pix2pixによる画像の学習について

前提・実現したいこと

pix2pixを用いてある画像から予測したある画像を生成したいです。

発生している問題・エラーメッセージ

こちらのGithubを参考にし、自前のデータセットで学習できるよう書き換えをしているのですが、

AssertionError: Height in the output should be positive.

というエラーが出て困っています。

試したこと

調べた際に『画像を畳み込んだら画像がなくなったのでは』と出てきたのでpaddingを調べたのですが、それを行っているConvolution2Dのところを見てもどうしたらいいのかが分からないです…。

class CBR内のL.Convolution2D(ch0, ch1, 4, 2, 1, initialW=w)をL.Convolution2D(ch0, ch1, 4, 1, 1, initialW=w)にしたところ、エラーが表示されなくなりましたが固まったようになり、動きません。
これはうまくいってて学習中なのか、無限ループかなにかに陥っているのかが分かりません。

python
1# -*- coding: utf-8 -*-
2"""
3Created on Wed Jan  9 16:52:30 2019
4
5@author: KNO3
6"""
7
8#!/usr/bin/env python
9
10from __future__ import print_function
11
12import numpy
13
14import chainer
15from chainer import cuda
16import chainer.functions as F
17import chainer.links as L
18
19# U-net https://arxiv.org/pdf/1611.07004v1.pdf
20
21# convolution-batchnormalization-(dropout)-relu
22class CBR(chainer.Chain):
23    def __init__(self, ch0, ch1, bn=True, sample='down', activation=F.relu, dropout=False):
24        self.bn = bn
25        self.activation = activation
26        self.dropout = dropout
27        layers = {}
28        w = chainer.initializers.Normal(0.02)
29        if sample=='down':
30            layers['c'] = L.Convolution2D(ch0, ch1, 4, 2, 1, initialW=w)
31        else:
32            layers['c'] = L.Deconvolution2D(ch0, ch1, 4, 2, 1, initialW=w)
33        if bn:
34            layers['batchnorm'] = L.BatchNormalization(ch1)
35        super(CBR, self).__init__(**layers)
36        
37    def __call__(self, x):
38        h = self.c(x)
39        if self.bn:
40            h = self.batchnorm(h)
41        if self.dropout:
42            h = F.dropout(h)
43        if not self.activation is None:
44            h = self.activation(h)
45        return h
46    
47class Encoder(chainer.Chain):
48    def __init__(self, in_ch):
49        layers = {}
50        w = chainer.initializers.Normal(0.02)
51        layers['c0'] = L.Convolution2D(in_ch, 64, 3, 1, 1, initialW=w)
52        layers['c1'] = CBR(64, 128, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
53        layers['c2'] = CBR(128, 256, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
54        layers['c3'] = CBR(256, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
55        layers['c4'] = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
56        layers['c5'] = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
57        layers['c6'] = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
58        layers['c7'] = CBR(512, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
59        super(Encoder, self).__init__(**layers)
60
61    def __call__(self, x):
62        hs = [F.leaky_relu(self.c0(x))]
63        for i in range(1,8):
64            hs.append(self['c%d'%i](hs[i-1]))
65        return hs
66
67class Decoder(chainer.Chain):
68    def __init__(self, out_ch):
69        layers = {}
70        w = chainer.initializers.Normal(0.02)
71        layers['c0'] = CBR(512, 512, bn=True, sample='up', activation=F.relu, dropout=True)
72        layers['c1'] = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=True)
73        layers['c2'] = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=True)
74        layers['c3'] = CBR(1024, 512, bn=True, sample='up', activation=F.relu, dropout=False)
75        layers['c4'] = CBR(1024, 256, bn=True, sample='up', activation=F.relu, dropout=False)
76        layers['c5'] = CBR(512, 128, bn=True, sample='up', activation=F.relu, dropout=False)
77        layers['c6'] = CBR(256, 64, bn=True, sample='up', activation=F.relu, dropout=False)
78        layers['c7'] = L.Convolution2D(128, out_ch, 3, 1, 1, initialW=w)
79        super(Decoder, self).__init__(**layers)
80
81    def __call__(self, hs):
82        h = self.c0(hs[-1])
83        for i in range(1,8):
84            h = F.concat([h, hs[-i-1]])
85            if i<7:
86                h = self['c%d'%i](h)
87            else:
88                h = self.c7(h)
89        return h
90
91    
92class Discriminator(chainer.Chain):
93    def __init__(self, in_ch, out_ch):
94        layers = {}
95        w = chainer.initializers.Normal(0.02)
96        layers['c0_0'] = CBR(in_ch, 32, bn=False, sample='down', activation=F.leaky_relu, dropout=False)
97        layers['c0_1'] = CBR(out_ch, 32, bn=False, sample='down', activation=F.leaky_relu, dropout=False)
98        layers['c1'] = CBR(64, 128, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
99        layers['c2'] = CBR(128, 256, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
100        layers['c3'] = CBR(256, 512, bn=True, sample='down', activation=F.leaky_relu, dropout=False)
101        layers['c4'] = L.Convolution2D(512, 1, 3, 1, 1, initialW=w)
102        super(Discriminator, self).__init__(**layers)
103
104    def __call__(self, x_0, x_1):
105        h = F.concat([self.c0_0(x_0), self.c0_1(x_1)])
106        h = self.c1(h)
107        h = self.c2(h)
108        h = self.c3(h)
109        h = self.c4(h)
110        #h = F.average_pooling_2d(h, h.data.shape[2], 1, 0)
111        return h
112