def test_prediction_csv(dir_images_name='database/test/images',
csv_name='test_prediction.csv'):
dir_images = np.sort(os.listdir(dir_images_name))
mask_otsu_all = []
mask_aux_all = []
score = []
with open(csv_name, 'w', newline='') as file:
writer = csv.writer(file)
writer.writerow(["ImageId", "EncodedPixels"])
for i in np.arange(np.size(dir_images)):
# - - - Llamada a la función 'skin_lesion_segmentation'
# Implementa el método propuesto para la segmentación de la lesión
# y proporciona a su salida la máscara predicha.
predicted_mask = skin_lesion_segmentation(dir_images_name + '/' +
dir_images[i])
# - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
# - - - Codificación RLE y escritura en fichero .csv
encoded_pixels = rle_encode(predicted_mask)
writer.writerow([dir_images[i][:-4], encoded_pixels])
print('Máscara ' + str(i) + ' codificada.')
def create_submission(model,
datagen,
orig_imgsize=(1918, 1280),
resize=False,
sort=True):
# Create the rle
timestr = time.strftime("%Y%m%d-%H%M%S")
submission_name = os.path.join(
SUBMISSION_PATH,
'carvana-sub-%s-%s.csv' % (model_func.__name__, timestr))
print("Saving Submission file to: %s" % submission_name)
# Open and memory efficiently write to the file
with open(submission_name, 'w') as csvfile:
# Initialize the header and writer
fieldnames = ['img', 'rle_mask']
writer = csv.DictWriter(csvfile, fieldnames=fieldnames)
writer.writeheader()
# Loop through all the test samples and encode them
for step in tqdm(range(datagen._generator.steps_per_epoch)):
batch_img, batch_long_ids = next(datagen)
batch_preds = model.predict_on_batch(batch_img)
if resize:
# Resize predictions
zoom_factor = (1., orig_imgsize[1] / batch_preds.shape[1],
orig_imgsize[0] / batch_preds.shape[2])
resized_batch_preds = ndi.interpolation.zoom(batch_preds,
zoom_factor,
order=0,
mode='nearest')
else:
# We need to remove the padding
padding = infer_padding(batch_preds.shape[1:3], orig_imgsize)
if padding != (0, 0, 0, 0):
left, right, top, bottom = padding
resized_batch_preds = batch_preds[:, top:batch_preds.
shape[2] - bottom,
left:batch_preds.
shape[2] - right]
else:
resized_batch_preds = batch_preds
# Threshold the image
resized_batch_preds = np.rint(resized_batch_preds)
assert resized_batch_preds.shape[1:] == (
orig_imgsize[1], orig_imgsize[0]
), "Resized images have wrong size of %s" % resized_batch_preds.shape
for i in range(resized_batch_preds.shape[0]):
sample_ind = step * BATCH_SIZE + i
if DEBUG:
print("Index: ", sample_ind)
print("Expected Long ID: ",
datagen._generator.dataset.long_ids[sample_ind])
print("Actual Long ID: ", batch_long_ids[i])
print("Resized: ", resize)
examine_prediction(
datagen._generator.dataset.long_ids[sample_ind],
batch_img[i],
batch_preds[i],
resized_batch_preds[i],
orig_imgsize=orig_imgsize)
assert datagen._generator.dataset.long_ids[
sample_ind] == batch_long_ids[
i], "Long IDs don't match! %s and %s" % (
datagen._generator.dataset.long_ids[sample_ind],
batch_long_ids[i])
# Write the row into csvfile
writer.writerow({
'img':
datagen._generator.dataset.long_ids[sample_ind] + '.jpg',
'rle_mask':
rle.rle_to_string(rle.rle_encode(resized_batch_preds[i]))
})
if sort:
print("Sorting CSVfile")
df = pd.read_csv(submission_name)
df = df.sort_values(by=['img'])
df.to_csv(submission_name, index=False)
print("Submission created at: %s" % submission_name)
total = len(img_list)
out_pred_rows = []
predictions = pd.read_csv('csvs/768_resnet_noship_prob.csv')
THRESHOLD = 0.5
if not INVERT_PRED:
has_ship_set = set(predictions.loc[predictions['Probability'] > THRESHOLD]
['ImageId'].tolist())
else:
has_ship_set = set(predictions.loc[predictions['Probability'] <= THRESHOLD]
['ImageId'].tolist())
full_submission = rle_encode(np.ones((768, 768, 1)).T)
for i in tqdm(range(total)):
if img_list[i] in has_ship_set:
out_pred_rows += [{
'ImageId': img_list[i],
'EncodedPixels': full_submission
}]
else:
out_pred_rows += [{'ImageId': img_list[i], 'EncodedPixels': None}]
submission_df = pd.DataFrame(out_pred_rows)[['ImageId', 'EncodedPixels']]
submission_df.to_csv('csvs/05_thresh_768_resnet_noship.csv', index=False)
print(submission_df.head())
# In[ ]:
aligned_segs = np.stack(tta_batch_reverse(first_segs), 3)
avg_seg = np.sum(aligned_segs, axis=3, keepdims=True)
num_augs = aligned_segs.shape[3]
comparison = THRESHOLD * TTA_INTERSECT_RATIO * num_augs
test_pred = np.greater(avg_seg, comparison).astype('float32')[0:,...]
for pred, c_img_name in zip(test_pred[0:,:,:,:], test_names):
# transpose rows/cols due to airbus data being col major order
ships_pred = instance_segment_mask(pred, smallest_size=SMALL_OBJ_SIZE)
if len(ships_pred) > 0:
for ship in ships_pred:
c_rle = rle_encode(np.transpose(ship, axes=[1,0,2]))
out_pred_rows += [{'ImageId': c_img_name, 'EncodedPixels': c_rle}]
else:
out_pred_rows += [{'ImageId': c_img_name, 'EncodedPixels': None}]
if i % 100 == 0:
gc.collect()
else:
out_pred_rows = [{'ImageId': c_img_name, 'EncodedPixels': None} for c_img_name in img_list]
submission_df = pd.DataFrame(out_pred_rows)[['ImageId', 'EncodedPixels']]
submission_df.to_csv('resnet34_768_ohem_best_loss_th_05_so_60_submission.csv', index=False)
print(submission_df.head())
# In[ ]:
print("--------Making Predictions-------\n")
predictions = predModel.predict(input_fn=pred_fn)
p = list(predictions)
p = np.reshape(p,(18000,101,101),'F')
p = np.round(p)
p_rle = []
"""
testP = p[:20]
for i in range(len(testP)):
implt=plt.imshow(testP[i],cmap='gray')
plt.show()
print(data['ids'][i])
"""
for i in range(len(p)):
p_rle.append(rle.rle_encode(p[i]))
if i%1000 == 0:
print("Converting RLEs:",i,"of 18000\n")
preds = {
'id': data['ids'],
'rle_mask' : p_rle
}
'''
df = pd.DataFrame.from_dict(preds)
print(df.head())
'''
#preds = [x for x in predictions]
#print(preds[:10])