Update

config.json

@@ -1,8 +1,12 @@
 {
   "arch_specifier": "no-align+fused-gelu-mlp",
   "architectures": [
-    "OpenVLAForActionPrediction"
+    "Emma-X"
   ],
+  "auto_map": {
+    "AutoConfig": "configuration_prismatic.OpenVLAConfig",
+    "AutoModelForVision2Seq": "modeling_prismatic.EmmaxForActionPrediction"
+  },
   "hf_llm_id": "meta-llama/Llama-2-7b-hf",
   "image_resize_strategy": "resize-naive",
   "image_sizes": [

configuration_prismatic.py

@@ -0,0 +1,140 @@
+"""
+configuration_prismatic.py
+
+HuggingFace-style configuration definition for Prismatic VLMs, inheriting from `transformers.PretrainedConfig`.
+Default configuration specifies `siglip-224px+7b`.
+"""
+
+from typing import Any, Dict, List, Optional
+
+from transformers import PretrainedConfig
+from transformers.models.auto import CONFIG_MAPPING
+
+# === Utilities for Mapping Prismatic names to HF names ===
+# fmt: off
+VISION_BACKBONE_TO_RESOLUTION: Dict[str, List[int]] = {
+    "clip-vit-l": [224], "siglip-vit-so400m": [224], "dinov2-vit-l": [224], "in1k-vit-l": [224],
+
+    "clip-vit-l-336px": [336],
+    "siglip-vit-so400m-384px": [384],
+
+    "dinoclip-vit-l-336px": [336, 336],
+    "dinosiglip-vit-so-224px": [224, 224],
+    "dinosiglip-vit-so-384px": [384, 384],
+}
+VISION_BACKBONE_TO_TIMM_ID: Dict[str, List[str]] = {
+    "clip-vit-l": ["vit_large_patch14_clip_224.openai"],
+    "clip-vit-l-336px": ["vit_large_patch14_clip_336.openai"],
+
+    "dinov2-vit-l": ["vit_large_patch14_reg4_dinov2.lvd142m"],
+    "in1k-vit-l": ["vit_large_patch16_224.augreg_in21k_ft_in1k"],
+
+    "siglip-vit-so400m": ["vit_so400m_patch14_siglip_224"],
+    "siglip-vit-so400m-384px": ["vit_so400m_patch14_siglip_384"],
+
+    "dinoclip-vit-l-336px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_large_patch14_clip_336.openai"],
+    "dinosiglip-vit-so-224px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_so400m_patch14_siglip_224"],
+    "dinosiglip-vit-so-384px": ["vit_large_patch14_reg4_dinov2.lvd142m", "vit_so400m_patch14_siglip_384"],
+}
+TIMM_OVERRIDE_ACT_LAYER: Dict[str, List[Optional[str]]] = {
+    "clip-vit-l": ["quick_gelu"], "clip-vit-l-336px": ["quick_gelu"],
+    "dinov2-vit-l": [None], "in1k-vit-l": [None],
+    "siglip-vit-so400m": [None], "siglip-vit-so400m-384px": [None],
+    "dinoclip-vit-l-336px": [None, "quick_gelu"],
+    "dinosiglip-vit-so-224px": [None, None], "dinosiglip-vit-so-384px": [None, None]
+}
+
+LLM_BACKBONE_TO_HF_PATH = {
+    "llama2-7b-pure": "meta-llama/Llama-2-7b-hf", "llama2-13b-pure": "meta-llama/Llama-2-13b-hf",
+    "llama2-7b-chat": "meta-llama/Llama-2-7b-chat-hf", "llama2-13b-chat": "meta-llama/Llama-2-13b-chat-hf",
+
+    "vicuna-v15-7b": "lmsys/vicuna-7b-v1.5", "vicuna-v15-13b": "lmsys/vicuna-13b-v1.5",
+
+    "mistral-v0.1-7b-pure": "mistralai/Mistral-7B-v0.1",
+    "mistral-v0.1-7b-instruct": "mistralai/Mistral-7B-Instruct-v0.1",
+
+    "phi-2-3b": "microsoft/phi-2",
+}
+LLM_BACKBONE_TO_HF_METACLASS = {
+    "llama2-7b-pure": "llama", "llama2-13b-pure": "llama", "llama2-7b-chat": "llama", "llama2-13b-chat": "llama",
+    "vicuna-v15-7b": "llama", "vicuna-v15-13b": "llama",
+
+    "mistral-v0.1-7b-pure": "mistral", "mistral-v0.1-7b-instruct": "mistral",
+
+    "phi-2-3b": "phi",
+}
+
+VALID_VISION_BACKBONES = set(VISION_BACKBONE_TO_RESOLUTION.keys())
+VALID_LLM_BACKBONES = set(LLM_BACKBONE_TO_HF_PATH)
+# fmt: on
+
+
+class PrismaticConfig(PretrainedConfig):
+    model_type: str = "prismatic"
