[Retiarii] support hypermodule: autoactivation

[QuanluZhang]

hypermodule is a (PyTorch) module which contains many architecture/hyperparameter candidates for this module. By using hypermodule, NNI will automatically find the best architecture/hyperparameter for this module. This follows the design philosophy of Retiarii that users write DNN model as a space.

This pr is implementing AutoActivation from the paper "Search for activation Functions"

• AutoActivation with one core unit
• AutoActivation with two or more core units
• add document and docstring
• add test

[sayakpaul]

@QuanluZhang do you have a minimal example that shows how to use this one?

[QuanluZhang]

@QuanluZhang do you have a minimal example that shows how to use this one?

I have not added the example in this pr. you can refer to the following example:

import random

from nni.retiarii import model_wrapper
import nni.retiarii.nn.pytorch as nn
import nni.retiarii.strategy as strategy
import nni.retiarii.evaluator.pytorch.lightning as pl
import torch.nn.functional as F
from nni.retiarii import serialize
from nni.retiarii.experiment.pytorch import RetiariiExeConfig, RetiariiExperiment, debug_mutated_model
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision.datasets import MNIST
#import nni


@model_wrapper
class Net(nn.Module):
    def __init__(self, hidden_size):
        #super(Net, self).__init__()
        super().__init__()
        self.conv1 = nn.Conv2d(1, 20, 5, 1)
        self.conv2 = nn.Conv2d(20, 50, 5, 1)
        self.fc1 = nn.Linear(4*4*50, hidden_size)
        #self.fc1 = nn.LayerChoice([
        #    nn.Linear(4*4*50, hidden_size),
        #    nn.Linear(4*4*50, hidden_size, bias=False)
        #])
        #repeat = nn.Linear(hidden_size, hidden_size)
        #self.repeat = nn.Repeat(repeat, [1, 4])
        #self.act_cell = nn.Cell([nn.ReLU(), nn.Dropout(), nn.Linear(hidden_size, hidden_size)], 3, 3)
        #self.fc2 = nn.Linear(hidden_size*3, 10)
        self.fc2 = nn.Linear(hidden_size, 10)
        self.act = nn.AutoActivation()

    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.repeat(x)
        #x = self.act_cell([x])
        x = self.act(self.fc2(x))
        return F.log_softmax(x, dim=1)


if __name__ == '__main__':
    base_model = Net(128)
    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
    train_dataset = serialize(MNIST, root='data/mnist', train=True, download=True, transform=transform)
    test_dataset = serialize(MNIST, root='data/mnist', train=False, download=True, transform=transform)
    trainer = pl.Classification(train_dataloader=pl.DataLoader(train_dataset, batch_size=100),
                                val_dataloaders=pl.DataLoader(test_dataset, batch_size=100),
                                max_epochs=1, gpus=1)

    # uncomment the following two lines to debug a generated model
    #debug_mutated_model(base_model, trainer, [])
    #exit(0)

    #simple_strategy = strategy.PolicyBasedRL()
    simple_strategy = strategy.GridSearch()

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

    exp_config = RetiariiExeConfig('local')
    exp_config.experiment_name = 'mnist_search'
    exp_config.trial_concurrency = 3
    exp_config.max_trial_number = 250
    exp_config.trial_gpu_number = 1
    exp_config.training_service.use_active_gpu = False
    exp_config.execution_engine = 'py'

    exp.run(exp_config, 8081 + random.randint(0, 100))
    print('Final model:')
    for model_code in exp.export_top_models(formatter='dict'):
        print(model_code)

[sayakpaul]

Sure. Will give it a try. Would you folks prefer to have Colab Notebooks as they don't require any non-trivial setups and might be easier for anyone to get started?

[ultmaster]

Looks great!

[ultmaster]

Suggest globals()(unary)(), which is semantically more clear that eval.

[ultmaster]

I don't think they should have the same label?

[QuanluZhang]

yes, you are right, this is for easy testing, i will remove the label then.

[ultmaster]

You can write one in the base class and Python will automatically inherit it.