Add transformer model to diffusion policy

lerobot/common/policies/diffusion/configuration_diffusion.py

@@ -98,7 +98,7 @@ class DiffusionConfig:
 
     # Inputs / output structure.
     n_obs_steps: int = 2
-    horizon: int = 16
+    horizon: int = 10
     n_action_steps: int = 8
 
     input_shapes: dict[str, list[int]] = field(
@@ -134,7 +134,7 @@ class DiffusionConfig:
     down_dims: tuple[int, ...] = (512, 1024, 2048)
     kernel_size: int = 5
     n_groups: int = 8
-    diffusion_step_embed_dim: int = 128
+    diffusion_step_embed_dim: int = 256
     use_film_scale_modulation: bool = True
     # Noise scheduler.
     noise_scheduler_type: str = "DDPM"
@@ -145,6 +145,14 @@ class DiffusionConfig:
     prediction_type: str = "epsilon"
     clip_sample: bool = True
     clip_sample_range: float = 1.0
+    # Transformer
+    use_transformer: bool = True
+    n_layer: int = 8
+    n_head: int = 4
+    p_drop_emb: float = 0.0
+    p_drop_attn: float = 0.3
+    causal_attn: bool = True
+    n_cond_layers: int = 0
 
     # Inference
     num_inference_steps: int | None = None
@@ -200,7 +208,7 @@ def __post_init__(self):
         # Check that the horizon size and U-Net downsampling is compatible.
         # U-Net downsamples by 2 with each stage.
         downsampling_factor = 2 ** len(self.down_dims)
-        if self.horizon % downsampling_factor != 0:
+        if not self.use_transformer and self.horizon % downsampling_factor != 0:
             raise ValueError(
                 "The horizon should be an integer multiple of the downsampling factor (which is determined "
                 f"by `len(down_dims)`). Got {self.horizon=} and {self.down_dims=}"

lerobot/common/policies/diffusion/modeling_diffusion.py

@@ -22,7 +22,7 @@
 
 import math
 from collections import deque
-from typing import Callable
+from typing import Callable, Tuple
 
 import einops
 import numpy as np
@@ -188,7 +188,12 @@ def __init__(self, config: DiffusionConfig):
             self._use_env_state = True
             global_cond_dim += config.input_shapes["observation.environment_state"][0]
 
-        self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
+        if config.use_transformer:
+            self.net = TransformerForDiffusion(config, cond_dim=global_cond_dim)
+        else:
+            self.net = DiffusionConditionalUnet1d(
+                config, global_cond_dim=global_cond_dim * config.n_obs_steps
+            )
 
         self.noise_scheduler = _make_noise_scheduler(
             config.noise_scheduler_type,
@@ -206,6 +211,20 @@ def __init__(self, config: DiffusionConfig):
         else:
             self.num_inference_steps = config.num_inference_steps
 
+    def get_optimizer(
+        self,
+        transformer_weight_decay: float = 1e-3,
+        rgb_encoder_weight_decay: float = 1e-6,
+        learning_rate: float = 1e-4,
+        betas: Tuple[float, float] = [0.9, 0.95],
+    ) -> torch.optim.Optimizer:
+        optim_groups = self.net.get_optim_groups(weight_decay=transformer_weight_decay)
+        optim_groups.append(
+            {"params": self.rgb_encoder.parameters(), "weight_decay": rgb_encoder_weight_decay}
+        )
+        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas)
+        return optimizer
+
     # ========= inference  ============
     def conditional_sample(
         self, batch_size: int, global_cond: Tensor | None = None, generator: torch.Generator | None = None
@@ -225,7 +244,7 @@ def conditional_sample(
 
         for t in self.noise_scheduler.timesteps:
             # Predict model output.
-            model_output = self.unet(
+            model_output = self.net(
                 sample,
                 torch.full(sample.shape[:1], t, dtype=torch.long, device=sample.device),
                 global_cond=global_cond,
@@ -324,7 +343,7 @@ def compute_loss(self, batch: dict[str, Tensor]) -> Tensor:
         noisy_trajectory = self.noise_scheduler.add_noise(trajectory, eps, timesteps)
 
         # Run the denoising network (that might denoise the trajectory, or attempt to predict the noise).
-        pred = self.unet(noisy_trajectory, timesteps, global_cond=global_cond)
+        pred = self.net(noisy_trajectory, timesteps, global_cond=global_cond)
 
