def fuse(self, result_dir):
train_json = self.load_json_file('./data/train_fusesize.json')
train_df = pd.DataFrame({'label_id': [x['class'] for x in train_json]})
results_file = glob(os.path.join(result_dir, '*.val.*.pkl'))
results_file.sort()
print('Total {} ensemble files.'.format(len(results_file)))
new_dict = {}
weight = [0.111111, 0.333333, 0.333333, 0.222222]
weight = [0.2] * 5
for i, res in tqdm(enumerate(results_file)):
res_dict = load_dict(res)
for k,v in res_dict.items():
v = v['logit']
if not k in new_dict.keys():
new_dict[k] = v / v.sum() * weight[i]
else:
new_dict[k] += v / v.sum() * weight[i]
# tmp = np.sqrt(new_dict[k] * (v / v.sum()))
# new_dict[k] = tmp / tmp.sum()
all_id_list = []
all_predicted = []
correct_count = 0
prob_list = []
label_list = []
for k,v in tqdm(new_dict.items(), desc='Fusing result'):
v /= v.sum()
prob_list.append(v)
top3_result = np.argsort(-v, axis=-1)[:3]
all_predicted.append('{:d} {:d} {:d}'.format(top3_result[0], top3_result[1], top3_result[2]))
all_id_list.append(k)
label = res_dict[k]['label']
label_list.append([label])
a = -np.sort(-v)[:10] / 6
correct_count += np.sum(top3_result == label)
datafame = pd.DataFrame({'id': all_id_list, 'predicted': all_predicted})
predict_csv = os.path.join(result_dir, 'ensemble.csv')
# datafame.to_csv(predict_csv, index=False, sep=',')
print('Top3 error rate: {:.6f}'.format(1 - correct_count / float(len(all_id_list))))
probs = np.stack(prob_list)
label = np.array(label_list)
# probability calibration
# calibrated_prob = self.calibrate_all(probs, train_df)
calibrated_prob = self.calibrate_probs(probs, train_df)
topk_predict_calib = np.argsort(-calibrated_prob, axis=-1)[:, 0:3]
err_calib = 1. - np.sum((topk_predict_calib == label), axis=-1).mean()
print('Top3 error rate (calibration): {:.6f}'.format(err_calib))
if False:
calibrated_prob_sp = self.calibrate_superlabel(train_json, calibrated_prob)
topk_predict_calib_sp = np.argsort(-calibrated_prob_sp, axis=-1)[:, 0:3]
err_calib_sp = 1. - np.sum((topk_predict_calib_sp == label), axis=-1).mean()
print('Top3 error rate (calibration hirechical label): {:.6f}'.format(err_calib_sp))
all_predicted_calib = []
for item in topk_predict_calib:
all_predicted_calib.append('{:d} {:d} {:d}'.format(item[0], item[1], item[2]))
datafame_calib = pd.DataFrame({'id': all_id_list, 'predicted': all_predicted_calib})
predict_csv_calib = os.path.join(result_dir, 'calib_ensemble_4.csv')
def cover_dict():
src = './log/ensemble/ep11.logit.test.pkl'
dst = './log/ensemble/ep11.test.ens100.logit.pkl'
cur_dict= load_dict(src)
new_dict = {}
for k,v in tqdm(cur_dict.items()):
new_dict[k] = {'logit': v, 'label': -1}
save_dict(new_dict, dst)
def load_feat_single(self, filename):
dict_feat_list = []
filename_list = glob(filename)
filename_list.sort()
for fn in filename_list:
print('Loading feature from {:s}'.format(fn))
dict_feat = load_dict(fn)
dict_feat_list.append(dict_feat)
return dict_feat_list
def knn(self, result_dir):
train_json = self.load_json_file('./data/train_fusesize.json')
train_df = pd.DataFrame({'label_id': [x['class'] for x in train_json]})
results_file_train = os.path.join(result_dir, 'train.logit.pkl')
print('Loading training file...')
