def submit(self, preds):
for i in range(len(self.x)):
file = os.path.basename(self.x[i]['image_file'])
for j, c in enumerate(preds[i]):
encoded_pixels = mask2rle(c)
im_cl = f'{file}_{j + 1}'
self.submit_res.append({
'ImageId_ClassId': im_cl,
'EncodedPixels': encoded_pixels
})
preds = preds.detach().cpu().numpy()
# Batch post processing
for p, file in zip(preds, image_file):
file = os.path.basename(file)
# Image postprocessing
for i in range(4):
p_channel = p[i]
imageid_classid = file + '_' + str(i + 1)
p_channel = (p_channel > thresholds[i]).astype(np.uint8)
if p_channel.sum() < min_area[i]:
p_channel = np.zeros(p_channel.shape, dtype=p_channel.dtype)
res.append({
'ImageId_ClassId': imageid_classid,
'EncodedPixels': mask2rle(p_channel)
})
df = pd.DataFrame(res)
df.to_csv('submission.csv', index=False)
df = pd.DataFrame(res)
df = df.fillna('')
df.to_csv('submission.csv', index=False)
df['Image'] = df['ImageId_ClassId'].map(lambda x: x.split('_')[0])
df['Class'] = df['ImageId_ClassId'].map(lambda x: x.split('_')[1])
df['empty'] = df['EncodedPixels'].map(lambda x: not x)
df[df['empty'] == False]['Class'].value_counts()
### matplotlib inline
df = pd.read_csv('submission.csv')[:40]
df['Image'] = df['ImageId_ClassId'].map(lambda x: x.split('_')[0])
cls_pred=torch.stack(cls_pred)
cls_pred = torch.mean(cls_pred, dim=0)
cls_pred = (np.copy(cls_pred) > threshold)
# Batch post processing
for p_cls,p, file in zip(cls_pred,preds, image_file):
file = os.path.basename(file)
# Image postprocessing
for i in range(4):
img_rle = ""
p_channel = p[i]
imageid_classid = file + '_' + str(i + 1)
if p_cls[i]:
p_channel = (p_channel > thresholds[i]).astype(np.uint8)
if p_channel.sum() >= min_area[i]:
img_rle = mask2rle(p_channel)
count[i]+=1
# p_channel = np.zeros(p_channel.shape, dtype=p_channel.dtype)
res[idx_res] = {'ImageId_ClassId': imageid_classid, 'EncodedPixels': img_rle}
idx_res = idx_res + 1
#else:
# p_channel = np.zeros(p_channel.shape, dtype=np.uint8)
# res.append({
# 'ImageId_ClassId': imageid_classid,
# 'EncodedPixels': img_rle
# })
df = pd.DataFrame(res)
columns = ['ImageId_ClassId','EncodedPixels']
df.to_csv('E:/pycharm_project/steel/code/submit/twotta.csv', index=False, columns=columns)
print(count[0],count[1],count[2],count[3])
preds = preds.detach().cpu().numpy()
# Batch post processing
for p, file in zip(preds, image_file):
file = os.path.basename(file)
# Image postprocessing
for i in range(4):
p_channel = p[i]
imageid_classid = file + "_" + str(i + 1)
p_channel = (p_channel > thresholds[i]).astype(np.uint8)
if p_channel.sum() < min_area[i]:
p_channel = np.zeros(p_channel.shape, dtype=p_channel.dtype)
res.append({
"ImageId_ClassId": imageid_classid,
"EncodedPixels": mask2rle(p_channel),
})
df = pd.DataFrame(res)
df.to_csv("submission.csv", index=False)
df = pd.DataFrame(res)
df = df.fillna("")
df.to_csv("submission.csv", index=False)
# In[22]:
import pdb
import os
import cv2
import torch