loss, val_loss, acc and val_acc do not update at a

问题:

I created an LSTM network for sequence classification (binary) where each sample has 25 timesteps and 4 features. The following is my keras network topology:

Above, the activation layer after Dense layer uses softmax function. I used binary_crossentropy for loss function and Adam as the optimizer to compile the keras model. Trained the model with batch_size=256, shuffle=True and validation_split=0.05, The following is the training log:

Train on 618196 samples, validate on 32537 samples
2017-09-15 01:23:34.407434: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:893] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2017-09-15 01:23:34.407719: I tensorflow/core/common_runtime/gpu/gpu_device.cc:955] Found device 0 with properties: 
name: GeForce GTX 1050
major: 6 minor: 1 memoryClockRate (GHz) 1.493
pciBusID 0000:01:00.0
Total memory: 3.95GiB
Free memory: 3.47GiB
2017-09-15 01:23:34.407735: I tensorflow/core/common_runtime/gpu/gpu_device.cc:976] DMA: 0 
2017-09-15 01:23:34.407757: I tensorflow/core/common_runtime/gpu/gpu_device.cc:986] 0:   Y 
2017-09-15 01:23:34.407764: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1045] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GeForce GTX 1050, pci bus id: 0000:01:00.0)
618196/618196 [==============================] - 139s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 2/50
618196/618196 [==============================] - 132s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 3/50
618196/618196 [==============================] - 134s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 4/50
618196/618196 [==============================] - 133s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 5/50
618196/618196 [==============================] - 132s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 6/50
618196/618196 [==============================] - 132s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 7/50
618196/618196 [==============================] - 132s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251
Epoch 8/50
618196/618196 [==============================] - 132s - loss: 4.3489 - acc: 0.7302 - val_loss: 4.4316 - val_acc: 0.7251

... and so on through 50 epochs with same numbers

So far, I have also tried using rmsprop, nadam optimizers and batch_size(s) 128, 512, 1024 but the loss, val_loss, acc, val_acc always remained same throughout all epochs, yielding accuracy in the range of 0.72 to 0.74 in my each attempt.

回答1:

The softmax activation makes sure the sum of the outputs is 1. It's useful for assuring that only one class among many classes will be output.

Since you have only 1 output (only one class), it's certainly a bad idea. You're probably ending up with 1 as result for all samples.

Use sigmoid instead. It goes well with binary_crossentropy.