dict_train = load_dict(results_file_train)
train_data = []
train_label = []
for k, v in dict_train.items():
train_data.append(v['logit'] / v['logit'].sum())
train_label.append(v['label'])
train_data = np.stack(train_data, axis=0)
train_label = np.array(train_label)
results_file_val = os.path.join(result_dir, 'val.logit.pkl')
print('Loading validation file...')
dict_val = load_dict(results_file_val)
val_data = []
val_label = []
for k, v in dict_val.items():
val_data.append(v['logit'] / v['logit'].sum())
val_label.append(v['label'])
val_data = np.stack(val_data, axis=0)
val_label = np.array(val_label)
print('Building KD tree...')
kdt = KDTree(train_data, leaf_size=30, metric='euclidean')
print('Predicting...')
idx = kdt.query(val_data, k=500, return_distance=False)
pred_tmp = train_label[idx]
count = np.stack([np.bincount(x, minlength=2019) for x in pred_tmp],
axis=0)
pred = np.argsort(-count, axis=-1)[:, 0:3]
err = 1. - np.sum(val_label[:, np.newaxis] == pred, axis=-1).mean()
print('Error: {:f}'.format(err))
pass
def calibrate_superlabel(self, train_json, probs):
splabel = load_dict('./data/fglabel2splabel.pkl')
sp_lab2idx_dict = {}
dict_fg2sp = {}
count = 0
for k, v in splabel.items():
tmp_str = ''.join([str(x) for x in v])
splabel[k] = tmp_str
if tmp_str not in sp_lab2idx_dict.keys():
sp_lab2idx_dict[tmp_str] = count
count += 1
for k, v in splabel.items():
tmp_str = ''.join([str(x) for x in v])
dict_fg2sp[k] = sp_lab2idx_dict[tmp_str]
topk_predict = np.argsort(-probs, axis=-1)[:, 0:3]
fgidx = np.array([dict_fg2sp[x] for x in range(self.num_class)])
new_probs = np.zeros_like(probs)
for i in tqdm(range(fgidx[-1])):
mask = fgidx == i
mask_float = mask / np.sum(mask)
mask_float = mask_float[np.newaxis, :]
tmpp = np.matmul(probs, mask_float.transpose())
new_probs += (probs + tmpp) * mask
return new_probs
def __init__(self, phase, is_transform=True, DEBUG=False):
self.img_size = 'fusesize'
self.img_root = '/data1/zhaoj/data/challenge/iMaterialist/image_' + self.img_size
self.img_list_root = './data'
self.phase = phase
self.fglabel2splabel_dict = load_dict('./data/fglabel2splabel.pkl')
reimgsize = 256
self.transform_dict = {
'resnet': {
'train':
transforms.Compose([
transforms.Resize(reimgsize),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
'test':
transforms.Compose([
transforms.Resize(reimgsize),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
},
# 'test': transforms.Compose([transforms.Resize(reimgsize), transforms.CenterCrop(224), transforms.ToTensor(),
# transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])},
'senet': {
'train':
transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
]),
'test':
transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
},
'inceptionresnet': {
'train':
transforms.Compose([
transforms.Resize(342),
transforms.RandomResizedCrop(299),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
]),
'test':
transforms.Compose([
transforms.Resize(342),
transforms.RandomResizedCrop(299),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
])
},
'nasnet': {
'train':
transforms.Compose([
transforms.Resize(378),
transforms.RandomResizedCrop(331),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
]),
'test':
transforms.Compose([
transforms.Resize(378),
transforms.RandomResizedCrop(331),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
])
}
}
if phase == 'train' and is_transform:
self.transform = self.transform_dict['inceptionresnet']['train']
# self.transform = transforms.Compose([transforms.Resize(512), transforms.RandomResizedCrop(448), transforms.RandomHorizontalFlip(), transforms.ToTensor(),
# transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
else:
self.transform = self.transform_dict['inceptionresnet']['test']
# self.transform = transforms.Compose([transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(),
# transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
# self.transform = transforms.Compose([transforms.Resize(512), transforms.RandomCrop(448), transforms.ToTensor(),
# transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
self.img_list = self.load_json_file()
if DEBUG:
self.img_list = self.img_list[:301]