import sys import os import cv2 import numpy as np import pandas as pd from timeit import default_timer as timer import psutil import platform import torch from torch.utils.data.dataset import Dataset from torch.utils.data import DataLoader from torch.utils.data.sampler import * import torch.nn.functional as F def time_to_str(t, mode='min'): if mode == 'min': t = int(t) / 60 hr = t // 60 min = t % 60 return '%2d hr %02d min' % (hr, min) elif mode == 'sec': t = int(t) min = t // 60 sec = t % 60 return '%2d min %02d sec' % (min, sec) else: raise NotImplementedError print('ram:', 'gb=%d'%int(psutil.virtual_memory()[0] / 1024 / 1024 / 1024)) print('cpu:', psutil.cpu_count()) print('gpu:', str(torch.cuda.get_device_properties(0))[22:-1]) print('') # setting ####################################################### if 0: # local image_dir = '/root/share1/kaggle/2020/cassava/data/train_images' df = pd.read_csv('/root/share1/kaggle/2020/cassava/data/train.csv') image_id = df.loc[:299, 'image_id'].values.tolist() label = df.loc[:299, 'label'].values checkpoint0 = [ '/root/share1/kaggle/2020/cassava/result/effb4/xx4-512/fold0/checkpoint/00011000_model.pth', '/root/share1/kaggle/2020/cassava/result/effb4/xx4-512/fold1/checkpoint/00011000_model.pth', '/root/share1/kaggle/2020/cassava/result/effb4/xx4-512/fold2/checkpoint/00010000_model.pth', ] checkpoint1 = [ '/root/share1/kaggle/2020/cassava/result/se-res50/xx2-512/fold0/checkpoint/00045000_model.pth', '/root/share1/kaggle/2020/cassava/result/se-res50/xx2-512/fold2/checkpoint/00012000_model.pth', ] if 1: # kaggle if 0: #(debug) image_dir = '../input/cassava-leaf-disease-classification/train_images' df = pd.read_csv('../input/cassava-leaf-disease-classification/train.csv') image_id = df.loc[:299, 'image_id'].values.tolist() label = df.loc[:299, 'label'].values else: image_dir = '../input/cassava-leaf-disease-classification/test_images' image_id = list(os.listdir(image_dir)) label = None sys.path.insert(0, '../input/cassava-dummy') checkpoint0 = [ # '../input/cassava-dummy/efb4_fold0_00011000_model.pth', # '../input/cassava-dummy/efb4_fold1_00011000_model.pth', # '../input/cassava-dummy/efb4_fold2_00010000_model.pth', # '../input/cassava-dummy/efb4_fold3_00009000_model.pth', # '../input/cassava-dummy/efb4_fold4_00009000_model.pth', '../input/cassava-dummy/efb4_ext_fold0_00050500_model.pth', '../input/cassava-dummy/efb4_ext_fold1_00015000_model.pth', '../input/cassava-dummy/efb4_ext_fold2_00014000_model.pth', '../input/cassava-dummy/efb4_ext_fold3_00014000_model.pth', '../input/cassava-dummy/efb4_ext_fold4_00013500_model.pth', # '../input/cassava-dummy/effb4_512_new_fold1_00008500_model.pth', # '../input/cassava-dummy/effb4_512_new_fold2_00009500_model.pth', # '../input/cassava-dummy/effb4_512_new_fold3_00009500_model.pth', # '../input/cassava-dummy/effb4_512_new_fold4_00009000_model.pth', ] checkpoint1 = [ # '../input/cassava-dummy/se50_fold0_00045000_model.pth', # '../input/cassava-dummy/se50_fold1_00012000_model.pth', # '../input/cassava-dummy/se50_fold2_00012000_model.pth', # '../input/cassava-dummy/se50_fold3_00011000_model.pth', # '../input/cassava-dummy/se50_fold4_00011000_model.pth', # '../input/cassava-dummy/se50_ext_fold0_00050500_model.pth', # '../input/cassava-dummy/se50_ext_fold1_00015000_model.pth', # '../input/cassava-dummy/se50_ext_fold2_00016000_model.pth', # '../input/cassava-dummy/se50_ext_fold3_00017500_model.pth', # '../input/cassava-dummy/se50_ext_fold4_00015500_model.pth', # '../input/cassava-dummy/se50_448_fold0_00014000_model.pth', # '../input/cassava-dummy/se50_448_fold1_00015000_model.pth', # '../input/cassava-dummy/se50_448_fold2_00015000_model.pth', # '../input/cassava-dummy/se50_448_fold3_00014000_model.pth', # '../input/cassava-dummy/se50_448_fold4_00014000_model.pth', ] ################################################################# from