submit2 = pd.read_csv('Christof/assets/sample_submission.csv')
# custom thresholds to match lb proportions
thresholds = np.linspace(0.95, 0.05, 101)
pred = draw_predict_map32.copy()
for j in tqdm(range(pred.shape[1])):
for t in thresholds:
pred[:, j] = (draw_predict_map32[:, j] > t).astype(int)
prop = np.mean(pred[:, j])
if prop >= desired[j]: break
print(j, '%3.2f' % t, '%6.4f' % desired[j], '%6.4f' % prop, j, )
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit2['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit2.to_csv(MODEL_PATH + 'submission_loss{}_lb_dist_adjusted_map32_8tta.csv'.format(exp_suffix), index=False)
from Christof.utils import f1_sub
sub_best = pd.read_csv('ens56d.csv')
sub2 = pd.read_csv(MODEL_PATH + 'submission_loss{}_lb_dist_adjusted_8tta.csv'.format(exp_suffix))
f1_sub(sub_best,submit)
f1_sub(sub_best,submit2)
sub2 = pd.read_csv(MODEL_PATH + 'submission_loss{}_lb_dist_adjusted_8tta.csv'.format(exp_suffix))
f1_sub(sub_best,sub2)
j,
'%3.2f' % t,
'%6.4f' % desired[j],
'%6.4f' % prop,
j,
)
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit3['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit3.to_csv(
MODEL_PATH +
'submission_loss{}_lb_dist_adjusted_mean_loss_map32_8tta.csv'.format(
exp_suffix),
index=False)
from Christof.utils import f1_sub
sub_best = pd.read_csv('ens56d.csv')
print(f1_sub(sub_best, submit))
print(f1_sub(sub_best, submit2))
print(f1_sub(sub_best, submit3))
s2 = pd.read_csv(
'Christof/models/GAPNet/13_ext/submission_loss_1_lb_dist_adjusted_8tta.csv'
)
f1_sub(sub_best, s2)
for i in range(best_sub.shape[0]):
for j in p0[i]: y0[i, int(j)] = 1
for i in range(28):
desired[i] = y0[:,i].mean()
thresholds = np.linspace(0.95, 0.05, 101)
pred = preds.copy()
for j in range(pred.shape[1]):
for t in thresholds:
pred[:, j] = (preds[:, j] > t).astype(int)
prop = np.mean(pred[:, j])
if prop >= desired[j]: break
print(j, '%3.2f' % t, '%6.4f' % desired[j], '%6.4f' % prop, j, )
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit = pd.read_csv('Christof/assets/sample_submission.csv')
submit['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
#submit.to_csv(MODEL_PATH + 'submission.csv', index=False)
from Christof.utils import f1_sub
best_sub = pd.read_csv('ens56d.csv')
f1_sub(best_sub,submit)
j,
'%3.2f' % t,
'%6.4f' % desired[j],
'%6.4f' % prop,
j,
)
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit2['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit2.to_csv(
MODEL_PATH +
'submission_loss{}_lb_dist_adjusted_map32_8tta.csv'.format(exp_suffix),
index=False)
from Christof.utils import f1_sub
sub_best = pd.read_csv('ens18.csv')
sub2 = pd.read_csv(
MODEL_PATH +
'submission_loss{}_lb_dist_adjusted_8tta.csv'.format(exp_suffix))
f1_sub(sub_best, sub2)
sub2 = pd.read_csv(
MODEL_PATH +
'submission_loss{}_lb_dist_adjusted_8tta.csv'.format(exp_suffix))
f1_sub(sub_best, sub2)
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit = pd.read_csv('Christof/assets/sample_submission.csv')
submit['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit.to_csv(MODEL_PATH + 'submission_5crop_stacked_650_freq.csv',
index=False)
from Christof.utils import f1_sub
