feature(xyy):add HPT model and test_hpt

ding/model/template/hpt.py

@@ -97,7 +97,8 @@ def init_cross_attn(self):
         """Initialize cross-attention module and learnable tokens."""
         token_num = 16
         self.tokens = nn.Parameter(torch.randn(1, token_num, 128) * INIT_CONST)
-        self.cross_attention = CrossAttention(128, heads=8, dim_head=64, dropout=0.1)
+        self.cross_attention = CrossAttention(
+            128, heads=8, dim_head=64, dropout=0.1)
 
     def compute_latent(self, x: torch.Tensor) -> torch.Tensor:
         """
@@ -112,10 +113,12 @@ def compute_latent(self, x: torch.Tensor) -> torch.Tensor:
         """
         # Using the Feature Extractor
         stem_feat = self.feature_extractor(x)
-        stem_feat = stem_feat.reshape(stem_feat.shape[0], -1, stem_feat.shape[-1])  # (B, N, 128)
+        stem_feat = stem_feat.reshape(
+            stem_feat.shape[0], -1, stem_feat.shape[-1])  # (B, N, 128)
         # Calculating latent tokens using CrossAttention
         stem_tokens = self.tokens.repeat(len(stem_feat), 1, 1)  # (B, 16, 128)
-        stem_tokens = self.cross_attention(stem_tokens, stem_feat)  # (B, 16, 128)
+        stem_tokens = self.cross_attention(
+            stem_tokens, stem_feat)  # (B, 16, 128)
         return stem_tokens
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -195,7 +198,8 @@ def forward(self, x: torch.Tensor, context: torch.Tensor, mask: Optional[torch.T
         h = self.heads
         q = self.to_q(x)
         k, v = self.to_kv(context).chunk(2, dim=-1)
-        q, k, v = map(lambda t: rearrange(t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
+        q, k, v = map(lambda t: rearrange(
+            t, "b n (h d) -> (b h) n d", h=h), (q, k, v))
         sim = torch.einsum("b i d, b j d -> b i j", q, k) * self.scale
 
         if mask is not None:

ding/model/template/tests/test_hpt.py

@@ -4,9 +4,9 @@
 from ding.model.template.hpt import HPT
 from ding.torch_utils import is_differentiable
 
-T, B = 3, 4
-obs_shape = [4, (8, ), (4, 64, 64)]  # Example observation shapes
-act_shape = [3, (6, ), [2, 3, 6]]  # Example action shapes
+T, B = 4
+obs_shape = [4, (8, )]
+act_shape = [3, (6, )]
 args = list(product(*[obs_shape, act_shape]))
 
 

dizoo/box2d/lunarlander/entry/lunarlander_dqn_example.py

@@ -19,16 +19,19 @@
 
 def main():
     logging.getLogger().setLevel(logging.INFO)
-    cfg = compile_config(main_config, create_cfg=create_config, auto=True, save_cfg=task.router.node_id == 0)
+    cfg = compile_config(main_config, create_cfg=create_config,
+                         auto=True, save_cfg=task.router.node_id == 0)
     ding_init(cfg)
 
     with task.start(async_mode=False, ctx=OnlineRLContext()):
         collector_env = SubprocessEnvManagerV2(
-            env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.collector_env_num)],
+            env_fn=[lambda: DingEnvWrapper(
+                gym.make("LunarLander-v2")) for _ in range(cfg.env.collector_env_num)],
             cfg=cfg.env.manager
         )
         evaluator_env = SubprocessEnvManagerV2(
-            env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.evaluator_env_num)],
+            env_fn=[lambda: DingEnvWrapper(
+                gym.make("LunarLander-v2")) for _ in range(cfg.env.evaluator_env_num)],
             cfg=cfg.env.manager
         )
 
@@ -38,7 +41,8 @@ def main():
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         model = DQN(**cfg.policy.model).to(device)
 
-        buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size)
+        buffer_ = DequeBuffer(
+            size=cfg.policy.other.replay_buffer.replay_buffer_size)
 
         # Pass the model into Policy
         policy = DQNPolicy(cfg.policy, model=model)

dizoo/box2d/lunarlander/entry/lunarlander_hpt_example.py

@@ -21,16 +21,19 @@
 
 def main():
     logging.getLogger().setLevel(logging.INFO)
-    cfg = compile_config(main_config, create_cfg=create_config, auto=True, save_cfg=task.router.node_id == 0)
+    cfg = compile_config(main_config, create_cfg=create_config,
+                         auto=True, save_cfg=task.router.node_id == 0)
     ding_init(cfg)
 
     with task.start(async_mode=False, ctx=OnlineRLContext()):
         collector_env = SubprocessEnvManagerV2(
-            env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.collector_env_num)],
+            env_fn=[lambda: DingEnvWrapper(
+                gym.make("LunarLander-v2")) for _ in range(cfg.env.collector_env_num)],
             cfg=cfg.env.manager
         )
         evaluator_env = SubprocessEnvManagerV2(
-            env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.evaluator_env_num)],
+            env_fn=[lambda: DingEnvWrapper(
+                gym.make("LunarLander-v2")) for _ in range(cfg.env.evaluator_env_num)],
             cfg=cfg.env.manager
         )
 
@@ -39,8 +42,10 @@ def main():
         # Migrating models to the GPU
         device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         # HPT introduces a Policy Stem module, which processes the input features using Cross-Attention.
-        model = HPT(cfg.policy.model.obs_shape, cfg.policy.model.action_shape).to(device)
-        buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size)
+        model = HPT(cfg.policy.model.obs_shape,
+                    cfg.policy.model.action_shape).to(device)
+        buffer_ = DequeBuffer(
+            size=cfg.policy.other.replay_buffer.replay_buffer_size)
 
         # Pass the model into Policy
         policy = DQNPolicy(cfg.policy, model=model)