def Main():
# 初期化
Initialize()
# 学習済みモデルを使用する場合
if gv.G.IsLoadPickle:
with open(gv.G.LoadPicklePath, "rb") as p:
gv.G.Model = pickle.load(p)
# End_With
# 学習する場合
else:
tr.Training()
# End_IfElse
# 実験
ex.UnknownDataExperiment(-1)
ex.KnownDataExperiment(-1)
# 保存
SaveModelAsPickle()
def Training():
print("\r\nTraining...\r\n")
for epoch in range(gv.G.EpochNum):
# 学習関係
totalLoss = 0
totalAccuracy = 0
for i in range(math.ceil((gv.G.DataSetSize - gv.G.Be.StartNowTrainNum) / gv.G.MinibatchSize)):
# ミニバッチ学習
x = Variable(gv.G.xp.array(gv.G.Be.GetFlattenData(gv.G.TrainDataIndex), dtype=gv.G.xp.float32))
t = Variable(gv.G.xp.array(gv.G.Be.GetFlattenData(gv.G.TrainLabelIndex), dtype=gv.G.xp.int32))
y = gv.G.Model(x)[gv.G.TargetLayer]
gv.G.Model.zerograds()
loss = F.softmax_cross_entropy(y, t)
acc = F.accuracy(y, t)
loss.backward()
gv.G.Optimizer.update()
totalLoss += loss.data * gv.G.MinibatchSize
totalAccuracy += acc.data * gv.G.MinibatchSize
# 途中経過表示
ydata = []
for ngo in y.data[0:10]:
ydata.append(int(np.argmax(chainer.cuda.to_cpu(ngo))))
# End_For
# print("t:{} y:{}".format(t[0:10], ydata) + " loss.data:{:*<8} acc.data:{:*<8}".format(
# round(float(loss.data), 6), round(float(acc.data), 6)))
print(f'\r{len(x)}')
if gv.G.IsDebugLog:
gv.G.LogFile.write("t:{} y:{} loss.data{:*<8} acc.data{:*<8}\r\n".format(
t[0:10], ydata, round(float(loss.data), 6), round(float(acc.data), 6)))
gv.G.LogFile.flush()
# End_If
# 次のデータセットへｲｸｿﾞ!
gv.G.Be.NextTrainData()
# End_For(ミニバッチの終了)
# エポック毎の誤差・Acc表示
print(
"---------------------------------------------↑{:^5}↑---------------------------------------------".format(epoch))
print("TotalLoss:{}".format(totalLoss / gv.G.Be.TrainSize))
print("TotalAcc:{}".format(totalAccuracy / gv.G.Be.TrainSize))
print("---------------------------------------------------------------------------------------------------\r\n\r\n")
# デバッグログ
if gv.G.IsDebugLog:
gv.G.LogFile.write(
"---------------------------------------------↑{:^5}↑---------------------------------------------\r\n".format(epoch))
gv.G.LogFile.write(
"TotalLoss:{}\r\n".format(
totalLoss / gv.G.TrainSize))
gv.G.LogFile.write(
"TotalAcc:{}\r\n".format(
totalAccuracy /
gv.G.TrainSize))
gv.G.LogFile.write(
"---------------------------------------------------------------------------------------------------\r\n\r\n\r\n")
gv.G.LogFile.flush()
# End_If
# 各エポック毎に認識実験する場合
if gv.G.IsTestRecognitionEachEpoch:
ex.UnknownDataExperiment(epoch)
ex.KnownDataExperiment(epoch)
# End_If
# add kauzya
if gv.G.is_every_save_pickle:
print(f' Save pickle: {epoch}')
dir_path = f'{gv.G.SavePicklePath}/epoch/'
if not os.path.exists(dir_path):
os.makedirs(dir_path)
with open(f'{dir_path}{epoch}_epoch_{gv.G.DateTimeNow}.pkl', "wb") as p:
pickle.dump(gv.G.Model, p)