+    is_composition: bool = False
+
+    def __init__(
+        self,
+        vision_backbone_id: str = "siglip-vit-so400m",
+        llm_backbone_id: str = "vicuna-v15-7b",
+        arch_specifier: str = "no-align+gelu-mlp",
+        use_fused_vision_backbone: Optional[bool] = None,
+        image_resize_strategy: str = "letterbox",
+        text_config: Optional[Dict[str, Any]] = None,
+        llm_max_length: int = 2048,
+        pad_token_id: int = 32000,
+        pad_to_multiple_of: int = 64,
+        output_projector_states: bool = False,
+        **kwargs: str,
+    ) -> None:
+        if vision_backbone_id not in VALID_VISION_BACKBONES:
+            raise ValueError(f"Vision backbone `{vision_backbone_id}` not in {VALID_VISION_BACKBONES = }")
+
+        if llm_backbone_id not in VALID_LLM_BACKBONES:
+            raise ValueError(f"LLM backbone `{llm_backbone_id}` not in {VALID_LLM_BACKBONES = }")
+
+        # Set Prismatic Configuration Fields
+        self.vision_backbone_id = vision_backbone_id
+        self.llm_backbone_id = llm_backbone_id
+        self.arch_specifier = arch_specifier
+        self.output_projector_states = output_projector_states
+
+        # [Contract] All vision backbone parameters are lists =>> supports fused backbones with different preprocessing
+        self.use_fused_vision_backbone = (
+            use_fused_vision_backbone
+            if use_fused_vision_backbone is not None
+            else any(self.vision_backbone_id.startswith(v) for v in ["dinoclip", "dinosiglip"])
+        )
+
+        self.timm_model_ids = VISION_BACKBONE_TO_TIMM_ID[self.vision_backbone_id]
+        self.timm_override_act_layers = TIMM_OVERRIDE_ACT_LAYER[self.vision_backbone_id]
+        self.image_sizes = VISION_BACKBONE_TO_RESOLUTION[self.vision_backbone_id]
+        self.image_resize_strategy = image_resize_strategy
+
+        self.hf_llm_id = LLM_BACKBONE_TO_HF_PATH[self.llm_backbone_id]
+        self.llm_max_length = llm_max_length
+        self.pad_token_id, self.pad_to_multiple_of = pad_token_id, pad_to_multiple_of
+
+        # [IMPORTANT] HF Utilities actually look for a `text_config` field... we need to use that specific naming!
+        self.text_config = (
+            CONFIG_MAPPING[LLM_BACKBONE_TO_HF_METACLASS[self.llm_backbone_id]](**text_config)
+            if text_config is not None
+            else CONFIG_MAPPING[LLM_BACKBONE_TO_HF_METACLASS[self.llm_backbone_id]]()
+        )
+
+        # Dispatch **kwargs to super() =>> note that `pad_token_id` collides, so we pass it in here as well...
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+
+class OpenVLAConfig(PrismaticConfig):
+    model_type: str = "openvla"
+
+    def __init__(
+        self,
+        norm_stats: Optional[Dict[str, Dict[str, Dict[str, Dict[str, List[float]]]]]] = None,
+        n_action_bins: int = 256,
+        **kwargs: str,
+    ) -> None:
+        self.norm_stats, self.n_action_bins = norm_stats, n_action_bins
+
+        super().__init__(**kwargs)

modeling_prismatic.py

@@ -0,0 +1,619 @@
+"""
+modeling_prismatic.py
+
+Core HuggingFace-style PrismaticPreTrainedModel and PrismaticForConditionalGeneration class definitions, inheriting
+from the default `transformers.PretrainedModel`. Meant to be standalone and self-contained, but exactly replicate the
+logic in `prismatic.models.vlms.prismatic.py`.
+
+Note =>> for the time being, not adding the custom HF "docstring" formatting.