         # Compute the loss.
         # The target is either the original trajectory, or the noise.
@@ -749,3 +768,264 @@ def forward(self, x: Tensor, cond: Tensor) -> Tensor:
         out = self.conv2(out)
         out = out + self.residual_conv(x)
         return out
+
+
+class TransformerForDiffusion(nn.Module):
+    def __init__(self, config: DiffusionConfig, cond_dim: int):
+        super().__init__()
+        self.config = config
+
+        # compute number of tokens for main trunk and condition encoder
+        if config.n_obs_steps is None:
+            config.n_obs_steps = config.horizon
+
+        t = config.horizon
+        t_cond = 1
+        t_cond += config.n_obs_steps
+
+        input_dim = config.output_shapes["action"][0]
+        # input embedding stem
+        self.input_emb = nn.Linear(input_dim, config.diffusion_step_embed_dim)
+        self.pos_emb = nn.Parameter(torch.zeros(1, t, config.diffusion_step_embed_dim))
+        self.drop = nn.Dropout(config.p_drop_emb)
+
+        # cond encoder
+        self.time_emb = DiffusionSinusoidalPosEmb(config.diffusion_step_embed_dim)
+        self.cond_obs_emb = None
+
+        self.cond_obs_emb = nn.Linear(cond_dim, config.diffusion_step_embed_dim)
+
+        self.cond_pos_emb = None
+        self.encoder = None
+        self.decoder = None
+
+        self.cond_pos_emb = nn.Parameter(torch.zeros(1, t_cond, config.diffusion_step_embed_dim))
+        if config.n_cond_layers > 0:
+            encoder_layer = nn.TransformerEncoderLayer(
+                d_model=config.diffusion_step_embed_dim,
+                nhead=config.n_head,
+                dim_feedforward=4 * config.diffusion_step_embed_dim,
+                dropout=config.p_drop_attn,
+                activation="gelu",
+                batch_first=True,
+                norm_first=True,
+            )
+            self.encoder = nn.TransformerEncoder(encoder_layer=encoder_layer, num_layers=config.n_cond_layers)
+        else:
+            self.encoder = nn.Sequential(
+                nn.Linear(config.diffusion_step_embed_dim, 4 * config.diffusion_step_embed_dim),
+                nn.Mish(),
+                nn.Linear(4 * config.diffusion_step_embed_dim, config.diffusion_step_embed_dim),
+            )
+        # decoder
+        decoder_layer = nn.TransformerDecoderLayer(
+            d_model=config.diffusion_step_embed_dim,
+            nhead=config.n_head,
+            dim_feedforward=4 * config.diffusion_step_embed_dim,
+            dropout=config.p_drop_attn,
+            activation="gelu",
+            batch_first=True,
+            norm_first=True,  # important for stability
+        )
+        self.decoder = nn.TransformerDecoder(decoder_layer=decoder_layer, num_layers=config.n_layer)
+
+        # attention mask
+        if config.causal_attn:
+            # causal mask to ensure that attention is only applied to the left in the input sequence
+            # torch.nn.Transformer uses additive mask as opposed to multiplicative mask in minGPT
+            # therefore, the upper triangle should be -inf and others (including diag) should be 0.
+            sz = t
+            mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
+            mask = mask.float().masked_fill(mask == 0, float("-inf")).masked_fill(mask == 1, float(0.0))
+            self.register_buffer("mask", mask)
+
+            # assume conditioning over time and observation both
+            p, q = torch.meshgrid(torch.arange(t), torch.arange(t_cond), indexing="ij")
+            mask = p >= (q - 1)  # add one dimension since time is the first token in cond
+            mask = mask.float().masked_fill(mask == 0, float("-inf")).masked_fill(mask == 1, float(0.0))
+            self.register_buffer("memory_mask", mask)
+        else:
+            self.mask = None
+            self.memory_mask = None
+
+        # decoder head
+        self.ln_f = nn.LayerNorm(config.diffusion_step_embed_dim)
+        self.head = nn.Linear(config.diffusion_step_embed_dim, input_dim)
+
+        # constants
+        self.t = t
+        self.t_cond = t_cond
+        self.horizon = config.horizon
+        self.n_obs_steps = config.n_obs_steps
+
+        # init
+        self.apply(self._init_weights)
+        # logger.info(
+        #     "number of parameters: %e", sum(p.numel() for p in self.parameters())
+        # )
+
+    def _init_weights(self, module):
+        ignore_types = (
+            nn.Dropout,
+            DiffusionSinusoidalPosEmb,
+            nn.TransformerEncoderLayer,
+            nn.TransformerDecoderLayer,
+            nn.TransformerEncoder,
+            nn.TransformerDecoder,
+            nn.ModuleList,
+            nn.Mish,
+            nn.Sequential,
+        )
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if isinstance(module, nn.Linear) and module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.MultiheadAttention):
+            weight_names = ["in_proj_weight", "q_proj_weight", "k_proj_weight", "v_proj_weight"]
+            for name in weight_names:
+                weight = getattr(module, name)
+                if weight is not None:
+                    torch.nn.init.normal_(weight, mean=0.0, std=0.02)
+
+            bias_names = ["in_proj_bias", "bias_k", "bias_v"]
+            for name in bias_names:
+                bias = getattr(module, name)
+                if bias is not None:
+                    torch.nn.init.zeros_(bias)
+        elif isinstance(module, nn.LayerNorm):
+            torch.nn.init.zeros_(module.bias)
