Fix models on python 3.11 (#1950)

Summary:
The CI is broken due to the recent default Python version upgrade.

Hereby we:
- Fix soft_actor_critic code change caused by upgrade of gym
- Upgrade omegaconf version due to Python and numpy version upgrade
- Upgrade detectron2 model version
- Move fambench_xlmr to canary_models

We are moving fambench_xlmr to canary_models because it depends on fairseq, and fairseq:
1. Depends on omegaconf < 2.1, which is not compatible with Python 3.11 (https://github.com/facebookresearch/fairseq/blob/main/setup.py#L183)
2. Detectron2, on the other hand, requires omegaconf >=2.1 < 2.4, so they can not co-exist. (https://github.com/facebookresearch/detectron2/blob/main/setup.py#L189)

We should suggest the owner of fairseq to update their code to adapt to omegaconf >=2.1.

Pull Request resolved: #1950

Reviewed By: davidberard98, msaroufim

Differential Revision: D49785274

Pulled By: xuzhao9

fbshipit-source-id: e46f56c28359f8a38e09e2b90d8e07d19b68d58a

torchbenchmark/models/soft_actor_critic/__init__.py

@@ -185,9 +185,9 @@ def __init__(self, test, device, batch_size=None, extra_args=[]):
 
     def get_module(self):
         model = self.agent.actor
-        state = self.train_env.reset()
+        state, _info = self.train_env.reset()
         action = self.agent.sample_action(state)
-        next_state, reward, done, info = self.train_env.step(action)
+        next_state, reward, done, info, _unused = self.train_env.step(action)
         self.buffer.push(state, action, reward, next_state, done)
         batch = self.buffer.sample(self.args.batch_size)
         state_batch, action_batch, reward_batch, next_state_batch, done_batch = batch
@@ -204,11 +204,11 @@ def train(self):
         niter = 1
         for step in range(niter):
             if done:
-                state = self.train_env.reset()
+                state, _info = self.train_env.reset()
                 steps_this_ep = 0
                 done = False
             action = self.agent.sample_action(state)
-            next_state, reward, done, info = self.train_env.step(action)
+            next_state, reward, done, info, _unused = self.train_env.step(action)
             self.buffer.push(state, action, reward, next_state, done)
             state = next_state
             steps_this_ep += 1
@@ -244,13 +244,13 @@ def eval(self) -> Tuple[torch.Tensor]:
             episode_return_history = []
             for episode in range(niter):
                 episode_return = 0.0
-                state = self.test_env.reset()
+                state, _info = self.test_env.reset()
                 done, info = False, {}
                 for step_num in range(self.args.max_episode_steps):
                     if done:
                         break
                     action = self.agent.forward(state)
-                    state, reward, done, info = self.test_env.step(action)
+                    state, reward, done, info, _unused = self.test_env.step(action)
                     episode_return += reward * (discount ** step_num)
                 episode_return_history.append(episode_return)
             retval = torch.tensor(episode_return_history)

torchbenchmark/util/framework/detectron2/requirements.txt

@@ -1,3 +1,3 @@
-git+https://github.com/facebookresearch/detectron2.git@57bdb21
-omegaconf==2.1.1
+git+https://github.com/facebookresearch/detectron2.git@1a4df4d
+omegaconf==2.3.0
 numpy