dongminlee94 / meta-learning-for-everyone Goto Github PK

• Best checkpoint
• Final checkpoint

The bug on RL^2's buffer needs to be fixed.

Describe the bug

• In general there is an epsilon in normalization equation practically avoid zero dividing error.

Expected behavior

• https://github.com/dongminlee94/meta-rl/blob/86f61bb1ce9b811456da24db8bd5dbb114346223/src/maml/algorithm/ppo.py#L91
• need to consider whether use the epsilon

아나콘다 프롬프트로 패키지 설치까지는 완료하였는데, 이후 jupyter notebook에서 입력해보니 torchmeta를 import하지 못하는 오류가 발생하고 있습니다.
혹시 해결방법이 있을까요?

Is your feature request related to a problem? Please describe.

A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like

A clear and concise description of what you want to happen.

Additional context

Add any other context or screenshots about the feature request here.

In the original MAML paper, the author implements a value function (vf) as a linear function and fits it with batch data every time.
Learn2Learn repo implement vf exactly same with the paper.
However, in Ray Project repo, vf is implemented as a 2 layered neural network and updated simultaneously with a policy network. In this case, vf is shared through tasks and iterations.

we need to determine how to implement this vf.

Is your feature request related to a problem? Please describe.

• In cheetah-dir environment, self.tasks uses dictionary unnecessarily.

directions = [-1, 1, -1, 1]
self.tasks = [{"direction": direction} for direction in directions]

https://github.com/dongminlee94/meta-rl/blob/develop/src/envs/half_cheetah_dir.py#L19

• self._goal is not used in both cheetah-dir and cheetah-vel environments

def __init__(self, num_tasks=2, seed=0):
        super().__init__(render=False)
        self.tasks = self.sample_tasks(num_tasks)
        self._goal_vel = self.tasks[0].get("velocity", 0.0)
        self._goal = self._goal_vel
        self._task = None
        self._alive = None
        self.rewards = None
        self.potential = None
        self.seed(seed)

'''
'''
'''

def reset_task(self, index):
        """Reset velocity target to index of task"""
        self._task = self.tasks[index]
        self._goal_vel = self._task["velocity"]
        self._goal = self._goal_vel
        self.reset()

Describe the solution you'd like

• Change self.tasks from dict to list
• Remove self._goal attribute

Additional context

Add any other context or screenshots about the feature request here.

• train_tasks: the total number of tasks to train
• num_sample_tasks: the number of sampled tasks for each iteration

This issue is to gather the common modules (e.g., sampler, buffer). Because the repo's codes currently have too many duplication codes.

현재 Colab에서 해당 코드 실행시 torchmeta가 import가 안되는 문제가 발생합니다.

파이썬 버전 문제인가 싶어서 python3.7로 실행했는데,
패키지 설치까진되고, 따로 import가 안됩니다.

혹시 해결하신분 계신가요?

Installing collected packages: urllib3, typing-extensions, tqdm, Pillow, ordered-set, numpy, idna, charset-normalizer, certifi, torch, requests, h5py, torchvision, torchmeta
Successfully installed Pillow-9.5.0 certifi-2022.12.7 charset-normalizer-3.1.0 h5py-3.8.0 idna-3.4 numpy-1.21.6 ordered-set-4.1.0 requests-2.30.0 torch-1.9.1 torchmeta-1.8.0 torchvision-0.10.1 tqdm-4.65.0 typing-extensions-4.5.0 urllib3-2.0.2
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv

Describe the bug

argv[0]= is occurred at the beginning of the training.

To Reproduce

Steps to reproduce the behavior:

1. run python {algorighm}_trainer.py
2. then argv[0]= is occurred on the terminal as followed
   

Expected behavior

A clear and concise description of what you expected to happen.

• It may occurred from somewhere of the PyBullet environment.

Additional context

Add any other context about the problem here.

Is your feature request related to a problem? Please describe.

All linters' configurations can be integrated into setup.cfg.

Describe the solution you'd like

https://github.com/annotation-ai/python-project-template/blob/main/setup.cfg
You can see the integrated configurations for black, isort, mypy, flake8, pylint, pytest.

This issue is to add type annotation to the repo's codes

This issue is an issue that must adds the list below.

