Nolearn/Lasagne神经网络没有开始训练

Question

我正在使用Theano 0.7,nolearn 0.6adev并lasagne 0.2.dev1在GPU(在IPython 3.2.1笔记本电脑中)训练神经网络.但是,由于第一层('reduc'),以下网络没有开始训练,等待几个小时后:

import theano
from lasagne import layers
from lasagne.updates import nesterov_momentum
from nolearn.lasagne import NeuralNet
from nolearn.lasagne import BatchIterator
from lasagne import nonlinearities
from lasagne import init
import numpy as np

testNet = NeuralNet(
    layers=[(layers.InputLayer, {"name": 'input', 'shape': (None, 12, 1000, )}),

            (layers.Conv1DLayer, {"name": 'reduc', 'filter_size': 1, 'num_filters': 4, 
                                  "nonlinearity":nonlinearities.linear,}),

            (layers.Conv1DLayer, {"name": 'conv1', 'filter_size': 25, 'num_filters': 100,
                                 'pad': 'same', }),
            (layers.MaxPool1DLayer, {'name': 'pool1', 'pool_size': 5, 'stride': 3}),



            (layers.Conv1DLayer, {"name": 'conv2', 'filter_size': 15, 'num_filters': 100,
                                 'pad': 'same', 
                                  'nonlinearity': nonlinearities.LeakyRectify(0.2)}),

            (layers.MaxPool1DLayer, {'name': 'pool2', 'pool_size': 5, 'stride': 2}),

            (layers.Conv1DLayer, {"name": 'conv3', 'filter_size': 9, 'num_filters': 100,
                                 'pad': 'same', 
                                  'nonlinearity': nonlinearities.LeakyRectify(0.2)}),

            (layers.MaxPool1DLayer, {'name': 'pool3', 'pool_size': 2}),

            (layers.Conv1DLayer, {"name": 'conv4', 'filter_size': 5, 'num_filters': 20,
                                 'pad': 'same', }),

            (layers.Conv1DLayer, {"name": 'conv5', 'filter_size': 3, 'num_filters': 20,
                                 'pad': 'same',}),


            (layers.DenseLayer, {"name": 'hidden1', 'num_units': 10, 
                                 'nonlinearity': nonlinearities.rectify}),

            (layers.DenseLayer, {"name": 'output', 'nonlinearity': nonlinearities.sigmoid, 
                                 'num_units': 5})
           ],

    # optimization method:
    update=nesterov_momentum,
    update_learning_rate=5*10**(-3),
    update_momentum=0.9,

    regression=True,  
    max_epochs=1000,
    verbose=1,
    )

testNet.fit(np.random.random([3000, 12, 1000]).astype(np.float32), 
            np.random.random([3000, 5]).astype(np.float32))

如果我注释掉第一层,训练将在几秒钟后开始.培训更复杂的网络也不是问题.是什么导致了这个问题？

编辑:奇怪的是,如果我删除conv4和conv5,训练也开始在合理的时间范围内.

编辑2:更奇怪的是,如果我在图层中将过滤器的大小更改为10 reduc,那么培训将在合理的时间内开始.如果之后我停止了单元格的执行,将此值更改为1,然后重新执行单元格,培训就可以了......

最后我开始使用另一个框架,但如果有人感兴趣,这里是我在lasagne用户组上开始的线程的链接.