Keras train_on_batch loss / accuracy 0

Question

I use a large dataset, so I'm trying to use train_on_batch (or match with epoch = 1)

model = Sequential()
model.add(LSTM(size,input_shape=input_shape,return_sequences=False))
model.add(Dense(output_dim))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=["accuracy"])

for e in range(nb_epoch):
    for batch_X, batch_y in batches:
        model.train_on_batch(batch_X,batch_y)
        # or
        # model.fit(batch_X,batch_y,batch_size=batch_size,nb_epoch=1,verbose=1,shuffle=True,)

But when training begins, it happens:

(0, 128)
Epoch 1/1
128/128 [==============================] - 2s - loss: 0.3262 - acc: 0.1130

(129, 257)
Epoch 1/1
128/128 [==============================] - 2s - loss: -0.0000e+00 - acc: 0.0000e+00

No matter how many eras I wait, this does not change. Even if I changed the batch size, the same thing will happen: the first batch has good values, and then it just goes into "loss: -0.0000e + 00 - acc: 0.0000e + 00".

Can someone help to understand what is happening here?

Thanks in advance