def __getitem__(self, i):
if self.mode is 'train' or self.mode is 'val':
image_id, mask, border = self.data.iloc[i]
# im_path = os.path.join(TEST_PATH, image_id)
# image = cv2.imread(im_path)
image = self.images[image_id]
mask = rle_decode(mask)
image, mask = self.apply_aug(image, mask)
image = torch.from_numpy(image).float().permute([2, 0, 1])
mask = torch.from_numpy(mask).float().unsqueeze(2).permute([2, 0, 1])
has_ship = torch.sum(mask) > 0
return image, mask, has_ship
elif self.mode is 'test':
image_id = self.images[i]
im_path = os.path.join(TEST_PATH, image_id)
image = cv2.imread(im_path)
image, _ = self.apply_aug(image)
image = torch.from_numpy(image).float().permute([2, 0, 1])
return (image, image_id)
else:
raise RuntimeError()
erosion = cv2.erode(mask, kernel, iterations=1)
return mask - erosion
print("Started preprocessing")
df = pd.read_csv('data/prediction.csv')
df['isempty'] = df.apply(lambda x: 1 if x.EncodedPixels is np.nan else 0,
axis=1)
ship_counts = df.groupby('ImageId').count()[['EncodedPixels']].rename(
{'EncodedPixels': 'ships'}, axis=1)
df = df.merge(ship_counts, right_index=True, left_on='ImageId')
df['msize'] = df[df.isempty == 0].apply(
lambda x: np.sum(rle_decode(x.EncodedPixels)), axis=1)
groupz = df.groupby('ImageId').groups
k = [(group, df.iloc[groupz[group]].EncodedPixels.values)
for group in tqdm.tqdm(groupz)]
kk = [(x[0], combine_masks(x[1])) for x in tqdm.tqdm(k)]
final = pd.DataFrame(kk)
border_mask = final.iloc[:, 1].apply(
lambda x: rle_encode(create_border_mask(rle_decode(x))))
final['border'] = border_mask
final.columns = ['ImageId', 'mask_nb', 'border']
final = df[['ImageId', 'ships']].merge(final,
left_on='ImageId',
right_on='ImageId')
final.to_csv("data/prediction_p.csv", index=False)
filenames = os.listdir(args.path)
dataframes = [pd.read_csv(os.path.join(args.path, name)) for name in filenames]
if threshold is None:
threshold = len(dataframes) // 2 + 1
test_samples = len(dataframes[0])
print(f"voting threshold = {threshold}")
pred_distr = {-1: 0, 0: 0, 1: 0, 2: 0, 3: 0}
encoded_pixels = []
encoded_pixels_ch = []
print("processing submissions")
for i in tqdm(range(test_samples)):
mask = np.zeros((350, 525))
for j in range(len(dataframes)):
if dataframes[j].iloc[i]["EncodedPixels"] is not np.nan:
mask += rle_decode(dataframes[j].iloc[i]["EncodedPixels"])
mask_final = np.where(mask >= threshold, 1, 0)
if mask_final.sum() == 0:
pred_distr[-1] += 1
encoded_pixels.append("")
else:
pred_distr[i % 4] += 1
r = mask2rle(mask_final)
encoded_pixels.append(r)
print(
f"empty={pred_distr[-1]} fish={pred_distr[0]} flower={pred_distr[1]} gravel={pred_distr[2]} sugar={pred_distr[3]}"
)
non_empty = pred_distr[0] + pred_distr[1] + pred_distr[2] + pred_distr[3]
total = non_empty + pred_distr[-1]
print(f"% of non-empty masks {round(non_empty / total, 4)}")
# Initialize mode and load trained weights
ckpt_path = "../output/weights/model_96_resnet18f2_0.pth"
device = torch.device("cuda")
model = smp.Unet(args.model,
classes=4,
encoder_weights='imagenet',
activation=None)
salt = model.to(device)
model.eval()
state = torch.load(ckpt_path, map_location=lambda storage, loc: storage)
model.load_state_dict(state["state_dict"])
# start prediction
predictions = []
for i, batch in enumerate(tqdm(test_loader)):
fnames, images = batch
batch_preds = torch.sigmoid(model(images.to(device)))
batch_preds = batch_preds.detach().cpu().numpy()
for fname, preds in zip(fnames, batch_preds):
for cls, pred in enumerate(preds):
pred, num = post_process(pred, best_threshold, min_size)
rle = rle_decode(pred)
name = fname + f"_{cls+1}"
predictions.append([name, rle])
# save predictions to submission.csv
df = pd.DataFrame(predictions,
columns=['ImageId_ClassId', 'EncodedPixels'])
df.to_csv("../output/submissions/submission.csv", index=False)
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from skimage.measure import label
from PIL import Image
import cv2
import torch
from torch.utils import data
from src.utils import rle_decode
print("Started preprocessing")
df = pd.read_csv('data/train_ship_segmentations_v2.csv')
df['isempty'] = df.apply(lambda x: 1 if x.EncodedPixels is np.nan else 0,
axis=1)
ship_counts = df.groupby('ImageId').count()[['EncodedPixels']].rename(
{'EncodedPixels': 'ships'}, axis=1)
df = df.merge(ship_counts, right_index=True, left_on='ImageId')
df['msize'] = df[df.isempty == 0].apply(
lambda x: np.sum(rle_decode(x.EncodedPixels)), axis=1)
df.to_csv('data/processed.csv', index=False)
print("Finished. Save to data/processed.csv")