def view_mistakes():
foldername = 'myval'
id_to_pred = get_decisions(
foldername,
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binary13b')
fileids = sb.clean_glob(
glob.glob(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/%s/images/*.png'
% foldername))
# imgs = sd.get_salt_labels(folder='myval')
# imgs = np.reshape(imgs, (imgs.shape[0], 101,101))
imgs = sd.get_salt_images(folder='myval')
val_pix_num = sd.salt_pixel_num(folder='myval')
valY = np.array(val_pix_num > 0).astype(int)
mistake_i = 0
for file_id_i, file_id in enumerate(fileids):
fileid, file_extension = os.path.splitext(file_id)
if valY[file_id_i] != id_to_pred[fileid]:
imgs[mistake_i, :, :] = imgs[file_id_i, :, :]
mistake_i += 1
print('made %i mistakes' % mistake_i)
X = np.swapaxes(np.swapaxes(imgs[:mistake_i, :, :], 0, 2), 0,
1).astype(float)
fig, ax = plt.subplots(1, 1)
tracker = ScrollThruPlot(ax, X, fig)
fig.canvas.mpl_connect('scroll_event', tracker.onscroll)
plt.show()
def eval_masks(
outpath='/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/predictions/myval/'
):
valX = get_salt_images(folder='myval')
fileids = sb.clean_glob(
glob.glob(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/myval/images/*.png'
))
model_dir = '/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binarypix1'
model = tf.estimator.Estimator(model_fn, model_dir=model_dir)
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[:384, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
for file_i, prediction in enumerate(tqdm(gen, total=384)):
p_label = prediction['mask']
pred = np.array(p_label).reshape((101, 101))
plt.imsave(outpath + fileids[file_i], pred, cmap=cm.gray)
# now get the tail
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[-32:, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
for file_i, prediction in enumerate(gen):
fileid = fileids[404 - 32 + file_i]
p_label = prediction['mask']
pred = np.array(p_label).reshape((101, 101))
plt.imsave(outpath + fileid, pred, cmap=cm.gray)
def kaggle_test(
outpath='/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/predictions/'
):
testX = get_salt_images(folder='test')
fileids = sb.clean_glob(
glob.glob(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/test/images/*.png'
))
model_dir = '/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binarypix1'
model = tf.estimator.Estimator(model_fn, model_dir=model_dir)
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': testX[:17984, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
id_to_pred = np.load(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binary1/test_bin_pred.npy'
).tolist()
with open(outpath + 'binarypix2.csv', 'a') as fd:
fd.write('id,rle_mask\n')
for file_i, prediction in enumerate(tqdm(gen, total=17984)):
fileid, file_extension = os.path.splitext(fileids[file_i])
p_label = prediction['mask']
pred = np.array(p_label).reshape(
(101, 101)).transpose().reshape(101**2) * id_to_pred[fileid]
fd.write('%s,%s\n' % (fileid, myrlestring(pred)))
# now get the tail
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': testX[-32:, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
for file_i, prediction in enumerate(gen):
if file_i >= 16:
fileid, file_extension = os.path.splitext(fileids[18000 - 32 +
file_i])
p_label = prediction['mask']
pred = np.array(p_label).reshape(
(101, 101)).transpose().reshape(101**
2) * id_to_pred[fileid]
fd.write('%s,%s\n' % (fileid, myrlestring(pred)))
def get_decisions(foldername='test', model_dir=None, outfile=None):
"""
change the first two vars to run on different sets
"""
valX = get_salt_images(folder=foldername)
fileids = sb.clean_glob(
glob.glob(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/%s/images/*.png'
% foldername))
setsize = len(fileids)
headsz = int(setsize / float(batch_size)) * batch_size
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[:headsz, :, :, :]},
batch_size=batch_size,
shuffle=False)
# '/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binary1'
bin_model = tf.estimator.Estimator(model_fn, model_dir=model_dir)
gen = bin_model.predict(input_fn)
id_to_pred = {}
for file_i, prediction in enumerate(tqdm(gen, total=headsz)):
fileid, file_extension = os.path.splitext(fileids[file_i])
id_to_pred[fileid] = prediction
# now get the tail
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[-batch_size:, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = bin_model.predict(input_fn)
for file_i, prediction in enumerate(gen):
idx = setsize - batch_size + file_i
fileid, file_extension = os.path.splitext(fileids[idx])
id_to_pred[fileid] = prediction
# '/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binary1/test_bin_pred'
if outfile:
np.save(outfile, id_to_pred)
return id_to_pred
def kaggle_summary(
outpath='/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/predictions/myval/'
):
valX = get_salt_images(folder='myval')
valY = get_salt_labels(folder='myval')
fileids = sb.clean_glob(
glob.glob(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/myval/images/*.png'
))
model_dir = '/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binarypix1'
model = tf.estimator.Estimator(model_fn, model_dir=model_dir)
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[:384, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
id_to_pred = np.load(
'/scratch0/ilya/locDoc/data/kaggle-seismic-dataset/models/binary13b/val_bin_pred.npy'
).tolist()
threshes = np.array(
[0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95])
tps = np.zeros(threshes.shape[0])
fps = np.zeros(threshes.shape[0])
fns = np.zeros(threshes.shape[0])
def tp_fp_fn_calc(gt, predicted, tps, fps, fns):
component1 = np.array(predicted).astype(bool)
component2 = gt.astype(bool)
overlap = component1 * component2 # Logical AND
union = component1 + component2 # Logical OR
iou = overlap.sum() / float(union.sum() + 1e-5)
mask_present_gt = np.any(gt > 0)
# these metrics methods return a tuple container
if mask_present_gt:
if np.all(predicted < 1):
fns += np.ones(threshes.shape[0])
else:
tps += (threshes < iou).astype(int)
else:
if np.any(predicted > 0):
fps += np.ones(threshes.shape[0])
for file_i, prediction in enumerate(tqdm(gen, total=384)):
fileid, file_extension = os.path.splitext(fileids[file_i])
p_label = prediction['mask']
p_label = np.array(p_label) * id_to_pred[fileid]
tp_fp_fn_calc(valY[file_i, :], p_label, tps, fps, fns)
# now get the tail
input_fn = tf.estimator.inputs.numpy_input_fn(
x={'images': valX[-32:, :, :, :]},
batch_size=batch_size,
shuffle=False)
gen = model.predict(input_fn)
for file_i, prediction in enumerate(gen):
idx = 404 - 32 + file_i
fileid, file_extension = os.path.splitext(fileids[idx])
p_label = prediction['mask']
p_label = np.array(p_label) * id_to_pred[fileid]
tp_fp_fn_calc(valY[idx, :], p_label, tps, fps, fns)
precisions = tps / (tps + fps + fns)
avg_precision = precisions.sum() / threshes.shape[0]
print('%d masks in dataset' % np.sum(valY.sum(axis=1) > 0))
for idx, thresh in enumerate(threshes):
print('tp at %f: %f' % (thresh, tps[idx]))
print('fp at %f: %f' % (thresh, fps[idx]))
print('fn at %f: %f' % (thresh, fns[idx]))
print('precisions at %f: %f' % (thresh, precisions[idx]))
print('avg precision: %f' % (avg_precision))