python - 训练准确度提高但验证准确度保持在 0.5，并且模型为每个验证样本预测几乎相同的类别
问题描述
我正在使用 keras tensorflow 后端在时间流逝的 IVF 胚胎图像数据集上实现 ResNet50 + LSTM 和注意力模型。
数据集由大约 220 个样本组成，我使用 85% - 15% 的训练/验证拆分（203 个用于训练，27 个用于验证）。
我的模型能够达到 0.80+ 的训练精度，但验证精度要么停留在 0.5 左右，要么仅停留在 0.5，验证损失几乎是训练损失的两倍。
这只是一个过度拟合的问题吗？
如果没有，我该如何调试和提高验证集的性能？
#我尝试过的事情：
我尝试添加正则化（L1，0.01）/Dropout 层（0.5）/减少神经元（1024 到 512 到 256），但它们都不起作用）
我还通过减去平均值和除以标准差来标准化我的数据。
我正在使用具有 1e-5 学习率且没有权重衰减的 Adam 优化器。图像在训练前被打乱。
#下面是我的模型和进度条的代码
＃进度条：
Epoch 1/40
150/150 [==============================] - 28s 189ms/step - loss: 2.1318 - acc: 0.5267 - val_loss: 4.8806 - val_acc: 0.5556
Epoch 00001: val_loss improved from inf to 4.88055, saving model to result/resnetmodel.hdf5
Epoch 2/40
150/150 [==============================] - 14s 94ms/step - loss: 1.9957 - acc: 0.5867 - val_loss: 4.8210 - val_acc: 0.5000
Epoch 00002: val_loss improved from 4.88055 to 4.82100, saving model to result/resnetmodel.hdf5
Epoch 3/40
150/150 [==============================] - 14s 94ms/step - loss: 1.8062 - acc: 0.6200 - val_loss: 4.9689 - val_acc: 0.5000
Epoch 00003: val_loss did not improve from 4.82100
Epoch 4/40
150/150 [==============================] - 14s 91ms/step - loss: 1.7516 - acc: 0.6267 - val_loss: 5.0284 - val_acc: 0.5000
Epoch 00004: val_loss did not improve from 4.82100
Epoch 5/40
150/150 [==============================] - 14s 94ms/step - loss: 1.6508 - acc: 0.7000 - val_loss: 4.9873 - val_acc: 0.4444
Epoch 00005: val_loss did not improve from 4.82100
Epoch 6/40
150/150 [==============================] - 14s 92ms/step - loss: 1.5003 - acc: 0.7733 - val_loss: 4.9800 - val_acc: 0.4444
Epoch 00006: val_loss did not improve from 4.82100
Epoch 7/40
150/150 [==============================] - 14s 96ms/step - loss: 1.4614 - acc: 0.7667 - val_loss: 4.9435 - val_acc: 0.5000
Epoch 00007: val_loss did not improve from 4.82100
Epoch 8/40
150/150 [==============================] - 14s 90ms/step - loss: 1.5480 - acc: 0.6800 - val_loss: 4.9345 - val_acc: 0.5000
Epoch 00008: val_loss did not improve from 4.82100
Epoch 9/40
150/150 [==============================] - 14s 93ms/step - loss: 1.4334 - acc: 0.7667 - val_loss: 5.0452 - val_acc: 0.5000
Epoch 00009: val_loss did not improve from 4.82100
Epoch 10/40
150/150 [==============================] - 14s 94ms/step - loss: 1.4344 - acc: 0.7667 - val_loss: 5.1768 - val_acc: 0.4444
Epoch 00010: val_loss did not improve from 4.82100
Epoch 11/40
150/150 [==============================] - 15s 98ms/step - loss: 1.3369 - acc: 0.8533 - val_loss: 5.1331 - val_acc: 0.4444
Epoch 00011: val_loss did not improve from 4.82100
Epoch 12/40
150/150 [==============================] - 14s 93ms/step - loss: 1.2834 - acc: 0.8133 - val_loss: 5.1265 - val_acc: 0.4444
Epoch 00012: val_loss did not improve from 4.82100
Epoch 13/40