+
+References [LLaVa, IDEFICS-2]:
+    => https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava/modeling_llava.py
+    => https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics2/modeling_idefics2.py
+"""
+
+import logging
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Callable, ClassVar, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import timm
+import tokenizers
+import torch
+import torch.nn as nn
+import transformers
+from timm.models.vision_transformer import LayerScale
+from transformers import AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
+from transformers.modeling_outputs import ModelOutput
+
+from PIL import Image
+
+from .configuration_prismatic import OpenVLAConfig, PrismaticConfig
+from solver import solver
+
+# Get Logger
+logger = logging.getLogger(__name__)
+
+
+# === PyTorch/HuggingFace Default IGNORE_INDEX (for CrossEntropyLoss labels)
+IGNORE_INDEX = -100
+
+
+# === Utility Functions for Monkey-Patching ===
+def unpack_tuple(fn: Callable[[Any], Tuple[Any]]) -> Callable[[Any], Any]:
+    def wrapper(*args: Any, **kwargs: Any) -> Any:
+        result = fn(*args, **kwargs)
+        return result[0] if isinstance(result, tuple) else result
+
+    return wrapper
+
+
+# HF Transformers overwrites parameters with names containing `gamma`; we're going to patch VisionBackbone.LayerScale.
+#   =>> TIMM :: https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L109
+#   =>> Transformers :: https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_utils.py#L3960
+def _ls_new_forward(self, x: torch.Tensor) -> torch.Tensor:
+    return x.mul_(self.scale_factor) if self.inplace else x * self.scale_factor
+
+
+def ls_apply_patch(ls_module: LayerScale):
+    ls_module.scale_factor = nn.Parameter(ls_module.gamma.clone())
+    ls_module.forward = _ls_new_forward.__get__(ls_module, LayerScale)
+    del ls_module.gamma
+
+
+# === Prismatic Vision Backbone (nn.Module) Definitions (w/ Fused Backbone Support) ===
+class PrismaticVisionBackbone(nn.Module):
+    def __init__(
+        self,
+        use_fused_vision_backbone: bool,
+        image_sizes: List[int],
+        timm_model_ids: List[str],
+        timm_override_act_layers: List[Optional[str]],
+    ) -> None:
+        super().__init__()
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+
+        # [Contract] Validate number of (fused) vision backbones, create "alpha" featurizer and Instantiate
+        #   =>> Note :: Monkey-Patch the `forward()` function of the backbone to ensure FSDP-compatibility
+        #               Hardcodes `get_intermediate_layers` to return the **SECOND-TO-LAST** layer patches!
+        assert len(timm_model_ids) <= 2, "Prismatic models only support up to 2 (fused) vision backbones!"
+        self.featurizer = timm.create_model(
+            timm_model_ids[0],
+            pretrained=False,
+            num_classes=0,
+            img_size=image_sizes[0],
+            act_layer=timm_override_act_layers[0],
+        )
+        self.featurizer.forward = unpack_tuple(
+            partial(self.featurizer.get_intermediate_layers, n={len(self.featurizer.blocks) - 2})
+        )
+        self.embed_dim = self.featurizer.embed_dim
+
+        # If `use_fused_vision_backbone` =>> create "beta" featurizer
+        if self.use_fused_vision_backbone:
+            self.fused_featurizer = timm.create_model(
+                timm_model_ids[1],
+                pretrained=False,
+                num_classes=0,
+                img_size=image_sizes[1],
+                act_layer=timm_override_act_layers[1],
+            )
+            self.fused_featurizer.forward = unpack_tuple(
+                partial(self.fused_featurizer.get_intermediate_layers, n={len(self.fused_featurizer.blocks) - 2})
+            )
+            self.embed_dim += self.fused_featurizer.embed_dim
+
+        # Patch `vision_backbone.featurizer` and `vision_backbone.fused_featurizer` with HF-Compatible LayerScale
+        for module in self.featurizer.modules():
+            if isinstance(module, LayerScale):
+                ls_apply_patch(module)
+
+        if self.use_fused_vision_backbone:
+            for module in self.fused_featurizer.modules():
+                if isinstance(module, LayerScale):
+                    ls_apply_patch(module)
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
+        """Run image (`pixel_values`) through featurizer; if channel-stacked, then dispatch and sequence stack."""
+        if not self.use_fused_vision_backbone:
+            return self.featurizer(pixel_values)
+
+        # Split `pixel_values :: [bsz, 2 * 3, resolution, resolution]` =>> featurize =>> channel stack
+        img, img_fused = torch.split(pixel_values, [3, 3], dim=1)
+        patches, patches_fused = self.featurizer(img), self.fused_featurizer(img_fused)
+
+        return torch.cat([patches, patches_fused], dim=2)
+
+
+# === Prismatic Projector (nn.Module) Definitions ===
+class PrismaticProjector(nn.Module):
+    def __init__(self, use_fused_vision_backbone: bool, vision_dim: int, llm_dim: int) -> None:
+        super().__init__()
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+        self.vision_dim, self.llm_dim = vision_dim, llm_dim
+
+        # Switch on `use_fused_vision_backbone` =>> use slightly different MLPs and projection factors!