model0 import Net as Net0 #efficient-b4 from model1 import Net as Net1 #se-reesnext-50 # model --- net0 = [] for f in checkpoint0: n = Net0().cuda() n.load_state_dict(torch.load(f, map_location=lambda storage, loc: storage)['state_dict'] , strict=True) net0.append(n) net1 = [] for f in checkpoint1: n = Net1().cuda() n.load_state_dict(torch.load(f, map_location=lambda storage, loc: storage)['state_dict'], strict=True) net1.append(n) print('load checkppoint ok! net0', len(net0)) print('load checkppoint ok! net1', len(net1)) # dataset --- image_size = 512 class CassavaDataset(Dataset): def __init__(self, ): pass def __len__(self): return len(image_id) def __getitem__(self, index): image = cv2.imread(image_dir + '/%s' % image_id[index]) image = cv2.resize(image, dsize=(image_size, image_size), interpolation=cv2.INTER_LINEAR) image = image[...,::-1] image = image.astype(np.float32) / 255 image = np.ascontiguousarray(image.transpose(2, 1, 0)) return image dataset = CassavaDataset() loader = DataLoader( dataset, sampler=SequentialSampler(dataset), batch_size=16, drop_last=False, num_workers=0, pin_memory=True, ) # start here! ------------------ probability = [] start_timer = timer() with torch.no_grad(): for t, batch in enumerate(loader): image = batch.cuda() p = [] for net in net0 + net1: # net.eval() logit = net(image) p.append(F.softmax(logit, -1)) # tta ---- if 0: logit = net(torch.flip(image, dims=(2,)).contiguous()) p.append(F.softmax(logit, -1)) logit = net(torch.flip(image, dims=(3,)).contiguous()) p.append(F.softmax(logit, -1)) logit = net(torch.flip(image, dims=(2,3)).contiguous()) p.append(F.softmax(logit, -1)) logit = net(image.permute(0,1,3,2).contiguous()) p.append(F.softmax(logit, -1)) # --------- p = torch.stack(p).mean(0) probability.append(p.data.cpu().numpy()) if t % 3 == 0: print('\r %8d / %d %s' % (t, len(loader), time_to_str(timer() - start_timer, 'sec')), end='', flush=True) print('') # ---------------------- probability = np.concatenate(probability) predict = probability.argmax(-1) df_submit = pd.DataFrame({'image_id': image_id, 'label': predict}) print('df_submit', df_submit.shape) print(df_submit) print('estimated time for 15,000 test images = %s'%time_to_str((timer() - start_timer)/len(predict)*15000, 'min')) df_submit.to_csv('submission.csv', index=False) # check if label is not None: correct = (predict == label).mean() print('correct', correct) print('probability\n', probability[:5]) print('predict\n', predict[:10]) ''' load checkppoint ok! 18 / 19 2 min 39 sec df_submit (300, 2) image_id label 0 1000015157.jpg 2 1 1000201771.jpg 3 2 100042118.jpg 4 3 1000723321.jpg 1 4 1000812911.jpg 3 .. ... ... 295 1052095724.jpg 0 296 1052118637.jpg 3 297 1052854295.jpg 0 298 1052881053.jpg 3 299 1052903541.jpg 3 [300 rows x 2 columns] estimated time for 15,000 test images = 2 hr 12 min correct 0.93 probability [[3.09394360e-01 1.13561094e-01 3.63957703e-01 3.47374426e-03 2.09613070e-01] [1.88316408e-05 1.86062753e-04 1.49351486e-03 9.98232782e-01 6.88106447e-05] [1.04819878e-03 2.20503122e-01 6.97037764e-03 1.38261551e-02 7.57652104e-01] [1.67905848e-04 9.92914855e-01 3.28231661e-04 3.20969219e-03 3.37923691e-03] [1.42867793e-05 1.11910940e-05 3.97972763e-04 9.99195218e-01 3.81269114e-04]] predict [2 3 4 1 3 3 2 0 0 3] Process finished with exit code 0 '''