+            torch.nn.init.ones_(module.weight)
+        elif isinstance(module, TransformerForDiffusion):
+            torch.nn.init.normal_(module.pos_emb, mean=0.0, std=0.02)
+            if module.cond_obs_emb is not None:
+                torch.nn.init.normal_(module.cond_pos_emb, mean=0.0, std=0.02)
+        elif isinstance(module, ignore_types):
+            # no param
+            pass
+        else:
+            raise RuntimeError("Unaccounted module {}".format(module))
+
+    def get_optim_groups(self, weight_decay: float = 1e-3):
+        """
+        This long function is unfortunately doing something very simple and is being very defensive:
+        We are separating out all parameters of the model into two buckets: those that will experience
+        weight decay for regularization and those that won't (biases, and layernorm/embedding weights).
+        We are then returning the PyTorch optimizer object.
+        """
+
+        # separate out all parameters to those that will and won't experience regularizing weight decay
+        decay = set()
+        no_decay = set()
+        whitelist_weight_modules = (torch.nn.Linear, torch.nn.MultiheadAttention)
+        blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
+        for mn, m in self.named_modules():
+            for pn, _ in m.named_parameters():
+                fpn = "{}.{}".format(mn, pn) if mn else pn  # full param name
+
+                if pn.endswith("bias"):
+                    # all biases will not be decayed
+                    no_decay.add(fpn)
+                elif pn.startswith("bias"):
+                    # MultiheadAttention bias starts with "bias"
+                    no_decay.add(fpn)
+                elif pn.endswith("weight") and isinstance(m, whitelist_weight_modules):
+                    # weights of whitelist modules will be weight decayed
+                    decay.add(fpn)
+                elif pn.endswith("weight") and isinstance(m, blacklist_weight_modules):
+                    # weights of blacklist modules will NOT be weight decayed
+                    no_decay.add(fpn)
+
+        # special case the position embedding parameter in the root GPT module as not decayed
+        no_decay.add("pos_emb")
+        # no_decay.add("_dummy_variable")
+        if self.cond_pos_emb is not None:
+            no_decay.add("cond_pos_emb")
+
+        # validate that we considered every parameter
+        # param_dict = {pn: p for pn, p in self.named_parameters()}
+        param_dict = dict(self.named_parameters())
+        inter_params = decay & no_decay
+        union_params = decay | no_decay
+        assert len(inter_params) == 0, "parameters {} made it into both decay/no_decay sets!".format(
+            str(inter_params)
+        )
+        assert (
+            len(param_dict.keys() - union_params) == 0
+        ), "parameters {} were not separated into either decay/no_decay set!".format(
+            str(param_dict.keys() - union_params),
+        )
+
+        # create the pytorch optimizer object
+        optim_groups = [
+            {
+                "params": [param_dict[pn] for pn in sorted(decay)],
+                "weight_decay": weight_decay,
+            },
+            {
+                "params": [param_dict[pn] for pn in sorted(no_decay)],
+                "weight_decay": 0.0,
+            },
+        ]
+        return optim_groups
+
+    def configure_optimizers(
+        self,
+        learning_rate: float = 1e-4,
+        weight_decay: float = 1e-3,
+        betas: Tuple[float, float] = (0.9, 0.95),
+    ):
+        optim_groups = self.get_optim_groups(weight_decay=weight_decay)
+        optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas)
+        return optimizer
+
+    def forward(self, sample: torch.Tensor, timestep: torch.Tensor, global_cond: torch.Tensor, **kwargs):
+        """
+        x: (B,T,input_dim)
+        timestep: (B,)
+        global_cond: (B, global_cond_dim)
+        output: (B,T,input_dim)
+        """
+        # 1. time
+        timesteps = timestep
+        batch_size = sample.shape[0]
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(batch_size)
+        time_emb = self.time_emb(timesteps).unsqueeze(1)
+        # (B,1,n_emb)
+
+        cond = einops.rearrange(global_cond, "b (s n) ... -> b s (n ...)", b=batch_size, s=self.n_obs_steps)
+        # (B,To,n_cond)
+
+        # process input
+        input_emb = self.input_emb(sample)
+
+        # encoder
+        cond_embeddings = time_emb
+        # (B,1,n_emb)
+
+        cond_obs_emb = self.cond_obs_emb(cond)
+        # (B,To,n_emb)
+        cond_embeddings = torch.cat([cond_embeddings, cond_obs_emb], dim=1)
+        # (B,To + 1,n_emb)
+
+        tc = cond_embeddings.shape[1]
+        position_embeddings = self.cond_pos_emb[:, :tc, :]  # each position maps to a (learnable) vector
+        x = self.drop(cond_embeddings + position_embeddings)
+        x = self.encoder(x)
+        memory = x
+        # (B,T_cond,n_emb)
+
+        # decoder
+        token_embeddings = input_emb
+        t = token_embeddings.shape[1]
+        position_embeddings = self.pos_emb[:, :t, :]  # each position maps to a (learnable) vector
+        x = self.drop(token_embeddings + position_embeddings)
+        # (B,T,n_emb)
+        x = self.decoder(tgt=x, memory=memory, tgt_mask=self.mask, memory_mask=self.memory_mask)
+        # (B,T,n_emb)
+
+        # head
+        x = self.ln_f(x)
+        x = self.head(x)
+        # (B,T,n_inp)
+        return x