keras でニューラルネット（最適なモデルを使用したい）

Question

下記にプログラムと出力結果を載せますが、出力結果のlossはどうやって求めていますか？

# 必要ライブラリのインポート
from keras.models import Sequential
from keras.layers import Dense,Dropout
from keras import losses
from keras import optimizers
from keras import callbacks
import os

# モデル生成
model = Sequential()

# 層の追加
layers=[
    Dense(128, activation='elu'),
    Dropout(0.01),
    Dense(128, activation='selu'),
    Dropout(0.01),
    Dense(64, activation='softmax'),
    Dense(1, activation='linear')
]
for layer in layers:
    model.add(layer)

# モデルの学習設定

model.compile(
    loss=losses.mean_squared_error,
    optimizer=optimizers.Adam(),
    metrics=['acc']
)

# モデルの学習
result = model.fit(
    X_train_n,
    y_train_n,
    batch_size=32,
    epochs=100, 
    callbacks=[
        callbacks.ModelCheckpoint(
            filepath = './model/best_model.h5', 
            monitor='loss', 
            save_best_only=True,
        )
    ]
)

Epoch 1/100
354/354 [==============================] - 0s 644us/step - loss: 0.1649 - acc: 0.0056
Epoch 2/100
354/354 [==============================] - 0s 116us/step - loss: 0.0638 - acc: 0.0056
Epoch 3/100
354/354 [==============================] - 0s 82us/step - loss: 0.0466 - acc: 0.0056
Epoch 4/100
354/354 [==============================] - 0s 82us/step - loss: 0.0442 - acc: 0.0056
Epoch 5/100
354/354 [==============================] - 0s 96us/step - loss: 0.0430 - acc: 0.0056
Epoch 6/100
354/354 [==============================] - 0s 90us/step - loss: 0.0421 - acc: 0.0056
Epoch 7/100
354/354 [==============================] - 0s 82us/step - loss: 0.0412 - acc: 0.0056
Epoch 8/100
354/354 [==============================] - 0s 85us/step - loss: 0.0400 - acc: 0.0056
Epoch 9/100
354/354 [==============================] - 0s 87us/step - loss: 0.0379 - acc: 0.0056
Epoch 10/100
354/354 [==============================] - 0s 90us/step - loss: 0.0355 - acc: 0.0056
Epoch 11/100
354/354 [==============================] - 0s 90us/step - loss: 0.0331 - acc: 0.0056
Epoch 12/100
354/354 [==============================] - 0s 101us/step - loss: 0.0306 - acc: 0.0056
・・・

下記のnp.mean(losses.mean_squared_error())で出力結果のLOSSと等しくなると思っていましたが、なりません。なぜでしょうか？
また、最適なモデルの使用方法は下記の使用方法で問題ありませんでしょうか？

# 学習での最終モデルを使用した損失算出
y_predict_n = model.predict(X_train_n)
np.mean(losses.mean_squared_error(y_predict_n,y_train_n))

# 最適モデルの損失算出
model = keras.models.load_model('./model/best_model.h5')
y_predict_n = model.predict(X_train_n)
np.mean(losses.mean_squared_error(y_predict_n,y_train_n))