def Exe_oof(model,origin_state, train_images,train_labels,test_images, optim_flg,epoch_num,str_name,num_acc,num_fold):
cv = StratifiedKFold(n_splits=Nsplits, shuffle=True, random_state=42)
oof_pred = np.zeros( (len(train_images),10 ) )
y_pred_max = np.zeros( (len(test_images),10 ) )
str_model = str_name +'_0_0000'
y_pred_flg=0
for fold, (train_idx, val_idx) in enumerate(cv.split(train_images, train_labels)):
print(fold)
print(num_fold)
if fold==num_fold:
print('True')
max_acc = num_acc
model.load_state_dict(origin_state)
#データを訓練用と検証用に分ける
train_x, valid_x = train_images[train_idx], train_images[val_idx]
train_y, valid_y = train_labels[train_idx], train_labels[val_idx]
#データをローダーに格納する。
train_loader = dataset.DataLoader(
Dataset(train_x, train_y, HoriFlip_p, RandCont_p, RandBrig_p, Blur_p,Cut_p, train=True), batch_size=64, shuffle=True)
valid_loader = dataset.DataLoader(
Dataset(valid_x, valid_y, HoriFlip_p, RandCont_p, RandBrig_p, Blur_p, Cut_p,train=False), batch_size=64, shuffle=False)
if optim_flg == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=lr_par)
else:
optimizer = optim.SGD(model.parameters(), lr=lr_par, weight_decay=0.01, momentum=0.9)
log = []
import datetime
print(datetime.datetime.now())
for epoch in range(epoch_num):
model.train()
with autograd.detect_anomaly():
train_loss, train_accuracy = perform(model, train_loader, optimizer)
model.eval()
with torch.no_grad():
valid_loss, valid_accuracy = perform(model, valid_loader, None)
print('[{}] train(loss/accuracy)={:.2f}/{:.2f}, valid(loss/accuracy)={:.2f}/{:.2f}'.format(epoch + 1, train_loss, train_accuracy, valid_loss, valid_accuracy))
if valid_accuracy > max_acc:
int_model = int(valid_accuracy*10000)
str_model = str_name +'_'+ str(fold) +'_'+ str(int_model)
DATA_DIR = 'drive/My Drive/ProbSpace/Ukiyoe/'+ str_name + '/' + str_model
torch.save(model.state_dict(), DATA_DIR)
#################################
# 検証用データの予測
#################################
model.eval
hogehoge_i = 1
for images, _ in valid_loader:
images = images.cuda()
with torch.no_grad():
if hogehoge_i>=1 :
preds = model(images)
hogehoge_i=0
else:
preds = torch.cat([preds,model(images)],dim=0)
dvs=torch.device('cpu')
preds_cpu = preds.to(dvs)
preds_np = preds_cpu.clone().numpy()
oof_pred[val_idx] = preds_np
#################################
# テスト用データの予測
#################################
dummy_labels = np.zeros( (test_images.shape[0], 1), dtype=np.int64)
test_loader = dataset.DataLoader(Dataset(test_images, dummy_labels,0.,0.,0.,0.,0.), batch_size=64, shuffle=False)
model.eval()
hogehoge_i = 1
for images, _ in test_loader:
images = images.cuda()
with torch.no_grad():
if hogehoge_i>=1 :
preds = model(images)
hogehoge_i=0
else:
preds = torch.cat([preds,model(images)],dim=0)
y_pred_max = preds
#################################
max_acc = valid_accuracy
print(datetime.datetime.now())
log.append( (epoch + 1, train_loss, train_accuracy, valid_loss, valid_accuracy) )
if y_pred_flg==0 :
y_pred = y_pred_max
y_pred_flg=1
else:
y_pred = torch.cat([y_pred,y_pred_max],dim=1)
print('Partial score of fold {} is: {}'.format(fold, max_acc))
）と、一つ前の出力（
）に、それぞれ重みをかけたものにバイアスを足して、最後に
）
![[商品価格に関しましては、リンクが作成された時点と現時点で情報が変更されている場合がございます。]](https://hbb.afl.rakuten.co.jp/hgb/0f083428.235a9254.0f083429.dafd0742/?me_id=1213310&item_id=19759104&m=https%3A%2F%2Fthumbnail.image.rakuten.co.jp%2F%400_mall%2Fbook%2Fcabinet%2F8434%2F9784297108434.jpg%3F_ex%3D80x80&pc=https%3A%2F%2Fthumbnail.image.rakuten.co.jp%2F%400_mall%2Fbook%2Fcabinet%2F8434%2F9784297108434.jpg%3F_ex%3D128x128&s=128x128&t=picttext "[商品価格に関しましては、リンクが作成された時点と現時点で情報が変更されている場合がございます。]")
