Chronos: Support TF2 in Chronos Forecaster

[TheaperDeng]

Import

Here we take LSTMForecaster as an example, same things will happen on LSTMForecaster, TCNForecaster, and Seq2SeqForecaster in the first round.

# which backend will be used by this will depend on a sys env CHRONOS_FORECASTER_BACKEND (default to 'torch')
# this is to support our legacy code
from bigdl.chronos.forecaster import LSTMForecaster

# These two will use the backend name stated in import
# tf2 is the same name as orca.learn
from bigdl.chronos.forecaster.pytorch import LSTMForecaster
from bigdl.chronos.forecaster.tf2 import LSTMForecater

Interface

I had a quick review of all APIs we have for pytorch and don't find any of them can not be supported in TF2. With the help of bigdl.nano.tf and bigdl.orca.learn.tf2. We should be able to support exactly the same API as pytorch.

class BaseTF2Forecaster(Forecaster):
    def __init__(..., optimizer="Adam",
                 loss="mse",
                 lr=0.001,
                 metrics=["mse"],
                 seed=None,
                 distributed=False,
                 workers_per_node=1,
                 distributed_backend="tf2"):

    no difference between pytorch
    distributed_backend will support “horovod”, “tf2” or “spark”
    ```
def fit(self, data, epochs=1, batch_size=32):
    ```
    no difference between pytorch 
    for data we will support
    1. a numpy ndarray tuple (x, y)
    2. a xshard item
    3. a tf.data.Dataset
    ```
def predict(self, data, batch_size=32, quantize=False):
    pass
def predict_with_onnx(self, data, batch_size=32):
    pass
def evaluate(self, data, batch_size=32, multioutput="raw_values", quantize=False):
    pass
def evaluate_with_onnx(self, data, batch_size=32, multioutput="raw_values"):
    pass
def save(self, checkpoint_file, quantize_checkpoint_file=None):
    pass
def load(self, checkpoint_file, quantize_checkpoint_file=None):
    pass
def to_local(self):
    pass
def get_model(self):
    pass
def build_onnx(self, thread_num=None, sess_options=None):
    pass
def export_onnx_file(self, dirname="model.onnx"):
    pass
def quantize(self, calib_data,
             val_data=None,
             metric=None,
             conf=None,
             framework='tensorflow',
             approach='static',
             tuning_strategy='bayesian',
             relative_drop=None,
             absolute_drop=None,
             timeout=0,
             max_trials=1):
    pass

[shane-huang]

"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.

[TheaperDeng]

"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.

Yeah, I have no strong opinion on tf or tf2. I choose tf2 just to align with orca. I think we can use tf for now.

[shane-huang]

• We need to ensure there won't be errors if user have only pytorch or tensorflow installed. - @jason-dai , do you still think a global flag is need to enable only pytorch or tensorflow models? maybe we don't need that anymore.
• We need install dependencies chronos[pytorch] chronos[tensorflow]. Together with lite and all, we may need pytorch-lite, tf-lite, pytorch-all, tf-all, etc.

[shane-huang]

"bigdl.chronos.forecaster.tf2": should we use "tf2" or just "tf" as nano ? For now we don't intend to support tf1. And even if we support tf1 in future we can make tf using tf 2.0 as default and use tf1 namespace for back compatibility.

Yeah, I have no strong opinion on tf or tf2. I choose tf2 just to align with orca. I think we can use tf for now.

@jason-dai what do you say?

[TheaperDeng]

• We need to ensure there won't be errors if user have only pytorch or tensorflow installed. - @jason-dai , do you still think a global flag is need to enable only pytorch or tensorflow models? maybe we don't need that anymore.
• We need install dependencies chronos[pytorch] chronos[tensorflow]. Together with lite and all, we may need pytorch-lite, tf-lite, pytorch-all, tf-all, etc.

I think we may have these installation option: For bigdl-chronos:

• default: bigdl-orca, bigdl-nano[pytorch]
• [all]: bigdl-orca[automl], bigdl-nano[pytorch, tensorflow]
• [pytorch]: bigdl-orca[automl], bigdl-nano[pytorch]
• [tensorflow]: bigdl-orca[automl], bigdl-nano[tensorflow]

For bigdl-chronos-lite:

• default: bigdl-nano
• [all]: bigdl-nano[pytorch, tensorflow]
• [pytorch]: bigdl-nano[pytorch]
• [tensorflow]: bigdl-nano[tensorflow]