best_sub = pd.read_csv('ens56d.csv')
f1_sub(best_sub, submit)
sub_650 = pd.read_csv(
'enstw43_10en36-642_10en39b-642_10hill2-636_10-646_2.4.csv')
f1_sub(sub_650, submit)
gap13_ext = pd.read_csv(
'Christof/models/GAPNet/13_ext/submission_loss_5fold_mean.csv')
gap13_ext_60_40 = pd.read_csv(
'Christof/models/GAPNet/13_ext_60_40/submission_loss_5fold_mean.csv')
f1_sub(gap13_ext_60_40, gap13_ext)
f1_sub(gap13_ext, submit)
preds2 = np.load('Christof/models/GAPNet/13_ext_60_40/pred_5fold.npy')
if prop >= desired[j]: break
print(
j,
'%3.2f' % t,
'%6.4f' % desired[j],
'%6.4f' % prop,
j,
)
print(pred[:5].astype(int))
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit = pd.read_csv('Christof/assets/sample_submission.csv')
submit['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit.to_csv(MODEL_PATH + 'submission_hc.csv', index=False)
from Christof.utils import f1_sub
good_sub = pd.read_csv('ens56d.csv')
print(f1_sub(good_sub, submit))
best_sub = pd.read_csv(
'enstw43_10en36-642_10en39b-642_10hill2-636_10-646_2.4.csv')
print(f1_sub(best_sub, submit))
sub2 = pd.read_csv('Christof/hillclimb/blends/2/submission_hc.csv')
print(f1_sub(best_sub, sub2))
preds[:, i] = preds[:, i] > threshs[i]
from sklearn.metrics import f1_score
f1_score(y0, preds, average='macro')
submit = pd.read_csv('Christof/assets/sample_submission.csv')
tta = 8
draw_predict = np.zeros((5, len(submit['Id']), 28))
for fold_id in range(5):
draw_predict[fold_id] = np.load(MODEL_PATH + f'pred{fold_id}.npy')
draw_predict1 = np.mean(draw_predict, axis=0)
np.save(MODEL_PATH + f'pred_5fold_3.npy', draw_predict1)
pred = draw_predict1.copy()
for i in range(28):
pred[:, i] = (pred[:, i] > threshs[i]).astype(int)
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit['Predicted'] = str_predict_label
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit.to_csv(MODEL_PATH + 'submission_loss_5fold_oof_treshs.csv', index=False)
from Christof.utils import f1_sub
best_sub = pd.read_csv('ens56d.csv')
f1_sub(best_sub, submit)
old_sub = pd.read_csv(MODEL_PATH + 'submission_loss_5fold_mean.csv')
f1_sub(best_sub, old_sub)
for id, prefix in enumerate(['ul','ur','mm','bl','br']):
draw_predict[id] = np.load(MODEL_PATH + f'pred1_{prefix}.npy')
thresholds = np.linspace(0.95, 0.05, 101)
pred = draw_predict[id].copy()
for j in range(pred.shape[1]):
for t in thresholds:
pred[:, j] = (draw_predict[id][:, j] > t).astype(int)
prop = np.mean(pred[:, j])
if prop >= desired[j]: break
label_predict = [np.arange(28)[score_predict == 1] for score_predict in pred]
str_predict_label = [' '.join(str(l) for l in lp) for lp in label_predict]
submit['Predicted'] = str_predict_label
print(f1_sub(best_sub, submit))
# np.save('draw_predict_InceptionV3.npy', score_predict)
submit.to_csv(MODEL_PATH + 'submission_5crop_mean.csv', index=False)
#draw_predict1 = np.mean(draw_predict,axis = 0)
draw_predict1 = draw_predict[0] + draw_predict[1] +1.5*draw_predict[2] +draw_predict[3]+draw_predict[4]
draw_predict1 = draw_predict1 /5.5
np.save(MODEL_PATH + f'5crop_mean.npy',draw_predict1)
# custom thresholds to match lb proportions
thresholds = np.linspace(0.95, 0.05, 101)
pred = draw_predict1.copy()