+        if not self.use_fused_vision_backbone:
+            self.fc1 = nn.Linear(self.vision_dim, self.llm_dim, bias=True)
+            self.fc2 = nn.Linear(self.llm_dim, self.llm_dim, bias=True)
+            self.act_fn1 = nn.GELU()
+        else:
+            initial_projection_dim = 4 * vision_dim
+            self.fc1 = nn.Linear(self.vision_dim, initial_projection_dim, bias=True)
+            self.fc2 = nn.Linear(initial_projection_dim, self.llm_dim, bias=True)
+            self.fc3 = nn.Linear(self.llm_dim, self.llm_dim, bias=True)
+            self.act_fn1 = nn.GELU()
+            self.act_fn2 = nn.GELU()
+
+    def forward(self, img_patches: torch.Tensor) -> torch.Tensor:
+        if not self.use_fused_vision_backbone:
+            projected_features = self.fc1(img_patches)
+            projected_features = self.act_fn1(projected_features)
+            projected_features = self.fc2(projected_features)
+        else:
+            projected_features = self.fc1(img_patches)
+            projected_features = self.act_fn1(projected_features)
+            projected_features = self.fc2(projected_features)
+            projected_features = self.act_fn2(projected_features)
+            projected_features = self.fc3(projected_features)
+
+        return projected_features
+
+
+# === Main HF Class Definitions ===
+@dataclass
+class PrismaticCausalLMOutputWithPast(ModelOutput):
+    """Base class for Prismatic casual (visually-conditioned) language model outputs; also exposes visual features."""
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+    # Additions for VLMs
+    projector_features: Optional[torch.FloatTensor] = None
+
+
+class PrismaticPreTrainedModel(PreTrainedModel):
+    config_class: PretrainedConfig = PrismaticConfig
+    base_model_prefix: str = "model"
+    supports_gradient_checkpointing: bool = True
+
+    _no_split_modules: ClassVar[List[str]] = ["PrismaticProjector"]
+    _skip_keys_device_placement: str = "past_key_values"
+    _supports_flash_attn_2: bool = True
+
+    def _init_weights(self, module: nn.Module) -> None:
+        # Important :: this HF ported version is *not* meant for training from scratch; only inference and fine-tuning!
+        #   => As such, this init_weights code is not correct; if training VLMs from scratch, use the main codebase at
+        #      https://github.com/TRI-ML/prismatic-vlms
+        std = (
+            self.config.initializer_range
+            if hasattr(self.config, "initializer_range")
+            else self.config.text_config.initializer_range
+        )
+
+        if hasattr(module, "class_embedding"):
+            module.class_embedding.data.normal_(mean=0.0, std=std)
+
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+    @property
+    def _supports_sdpa(self) -> bool:
+        """Check LLM supports SDPA Attention"""
+        return self.language_model._supports_sdpa
+
+
+class PrismaticForConditionalGeneration(PrismaticPreTrainedModel):
+    def __init__(self, config: PrismaticConfig) -> None:
+        super().__init__(config)
+
+        # [Validation] Lightweight Validate on `config` Fields + Dependency Versions
+        if config.use_fused_vision_backbone is None:
+            raise ValueError("Missing config field `use_fused_vision_backbone`")
+
+        if timm.__version__ not in {"0.9.10", "0.9.11", "0.9.12", "0.9.16"}:
+            raise NotImplementedError(
+                "TIMM Version must be >= 0.9.10 and < 1.0.0 (breaking); please raise a GitHub Issue "
+                "if you urgently need support for latest TIMM versions."
+            )
+
+        if (transformers.__version__ != "4.40.1") or (tokenizers.__version__ != "0.19.1"):
+            logger.warning(
+                f"Expected `transformers==4.40.1` and `tokenizers==0.19.1` but got "
+                f"`transformers=={transformers.__version__}` and `tokenizers=={tokenizers.__version__}`; "
+                f"there might be inference-time regressions due to dependency changes. If in doubt, please"
+                f"use the above versions."