150/150 [==============================] - 14s 91ms/step - loss: 1.3007 - acc: 0.8200 - val_loss: 5.1941 - val_acc: 0.4444
Epoch 00013: val_loss did not improve from 4.82100
Epoch 14/40
150/150 [==============================] - 14s 94ms/step - loss: 1.2358 - acc: 0.8533 - val_loss: 5.3716 - val_acc: 0.4444
Epoch 00014: val_loss did not improve from 4.82100
Epoch 15/40
150/150 [==============================] - 14s 92ms/step - loss: 1.2823 - acc: 0.8000 - val_loss: 5.3877 - val_acc: 0.4444
纪元 00015：val_loss 没有从 4.82100 改进
纪元 00015：提前停止
评价：----加载数据
----prediction_on_eval-----
事实= [1。0.]，预测=[0.03809702 0.96190304]
事实= [1。0.]，预测=[0.9803326 0.0196674]
事实= [1。0.]，预测=[9.9986279e-01 1.3717638e-04]
事实= [1。0.]，预测=[0.98158103 0.01841903]
事实= [1。0.]，预测=[0.99492776 0.00507224]
事实= [1。0.]，预测=[0.70435154 0.29564843]
事实= [1。0.]，预测=[4.1277369e-04 9.9958724e-01]
事实= [1。0.]，预测=[0.9818978 0.01810225]
事实= [1。0.]，预测=[0.91195923 0.08804072]
事实= [0。1.]，预测=[0.986312 0.013688]
事实= [0。1.]，预测=[0.9985434 0.00145668]
事实= [0。1.]，预测=[0.80424094 0.195759]
事实= [0。1.]，预测=[0.9214819 0.07851809]
事实= [0。1.]，预测=[0.03754392 0.96245605]
事实= [0。1.]，预测=[9.9976009e-01 2.3989924e-04]
事实= [0。1.]，预测=[0.98681134 0.01318868]
事实= [0。1.]，预测=[0.9984666 0.0015334]
事实= [0。1.]，预测=[0.7229417 0.27705824]
#这是我的模型：
x =Input(shape = (40, config.img_shape, config.img_shape, config.img_channel))
if config.base_model == "inception_v3":
cnn = InceptionV3(weights = None, include_top=False, pooling = "avg")
elif config.base_model == 'ResNet50':
cnn = ResNet50(weights = None, include_top=False, pooling = "avg")
cnn.load_weights(config.pretrained_path)
for layer in cnn.layers:
layer.trainable = False
extracted_features = TimeDistributed(cnn)(x)
activations = Bidirectional(LSTM(config.num_units_lstm, return_sequences=True, recurrent_activation = 'relu', recurrent_initializer = 'glorot_uniform', name='Bidirectional_LSTM'))(extracted_features)
activations = Dropout(0.5)(activations)
attention = TimeDistributed(Dense(1, activation='tanh'),name = "context_vector")(activations)
attention = Flatten()(attention)
attention = Activation('softmax', name = "conext_weights")(attention)
attention = RepeatVector(config.num_units_lstm*2)(attention)
attention = Permute([2, 1])(attention)
sent_representation = merge.multiply([activations, attention])
sent_representation = Lambda(lambda xin: K.sum(xin, axis=1))(sent_representation)
sent_representation = BatchNormalization()(sent_representation)
prediction = Dense(config.num_classes, activation = 'softmax')(sent_representation)
model = Model(inputs = x, outputs = prediction)
解决方案
数据集由大约 220 个样本组成，我使用 85% - 15% 的训练/验证拆分（203 个用于训练，27 个用于验证）。
这只是一个过度拟合的问题吗？
听起来很有可能，是的。对于这样一个深度网络，220 个样本是非常非常小的数据集。真的不太可能从这么小的一组数据中学会很好地概括。
如果没有，我该如何调试和提高验证集的性能？
在理想情况下，再获取 100,000 个左右的样本并将它们添加到您的数据集中！
接受这一点可能是不切实际的，您可以尝试以下策略之一或组合：