• Early stopping condition
• Models, buffers, and (best, final) checkpoints saving
• Models, buffers, and (best, final) checkpoints loading
• Train tasks and test tasks distribution (optional)

The issue is related to #22

Currently get_action is misplaced in sampler.py.
To remove it, it is needed to check whether there are any problems in putting agent into the sampler.py.

recommended code: self.agent.policy = inner_policy

need to determine how to implement meta-test process of MAML algorithm.

• whether to iteratively perform second-order optimization
• how about in other MAML repos?

• variable name (num, cur, max)
• sampler.py
• buffer.py
• add PATH to the make file
• change directory (src, env)

Change the code from

def collect_train_data(self, task_index, max_samples, update_posterior, add_to_enc_buffer):
        """Data collecting for meta-train"""
        self.agent.encoder.clear_z()
        self.agent.policy.is_deterministic = False

        cur_samples = 0
        while cur_samples < max_samples:
            trajs, num_samples = self.sampler.obtain_samples(
                max_samples=max_samples - cur_samples,
                update_posterior=update_posterior,
                accum_context=False,
            )
            cur_samples += num_samples

            self.rl_replay_buffer.add_trajs(task_index, trajs)
            if add_to_enc_buffer:
                self.encoder_replay_buffer.add_trajs(task_index, trajs)

            if update_posterior:
                context_batch = self.sample_context([task_index])
                self.agent.encoder.infer_posterior(context_batch)

def obtain_samples(self, max_samples, update_posterior, accum_context=True):
        """Obtain samples up to the number of maximum samples"""
        trajs = []
        cur_samples = 0

        while cur_samples < max_samples:
            traj = self.rollout(accum_context=accum_context)
            trajs.append(traj)
            cur_samples += len(traj["cur_obs"])
            self.agent.encoder.sample_z()

            if update_posterior:
                break
        return trajs, cur_samples

to

def collect_train_data(self, task_index, max_samples, update_posterior, add_to_enc_buffer):
        """Data collecting for meta-train"""
        self.agent.encoder.clear_z()
        self.agent.policy.is_deterministic = False

        trajs, num_samples = self.sampler.obtain_samples(
            max_samples=max_samples,
            accum_context=False,
        )

        self.rl_replay_buffer.add_trajs(task_index, trajs)
        if add_to_enc_buffer:
            self.encoder_replay_buffer.add_trajs(task_index, trajs)

        if update_posterior:
            context_batch = self.sample_context([task_index])
            self.agent.encoder.infer_posterior(context_batch)

def obtain_samples(self, max_samples, accum_context=True):
        """Obtain samples up to the number of maximum samples"""
        trajs = []
        cur_samples = 0

        while cur_samples < max_samples:
            traj = self.rollout(accum_context=accum_context)
            trajs.append(traj)
            cur_samples += len(traj["cur_obs"])
            self.agent.encoder.sample_z()
        return trajs, cur_samples

Then, experiment PEARL

Describe the bug

Killed error is occurred on laptop notebooks to be the cause of out of memory

To Reproduce

Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

Expected behavior

A clear and concise description of what you expected to happen.

Additional context

Add any other context about the problem here.

Change from python config files to yaml config files

Hi, friends. I happen to find this great repo! It seems much elegant than most other meta-rl implementations. I wonder if the Feature branch is tested on environments like cheetah and can be readily used for other environments. Thanks!

torch.lstsq() is deprecated in favor of torch.linalg.lstsq() in Pytorch version >= 1.9.0.

for more stable codes we might modify the function for users with higher version of Pytorch as followed.

if hasattr(torch, 'lstsq'):  # Required for torch < 1.9.0
    coeffs = torch.lstsq(b, A).solution
else:
    coeffs = torch.linalg(A, b).solution

Describe the bug

README shows package information not aligned.

• no package information about PyTorch in both requirements.txt and setup.py.
• no mypy setting

In MAML we uses "Higher" module which is a library providing support for higher-order optimization developed by Facebook.
https://github.com/facebookresearch/higher
It turns existing torch.nn.Module instances "stateless", meaning that changes to the parameters thereof can be tracked.

Therefore, it is needed to be checked whether the implemented Higher module really tracks the parameters of policy through the outer loop and inner loop.

@all-contributors please add @dongminlee94 for infrastructure, tests and code