marsaki/nasnet

reimplementation of "Learning Transferable Architectures for Scalable Image Recognition" using mnist dataset, include controller

NasNet 2018

This code is the reimplementation of "Learning Transferable Architectures for Scalable Image Recognition", including the training process of controller. This code contains three algorithms to search model, Random Search, Policy Gradient and PPO.

Requirements

Python >= 3.6.7, PyTorch == 0.4.0

Architecture Search

python train_search.py --cutout --algorithm RS  #use random search
python train_search.py --cutout --algorithm PG  #use policy gradient
python train_search.py --cutout --algorithm PPO #use PPO

Note the validation performance in this step does not indicate the final performance of the architecture. One must train the obtained genotype/architecture from scratch using full-sized models. Also the default setting is training with 20 processes and 3 GPU. Change the processes to 10:

python train_search.py --cutout --episodes 10

or modify code in random_search.py, policy_gradient.py and PPO.py .

Architecture Evaluation

Because of the limitation of time and computation resource, I didn't train the candidate genotypes/architectures from scratch.

Results

python draw.py

We can see RL search is better than random search, also PPO is more stable and faster.

nasnet's Issues

What are node decoders?

policy_gradient.py's problems

hello.
I have two questions in policy_gradient.py to ask you.
first:
def cal_loss(self, actions_p, actions_log_p, worker, baseline):
reward = worker.acc - baseline
policy_loss = -1 * torch.sum(actions_log_p * reward)
entropy = -1 * torch.sum(actions_p * actions_log_p)
entropy_bonus = -1 * entropy * self.entropy_weight
what does entory mean?
second:
for episode in range(self.episodes):
worker = results_queue.get()
# worker.actions_p = torch.Tensor(worker.actions_p).to(self.device)
worker.actions_index = torch.LongTensor(worker.actions_index).to(self.device)
workers.append(worker)

worker.actions_index = torch.LongTensor(worker.actions_index).to(self.device) ,the worker.actions_index does't change?

What the step means in controller

What does the step mean in the controller?

for node in range(self.steps):

I think it's a number of blocks in each cell. In the original paper, there are 5 blocks in either normal cell or reduction cell.

why it is 4?