我正在使用平衡的CT扫描数据集。我使用keras层（即squeezenet）构建我的自定义网络。
我正在此网络上训练我的模型，但是我的训练精度停滞不前，验证精度为0。我经历了一些代码，并建议将relu更改为tanh。但是没有任何改善。
我对新数据集的谓词始终是同一类。
def build_squeezenet（input_shape）：
input_layer = Input(shape = input_shape) # ( 50,50,1)
out = Conv2D(96, kernel_size =(3, 3), activation='relu')(input_layer)
out = MaxPooling2D(pool_size=2, strides=None, padding='valid')(out)
# no mxpooling done since small images
out = fire_module(out,squeeze =16 , expansion =64)
out = fire_module(out,squeeze = 16 , expansion=64)
out = fire_module(out,squeeze= 32,expansion = 128)
out = fire_module(out,squeeze= 32,expansion = 128)
out = fire_module(out , squeeze = 48 ,expansion=192)
out = fire_module(out,squeeze = 48 , expansion = 192)
out = fire_module(out,squeeze = 64 , expansion = 256)
out = fire_module(out,squeeze = 64 , expansion = 256)
out = Dropout(0.5)(out)
out = Conv2D(2,kernel_size=(1,1),padding='valid',activation='relu')(out)
out = GlobalAveragePooling2D()(out)
out = Dense(2, activation="softmax")(out)
model = Model(input_layer, out, name='squeezenet')
return model
def fire_module（input_layer，squeeze = 16，expansion = 32）：
fire_sq = Conv2D(squeeze,1,1,activation= 'relu')(input_layer)
fire_exp1 = Conv2D(expansion ,kernel_size=(1,1),activation='relu',padding='valid')(fire_sq)
fire_exp2 = Conv2D(expansion ,kernel_size=(3,3),activation='relu',padding='same')(fire_sq)
out = concatenate([fire_exp1,fire_exp2], axis=3)
return out
## training of the model
model_dw = build_squeezenet(input_shape= (50,50,1))
sgd = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0,
amsgrad=False)
model_dw.compile(
optimizer=sgd, loss='categorical_crossentropy',
)
model_dw.summary()
model_dw.fit(x=images,
y=labels,
batch_size=128,
epochs=10,
verbose=1,
validation_split=0.2,
validation_data=(val_images,val_labels),
shuffle=True)
Train on 4156 samples, validate on 1040 samples
Epoch 1/10
4156/4156 [==============================] - 17s 4ms/step - loss: 0.6661 -
acc: 0.6251 - val_loss: 0.9930 - val_acc: 0.0000e+00
Epoch 2/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6620 -
acc: 0.6251 - val_loss: 0.9439 - val_acc: 0.0000e+00
Epoch 3/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6597 -
acc: 0.6215 - val_loss: 0.9616 - val_acc: 0.0000e+00
Epoch 4/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6443 -
acc: 0.6251 - val_loss: 0.8945 - val_acc: 0.0000e+00
Epoch 5/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6336 -
acc: 0.6251 - val_loss: 0.9328 - val_acc: 0.0000e+00
Epoch 6/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6384 -
acc: 0.6251 - val_loss: 0.8628 - val_acc: 0.0000e+00
Epoch 7/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6619 -
acc: 0.6251 - val_loss: 0.9726 - val_acc: 0.0000e+00
Epoch 8/10
4156/4156 [==============================] - 14s 3ms/step - loss: 0.6608 -
acc: 0.6251 - val_loss: 0.8752 - val_acc: 0.0000e+00
Epoch 9/10
4156/4156 [==============================] - 15s 3ms/step - loss: 0.6627 -
acc: 0.6251 - val_loss: 1.0165 - val_acc: 0.0000e+00
Epoch 10/10
4156/4156 [==============================] - 15s 4ms/step - loss: 0.6622 -
acc: 0.6251 - val_loss: 1.0553 - val_acc: 0.0000e+00