Add TupleConstruct to allow (a, b)=SomeModule()

[tczhangzhi]

NNI did not support the prim::TupleConstruct operation, causing the following code to report an error when outputting tuple data in forward:

import random

import nni.retiarii.nn.pytorch as nn
import torch.nn.functional as F
from nni.retiarii.experiment import RetiariiExeConfig, RetiariiExperiment
from nni.retiarii.strategies import RandomStrategy
from nni.retiarii.trainer import PyTorchImageClassificationTrainer


class Net(nn.Module):
    def __init__(self, hidden_size):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(1, 20, 5, 1)
        self.conv2 = nn.Conv2d(20, 50, 5, 1)
        self.fc1 = nn.LayerChoice([
            nn.Linear(4*4*50, hidden_size),
            nn.Linear(4*4*50, hidden_size, bias=False)
        ])
        self.fc2 = nn.Linear(hidden_size, 10)

    def forward(self, x):
        x = F.relu(self.conv1(x))
        x = F.max_pool2d(x, 2, 2)
        x = F.relu(self.conv2(x))
        x = F.max_pool2d(x, 2, 2)
        x = x.view(-1, 4*4*50)
        x = F.relu(self.fc1(x))
        x = self.fc2(x)
        =========== tuple output start ===========
        return F.log_softmax(x, dim=1), x
        =========== tuple output end ===========


if __name__ == '__main__':
    base_model = Net(128)
    # some customized trainer
    trainer = PyTorchImageClassificationTrainer(base_model, dataset_cls="MNIST",
                                                dataset_kwargs={"root": "data/mnist", "download": True},
                                                dataloader_kwargs={"batch_size": 32},
                                                optimizer_kwargs={"lr": 1e-3},
                                                trainer_kwargs={"max_epochs": 1})

    simple_startegy = RandomStrategy()

    exp = RetiariiExperiment(base_model, trainer, [], simple_startegy)

    exp_config = RetiariiExeConfig('local')
    exp_config.experiment_name = 'mnist_search'
    exp_config.trial_concurrency = 2
    exp_config.max_trial_number = 10
    exp_config.training_service.use_active_gpu = False

    exp.run(exp_config, 7081 + random.randint(0, 100))

The error message is as follows:

Traceback (most recent call last):
  File "test.py", line 50, in <module>
    exp.run(exp_config, 7081 + random.randint(0, 100))
  File "/data/zhangzhi/nni/nni/retiarii/experiment.py", line 173, in run
    super().run(port, debug)
  File "/data/zhangzhi/nni/nni/experiment/experiment.py", line 181, in run
    self.start(port, debug)
  File "/data/zhangzhi/nni/nni/retiarii/experiment.py", line 158, in start
    self._start_strategy()
  File "/data/zhangzhi/nni/nni/retiarii/experiment.py", line 124, in _start_strategy
    base_model_ir = convert_to_graph(script_module, self.base_model)
  File "/data/zhangzhi/nni/nni/retiarii/converter/graph_gen.py", line 522, in convert_to_graph
    GraphConverter().convert_module(script_module, module, module_name, model)
  File "/data/zhangzhi/nni/nni/retiarii/converter/graph_gen.py", line 483, in convert_module
    module_name, ir_model, ir_graph)
  File "/data/zhangzhi/nni/nni/retiarii/converter/graph_gen.py", line 349, in handle_graph_nodes
    handle_single_node(node)
  File "/data/zhangzhi/nni/nni/retiarii/converter/graph_gen.py", line 344, in handle_single_node
    raise RuntimeError('Unsupported kind: {}'.format(node.kind()))
RuntimeError: Unsupported kind: prim::TupleConstruct

After PR, the generated operation is as follows:

__TupleConstruct32 = (__log_softmax31, __fc2)
return __TupleConstruct32

This PR is partly related to #2756 (requiring the support for TupleUnpack and TupleConstruct) and #3340 (working on TupleUnpack).

[tczhangzhi]

BTW, I found this bug during adding regularization terms (of loss functions) depending on some intermediate variables. I also fixed other problems but they are not related to this issue, so I will submit them in another PRs.