SIGNATEの画像ラベリング（２０種類）の暫定評価向上について

#質問
今回は初挑戦でSIGNATEの練習問題である画像ラベリング(20種類)に現在挑戦しています。
今まで１０回ほど投稿していきましたが、投稿するたびに暫定評価値が上がっていっている状態です。
(※今回評価値は0以上の値をとり、精度が高いほど小さな値になる)

１回目の投稿では、暫定評価が約3.8、10回目では約4.7になっています。
しかし、model lossを出力させてみると、明らかに１回目より、１０回目の方がlossの値は小さくなっています。
ではなぜ、今回暫定評価はlossの値が高い1回目の方が好成績なのかがよくわかりません。

アドバイス等あればよろしくお願いいたします。

#1回目のコード

python
1import numpy as np
2import pandas as pd
3from matplotlib import pyplot as plt
4from skimage import io
5from keras import utils
6from keras.models import Sequential
7from keras.layers import Dense, Flatten, Activation
8from keras.layers import Conv2D, MaxPooling2D
9from keras import optimizers
10
11labels= pd.read_csv("train_master.tsv", sep="\t")
12master = pd.read_csv("label_master.tsv", sep="\t")
13sample = pd.read_csv("sample_submit.csv", header=None, sep=",")
14
15train_images = []
16for fname in labels["file_name"]:
17    path = "./train/" + fname
18    img = io.imread(path)
19    train_images.append(img)
20train_images = np.array(train_images)
21print(type(train_images), train_images.shape)
22
23test_images = []
24for fname in sample[0]:
25    path = "./test/" + fname
26    img = io.imread(path)
27    test_images.append(img)
28test_images = np.array(test_images )
29print(type(test_images ), test_images.shape)
30
31train_images = train_images / 255
32test_images = test_images / 255
33
34y = labels["label_id"]
35
36y_categorical = utils.to_categorical(y)
37y_categorical
38
39X_con_image, X_ver_image = np.split(train_images, [40000])
40y_con_label, y_ver_label = np.split(y_categorical, [40000])
41
42model = Sequential()
43
44model.add(Conv2D(filters=6, kernel_size=(3,3), padding="same", input_shape=(32,32,3)))
45model.add(Activation("sigmoid"))
46model.add(MaxPooling2D(pool_size=(2,2)))
47
48model.add(Conv2D(filters=12, kernel_size=(3,3), padding="same"))
49model.add(Activation("sigmoid"))
50model.add(MaxPooling2D(pool_size=(2,2)))
51
52model.add(Flatten())
53
54model.add(Dense(units=120))
55model.add(Activation("sigmoid"))
56model.add(Dense(units=60))
57model.add(Activation("sigmoid"))
58model.add(Dense(units=20))
59model.add(Activation("softmax"))
60
61model.compile(loss="categorical_crossentropy",
62             optimizer=optimizers.SGD(lr=0.05, momentum=0.9, decay=0.0, nesterov=True))
63
64batch_size=100
65epochs=20
66
67history = model.fit(X_con_image, y_con_label,
68                    batch_size=batch_size,
69                    epochs=epochs,
70                    verbose=1,
71                    validation_data=(X_ver_image, y_ver_label))

SIGNATEが出した暫定評価: 3.81172

#10回目のコード

python
1import numpy as np
2import pandas as pd
3from matplotlib import pyplot as plt
4from skimage import io
5from keras import utils
6from keras.models import Sequential
7from keras.layers import Dense, Flatten,Dropout, Activation
8from keras.layers.normalization import BatchNormalization
9from keras.layers.pooling import GlobalAveragePooling2D
10from keras.layers import Conv2D, MaxPooling2D
11from keras import optimizers
12
13labels= pd.read_csv("train_master.tsv", sep="\t")
14master = pd.read_csv("label_master.tsv", sep="\t")
15sample = pd.read_csv("sample_submit.csv", header=None, sep=",")
16
17train_images = []
18for fname in labels["file_name"]:
19    path = "./train/" + fname
20    img = io.imread(path)
21    train_images.append(img)
22train_images = np.array(train_images)
23print(type(train_images), train_images.shape)
24
25test_images = []
26for fname in sample[0]:
27    path = "./test/" + fname
28    img = io.imread(path)
29    test_images.append(img)
30test_images = np.array(test_images )
31print(type(test_images ), test_images.shape)
32
33train_images = train_images / 255
34test_images = test_images / 255
35
36y = labels["label_id"]
37
38y_categorical = utils.to_categorical(y)
39y_categorical
40
41X_con_image, X_ver_image = np.split(train_images, [40000])
42y_con_label, y_ver_label = np.split(y_categorical, [40000])
43
44model = Sequential()
45
46model.add(Conv2D(filters=64, kernel_size=(3,3), padding="same", input_shape=(32,32,3)))
47model.add(Activation("relu"))
48model.add(BatchNormalization())
49model.add(Dropout(0.25))
50
51model.add(Conv2D(filters=64, kernel_size=(3,3), padding="same"))
52model.add(Activation("relu"))
53model.add(BatchNormalization())
54model.add(Dropout(0.25))
55
56model.add(Conv2D(filters=64, kernel_size=(3,3), padding="same"))
57model.add(Activation("relu"))
58model.add(BatchNormalization())
59model.add(Dropout(0.25))
60
61model.add(Conv2D(filters=64, kernel_size=(3,3), padding="same"))
62model.add(Activation("relu"))
63model.add(BatchNormalization())
64model.add(Dropout(0.25))
65
66model.add(Conv2D(filters=64, kernel_size=(3,3), padding="same"))
67model.add(Activation("relu"))
68model.add(BatchNormalization())
69model.add(Dropout(0.25))
70
71model.add(MaxPooling2D(pool_size=(2, 2)))
72model.add(GlobalAveragePooling2D())
73
74model.add(Dense(units=20))
75model.add(Activation("softmax"))
76
77model.compile(loss="categorical_crossentropy",
78             optimizer=optimizers.rmsprop(lr=0.0001, decay=1e-6),
79              metrics=["accuracy"])
80
81batch_size=32
82epochs=20
83
84history = model.fit(X_con_image, y_con_label,
85                    batch_size=batch_size,
86                    epochs=epochs,
87                    verbose=1,
88                    validation_data=(X_ver_image, y_ver_label))