ClassType'] == cls),
'MultipolygonWKT'].iloc[0]) for s in subms
]
if preset.get('pixelize', False):
mask = sum(poly_to_mask(p, (9000, 9000), xymax)
for p in polys) > 0.5
if 'mask_postprocess' in preset:
mask = preset['mask_postprocess'](mask)
res = mask_to_poly(mask, xymax, min_area=1.0, threshold=1.0)
else:
try:
res = unary_union(polys)
except:
print "Error, using first poly"
res = polys[0]
res = res.buffer(0)
for image_id in full_train_image_ids:
start_time = time.time()
sys.stdout.write(" Processing %s... " % image_id)
sys.stdout.flush()
pred = predict_mask(image_id)
xymax = (grid_sizes.loc[image_id, 'xmax'], grid_sizes.loc[image_id,
'ymin'])
for ci, cls in enumerate(classes):
true_poly = shapely.wkt.loads(
train_wkt.loc[(train_wkt['image_id'] == image_id) &
(train_wkt['cls'] == cls + 1),
'multi_poly_wkt'].iloc[0])
true_mask = poly_to_mask(true_poly, pred.shape[1:], xymax)
pred_mask = pred[ci]
cls_pixel_inter = (pred_mask & true_mask).sum()
cls_pixel_union = (pred_mask | true_mask).sum()
pixel_intersections[cls] += cls_pixel_inter
pixel_unions[cls] += cls_pixel_union
pred_poly = mask_to_poly(pred_mask, xymax, **cls_opts.get(cls, {}))
poly_intersections[cls] += pred_poly.intersection(true_poly).area
poly_unions[cls] += pred_poly.union(true_poly).area
print "Done in %d seconds" % (time.time() - start_time)