add word2vec test for the new API

[helinwang]

No description provided.

[sidgoyal78]

Can we maybe have something like fluid.params.load() (since we have trainer.params.save() above) ?

[helinwang]

Thanks for thinking about it! If we aim for symmetry, params.save is a method of a params instance, in the same time fluid.params.load is a static method, which is not very symmetric?

[cs2be]

Sorry didn't catch this before, but place is not used, it should be passed into fluid.Inferencer

[helinwang]

Thanks, done.

[reyoung]

The metrics, like error rate, auc, are not defined in this demo. How do we define a error rate in our new designed API?

By just adding fluid.layers.accury(input=predict_word, label=label) ? Should we fetch metrics by default?

[reyoung]

If we passing the optimizer instance to a trainer, we cannot use learning rate schedule since the learning_rate is a variable of learning rate schedule method, which is designed by @jacquesqiao

Perhaps, we can pass a lambda here or an instance here.

lambda could be

trainer = fluid.Trainer(optimizer=lambda: fluid.optimizer.SGD(learning_rate=fluid.layers.fill_constant(1e-3)))

[helinwang]

Cool! I realized that we may need to freeze parameter when reading your GAN example.
Another possibility is we pass in a optimize function?

def optimize(loss):
    opt = fluid.optimizer.SGD(learning_rate=fluid.layers.fill_constant(1e-3)))
    return opt.minimize(
        loss=loss,
        parameter_list=[
            p.name for p in g_program.global_block().all_parameters()
        ])

trainer = fluid.Trainer(..., optimize=optimize)

EDIT: never mind, I just realized that the user would not have access to p.name for p in g_program.global_block().all_parameters().

[reyoung]

partial(train_network, is_sparse) ?

[helinwang]

Yes, fluid.Trainer's first argument should be a function without any arguments. partial(train_network, is_sparse) binds is_sparse to train_network, creating a function without any argument.