在 Theano/lasagne 中训练神经网络时实时情节损失

Live plot losses while training neural net in Theano/lasagne

我正在 Theano 和千层面中训练神经网络,运行 iPython 笔记本中的代码。我喜欢在每次迭代时显示训练和有效损失,如下所示:

epoch    train loss    valid loss    train/val    valid acc  dur
-------  ------------  ------------  -----------  -----------  -----
  1       0.53927       0.22774      2.36794      0.93296  5.45s
  2       0.28789       0.16561      1.73840      0.95033  5.40s

但我也想看看live/dynamic两次损失的情节。有内置的方法吗?

我已经尝试创建一个自定义 class 并将其添加到我的网络 on_epoch_finished,但要么我在每次迭代时得到一个新图(我想要一个,更新),要么我必须在每次迭代时删除以前的输出,因此看不到文本输出(我想保留)。

我终于设法按照我的意愿更新了损失图,使用以下 class:

from IPython import display
from matplotlib import pyplot as plt
import numpy as np
from lasagne import layers
from lasagne.updates import nesterov_momentum
from nolearn.lasagne import NeuralNet
from lasagne import nonlinearities

class PlotLosses(object):
    def __init__(self, figsize=(8,6)):
        plt.plot([], []) 

    def __call__(self, nn, train_history):
        train_loss = np.array([i["train_loss"] for i in nn.train_history_])
        valid_loss = np.array([i["valid_loss"] for i in nn.train_history_])

        plt.gca().cla()
        plt.plot(train_loss, label="train") 
        plt.plot(valid_loss, label="test")

        plt.legend()
        plt.draw()

和要重现的代码示例:

net_SO = NeuralNet(
    layers=[(layers.InputLayer, {"name": 'input', 'shape': (None, 1, 28, 28)}),
            (layers.Conv2DLayer, {"name": 'conv1', 'filter_size': (3,3,), 'num_filters': 5}),
            (layers.DropoutLayer, {'name': 'dropout1', 'p': 0.2}),
            (layers.DenseLayer, {"name": 'hidden1', 'num_units': 50}),
            (layers.DropoutLayer, {'name': 'dropout2', 'p': 0.2}),
            (layers.DenseLayer, {"name": 'output', 'nonlinearity': nonlinearities.softmax, 'num_units': 10})],
    # optimization method:
    update=nesterov_momentum,
    update_learning_rate=10**(-2),
    update_momentum=0.9,

    regression=False,  
    max_epochs=200, 
    verbose=1,

    on_epoch_finished=[PlotLosses(figsize=(8,6))], #this is the important line
    )

net_SO.fit(X, y) #X and y from the MNIST dataset