def predict_test_majority():
"""
Majority voting method.
"""
labels = np.empty((len(models), 61191, 17))
for m_idx, model in enumerate(models):
name = str(model).split()[1]
print('predicting model {}'.format(name))
net = nn.DataParallel(model().cuda())
net.load_state_dict(
torch.load(
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/model/full_data_{}_10xlr.pth'
.format(name)))
net.eval()
preds = np.zeros((61191, 17))
for t in transforms:
test_dataloader.dataset.images = t(test_dataloader.dataset.images)
print(t, name)
pred = predict(net, dataloader=test_dataloader)
preds = preds + pred
# get predictions for the single model
preds = preds / len(transforms)
np.savetxt(
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/submission_probs/full_data_{}_10xlr_224and256.txt'
.format(name), preds)
# get labels
preds = (preds > thresholds[name]).astype(int)
labels[m_idx] = preds
# majority voting
labels = labels.sum(axis=0)
labels = (labels >= (len(models) // 2)).astype(int)
pred_csv(predictions=labels,
name='majority_voting_ensembles_split_data_10xlr_224and256')
def predict_test_averaging(t):
preds = np.zeros((61191, 17))
# imgs = test_dataloader.dataset.images.copy()
# iterate over models
for index, model in enumerate(models):
name = str(model).split()[1]
net = nn.DataParallel(model().cuda())
net.load_state_dict(
torch.load(
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/model/{}.pth'.
format(name)))
net.eval()
# iterate over transformations
for transformation in transforms:
# imgs = transformation(imgs)
test_dataloader.dataset.images = transformation(
test_dataloader.dataset.images)
pred = predict(dataloader=test_dataloader, net=net)
preds = preds + pred
preds = preds / (len(models) * len(transforms))
# preds = preds / len(models)
pred_csv(predictions=preds,
threshold=t,
name='transforms-res50-152dense-ensembels')
def test():
net = nn.DataParallel(densenet161().cuda())
net.load_state_dict(torch.load('models/densenet161.pth'))
net.eval()
dataset = KgForestDataset(split='test-61191',
transform=Compose([
Lambda(lambda x: toTensor(x)),
Normalize(mean=mean, std=std)
]),
height=256,
width=256,
label_csv=None)
test_loader = DataLoader(dataset,
batch_size=512,
shuffle=False,
pin_memory=True)
probs = predict(net, test_loader)
# probs = np.empty((61191, 17))
# current = 0
# for batch_idx, (images, im_ids) in enumerate(test_loader):
# num = images.size(0)
# previous = current
# current = previous + num
# logits = net(Variable(images.cuda(), volatile=True))
# prob = F.sigmoid(logits)
# probs[previous:current, :] = prob.data.cpu().numpy()
# print('Batch Index ', batch_idx)
pred_csv(probs, name='densenet161', threshold=BEST_THRESHOLD)
def test_majority_blender():
label_files = [f for f in glob.glob('submission_probs/*.txt') if 'blender' in f]
label = np.zeros((61191, 17))
for f in label_files:
l = np.loadtxt(f)
label = l + label
label = (label >= len(transforms)).astype(int)
pred_csv(label, 'blender')
def predict_test(t):
preds = np.zeros((61191, 17))
for index, model in enumerate(models):
name = str(model).split()[1]
net = nn.DataParallel(model().cuda())
net.eval()
net.load_state_dict(torch.load('models/{}.pth'.format(name)))
pred = predict(dataloader=test_dataloader, net=net)
preds = preds + pred
preds = preds / len(models)
pred_csv(predictions=preds, threshold=t, name='ensembles')
def predict_test_majority():
"""
Majority voting method.
"""
labels = np.empty((len(models), 61191, 17))
for m_idx, model in enumerate(models):
print('Loading name {}'.format(model))
name = str(model).split()[1]
# /home/jxu7/Research/planet-competition/submission_preds/full_data_resnet34_planet.txt
preds = np.loadtxt('submission_preds/full_data_{}.txt'.format(name))
labels[m_idx] = preds
labels = np.sum(labels, axis=0)
if len(models) % 2 == 0:
labels = (labels > (len(models) // 2)).astype(int)
elif len(models) == 1:
labels = labels
else:
labels = (labels >= (len(models) // 2)).astype(int)
# for m_idx, model in enumerate(models):
# name = str(model).split()[1]
# threshold = thresholds[name]
# print('predicting model {}'.format(name))
# print('threshold is', threshold)
# net = nn.DataParallel(model().cuda())
# net.load_state_dict(torch.load('models/full_data_{}.pth'.format(name)))
# net.eval()
# preds = np.zeros((61191, 17))
# for t in transforms:
# test_dataloader.dataset.images = t(test_dataloader.dataset.images)
# print(t, name)
# p = predict(net, dataloader=test_dataloader)
# preds = preds + (p > threshold).astype(int)
# # get predictions for the single model
# # preds = preds/len(transforms)
# # np.savetxt('submission_probs/full_data_{}.txt'.format(name), preds)
# # get labels
# # preds = (preds > thresholds[m_idx]).astype(int)
# preds = (preds > (len(transforms)//2)).astype(int)
# np.savetxt('submission_preds/full_data_{}.txt'.format(name), preds)
# labels[m_idx] = preds
# majority voting
# labels = labels.sum(axis=0)
# labels = (labels >= (len(models)//2)).astype(int)
pred_csv(predictions=labels, name='resnet101_resnet152_densenets')
def predict_test_majority():
"""
Majority voting method.
