预测结果为全灰图

[simplew2011]

• 默认数据集，训练5轮
• 预测发现全部是灰图，没有结果

看loss比较大，loss: 0.5252 - accuracy: 0.7814，就是没有模型收敛。
修改loss，lr，Adam(lr = 1e-5)，steps_per_epoch=2000, epochs=20，batch_size=32。
loss缓慢下降，正常收敛，预测正常
关键是batch_size和lr需要微调，多调整几次看loss有没有下降趋势，没有的话中断训练，继续修改

2022-09-13 05:10:38.208563: I tensorflow/stream_executor/cuda/cuda_dnn.cc:384] Loaded cuDNN version 8201
2000/2000 [==============================] - ETA: 0s - loss: 0.2607 - accuracy: 0.8837
Epoch 1: loss improved from inf to 0.26072, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 858s 424ms/step - loss: 0.2607 - accuracy: 0.8837
Epoch 2/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1718 - accuracy: 0.9249
Epoch 2: loss improved from 0.26072 to 0.17182, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 849s 425ms/step - loss: 0.1718 - accuracy: 0.9249
Epoch 3/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1499 - accuracy: 0.9342
Epoch 3: loss improved from 0.17182 to 0.14994, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 850s 425ms/step - loss: 0.1499 - accuracy: 0.9342
Epoch 4/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1355 - accuracy: 0.9406
Epoch 4: loss improved from 0.14994 to 0.13549, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 851s 426ms/step - loss: 0.1355 - accuracy: 0.9406
Epoch 5/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1243 - accuracy: 0.9455
Epoch 5: loss improved from 0.13549 to 0.12432, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 841s 421ms/step - loss: 0.1243 - accuracy: 0.9455
Epoch 6/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1154 - accuracy: 0.9495
Epoch 6: loss improved from 0.12432 to 0.11543, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 841s 420ms/step - loss: 0.1154 - accuracy: 0.9495
Epoch 7/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1082 - accuracy: 0.9527
Epoch 7: loss improved from 0.11543 to 0.10821, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 839s 419ms/step - loss: 0.1082 - accuracy: 0.9527
Epoch 8/20
2000/2000 [==============================] - ETA: 0s - loss: 0.1022 - accuracy: 0.9554
Epoch 8: loss improved from 0.10821 to 0.10224, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 840s 420ms/step - loss: 0.1022 - accuracy: 0.9554
Epoch 9/20
2000/2000 [==============================] - ETA: 0s - loss: 0.0972 - accuracy: 0.9576
Epoch 9: loss improved from 0.10224 to 0.09725, saving model to unet_membrane.hdf5
2000/2000 [==============================] - 843s 422ms/step - loss: 0.0972 - accuracy: 0.9576

[anson0626150]

so do I

[Sisyphka]

really appreciate, you fix the problem that always perplex me, using the measures, Unet work

Epoch 4/25
500/500 [==============================] - ETA: 0s - loss: 0.0701 - accuracy: 0.9694
Epoch 4: loss improved from 0.08390 to 0.07009, saving model to unet_membrane.hdf5