"""
labels = np.empty((len(models), 61191, 17))
for m_idx, model in enumerate(models):
print('Loading name {}'.format(model))
name = str(model).split()[1]
# /home/jxu7/Research/planet-competition/submission_preds/full_data_resnet34_planet.txt
preds = np.loadtxt('/mnt/home/dunan/Learn/Kaggle/planet_amazon/submission_probs/full_data_{}_single_10xlr_224and256.txt'.format(name))
labels[m_idx] = preds
labels = np.sum(labels, axis=0)
if len(models) % 2 == 0:
labels = (labels > (len(models)//2)).astype(int)
elif len(models) == 1:
labels = labels
else:
labels = (labels >= (len(models) // 2)).astype(int)
pred_csv(predictions=labels, name='resnet50152_densenet161169201')
def predict_test_averaging(t):
preds = np.zeros((61191, 17))
# imgs = test_dataloader.dataset.images.copy()
# iterate over models
for index, model in enumerate(models):
name = str(model).split()[1]
net = nn.DataParallel(model().cuda())
net.load_state_dict(torch.load('models/{}.pth'.format(name)))
net.eval()
# iterate over transformations
for transformation in transforms:
# imgs = transformation(imgs)
test_dataloader.dataset.images = transformation(
test_dataloader.dataset.images)
pred = predict(dataloader=test_dataloader, net=net)
preds = preds + pred
preds = preds / (len(models) * len(transforms))
# preds = preds / len(models)
np.savetxt('submission_probs/fpn_152', preds)
pred_csv(predictions=preds, threshold=t, name='fpn-152')
def predict_test(t):
preds = np.zeros((61191, 17))
# imgs = test_dataloader.dataset.images.copy()
# iterate over models
for index, model in enumerate(models):
name = str(model).split()[1]
net = nn.DataParallel(model().cuda())
net.load_state_dict(torch.load('models/{}.pth'.format(name)))
net.eval()
# iterate over transformations
for transformation in transforms:
# imgs = transformation(imgs)
test_dataloader.dataset.images = transformation(
test_dataloader.dataset.images)
pred = predict(dataloader=test_dataloader, net=net)
preds = preds + pred
preds = preds / (len(models) * len(transforms))
# preds = preds / len(models)
pred_csv(predictions=preds,
threshold=t,
name='transforms-resnet152_densenet161_densent169-ensembels')
0.22366667, 0.20366667, 0.17166667, 0.14633333, 0.20066667, 0.18966667,
0.197, 0.20166667, 0.17233333, 0.21466667, 0.15566667, 0.197,
0.16366667, 0.149, 0.25366667, 0.18333333, 0.26033333
]
]
models = [
resnet18_planet,
resnet34_planet,
resnet50_planet,
resnet152_planet,
densenet121,
densenet161,
densenet169,
]
labels = np.empty((len(models), 61191, 17))
for m_idx, model in enumerate(models):
name = str(model).split()[1]
preds = np.loadtxt(
'/mnt/home/dunan/Learn/Kaggle/planet_amazon/submission_probs/full_data_{}.txt'
.format(name))
# get labels
preds = (preds > thresholds[m_idx]).astype(int)
labels[m_idx] = preds
# majority voting
labels = labels.sum(axis=0)
labels = (labels >= (len(models) // 2)).astype(int)
pred_csv(predictions=labels, name='majority_voting_ensembles_full_data')
import pandas as pd
import numpy as np
import glob
from data import kgdataset
from data.kgdataset import KgForestDataset
kgdataset.KAGGLE_DATA_DIR = '../../../kaggle'
from util import name_idx, pred_csv
filenames = glob.glob('submissions/*.csv')
print('\n'.join(filenames))
# read all csv files and convert tags to labels
dfs = [pd.read_csv(name) for name in filenames]
labels = np.empty((8, 61191, 17))
for df_idx, df in enumerate(dfs):
print(df_idx)
for row, tag in enumerate(df['tags']):
label = np.zeros(17)
idx = [name_idx()[name] for name in tag.split(' ')]
label[idx] = 1
labels[df_idx, row, :] = label
majority_voting = labels.sum(axis=0)
majority_voting = (majority_voting >= 4).astype(int)
pred_csv(majority_voting, name='sub_ensembles')