+            )
+
+        # Instantiate PrismaticVisionBackbone (w/ Potential Fused Backbone)
+        self.vision_backbone = PrismaticVisionBackbone(
+            config.use_fused_vision_backbone, config.image_sizes, config.timm_model_ids, config.timm_override_act_layers
+        )
+
+        # Create Multimodal Projector
+        self.projector = PrismaticProjector(
+            config.use_fused_vision_backbone,
+            vision_dim=self.vision_backbone.embed_dim,
+            llm_dim=config.text_config.hidden_size,
+        )
+
+        # Instantiate LLM Backbone
+        self.language_model = AutoModelForCausalLM.from_config(
+            config.text_config, attn_implementation=config._attn_implementation
+        )
+        self.vocab_size = config.text_config.vocab_size
+        self.pad_token_id = config.pad_token_id
+
+        # HF Boilerplate =>> initializes weights via `_init_weights()` and sets gradient checkpointing
+        self.post_init()
+
+    # === `PreTrainedModel` Boilerplate ===
+    def get_input_embeddings(self) -> nn.Module:
+        return self.language_model.get_input_embeddings()
+
+    def set_input_embeddings(self, value: nn.Module) -> None:
+        self.language_model.set_input_embeddings(value)
+
+    def get_output_embeddings(self) -> nn.Module:
+        return self.language_model.get_output_embeddings()
+
+    def set_output_embeddings(self, new_embeddings: nn.Module) -> None:
+        self.language_model.set_output_embeddings(new_embeddings)
+
+    def get_decoder(self) -> nn.Module:
+        return self.language_model.get_decoder()
+
+    def set_decoder(self, decoder: nn.Module) -> None:
+        self.language_model.set_decoder(decoder)
+
+    def tie_weights(self) -> None:
+        self.language_model.tie_weights()  # Note: `Llama-2` and `Mistral` don't tie weights (no-op)
+
+    def resize_token_embeddings(
+        self, new_num_tokens: Optional[int] = None, pad_to_multiple_of: Optional[int] = None
+    ) -> nn.Embedding:
+        updated_embeddings = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
+
+        # Update config/instance variables
+        self.config.text_config.vocab_size = updated_embeddings.num_embeddings
+        self.vocab_size = updated_embeddings.num_embeddings
+
+        return updated_embeddings
+
+    # === Core Prismatic VLM `forward()` Logic ===
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        output_projector_features: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, PrismaticCausalLMOutputWithPast]:
+        """Run a forward pass through the VLM, returning a PrismaticCausalLMOutputWithPast instance."""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        output_projector_features = output_projector_features if output_projector_features is not None else False
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # Respect `use_cache` only if not training (even if `gradient_checkpointing` is off)
+        use_cache = use_cache and not self.training
+
+        # Instantiate Placeholder for Projector Features
+        projected_patch_embeddings = None
+
+        # Note :: We only support forward passes with the following cases:
+        #   => Cached Generation :: (input_ids.shape[1] == 1) and (past_key_values is not None)
+        #   => Unimodal Forward :: (pixel_values is None)
+        #   => Multimodal Forward :: (pixel_values is not None) and (input_ids/embeds.shape[0] == pixel_values.shape[0])
+
+        # === Handle Generation with Cache (`input_ids.shape[1] == 1`) =>> requires `past_keys_values` ===
+        if input_ids.shape[1] == 1:
+            assert input_ids.shape[0] == 1, "Generation is only currently supported for batch size of 1!"
+            assert past_key_values is not None, "You must provide `past_key_values` during cached generation!"
+            assert labels is None, "Unexpected key `labels` provided during cached generation!"
+
+            language_model_output = self.language_model(
+                input_ids=input_ids,
+                attention_mask=None,
+                position_ids=None,
+                past_key_values=past_key_values,
+                inputs_embeds=None,
+                labels=None,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Handle Unimodal Forward ===
+        elif pixel_values is None:
+            assert (input_ids is not None) and (inputs_embeds is None), "Missing `input_ids` in language-only forward!"
+            assert past_key_values is None, "Unexpected key `past_key_values` provided during language-only forward!"
+
+            language_model_output = self.language_model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=None,
+                labels=labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Handle Multimodal Forward ===
+        elif (input_ids.shape[0] == pixel_values.shape[0]) or (inputs_embeds.shape[0] == pixel_values.shape[0]):
+            assert past_key_values is None, "Unexpected key `past_key_values` provided during language-only forward!"
+
+            # Visual Feature Extraction
+            patch_features = self.vision_backbone(pixel_values)
+
+            # Projection Logic =>> Update Attention Mask
+            projected_patch_embeddings = self.projector(patch_features)
+            projected_patch_attention_mask = None
+            if attention_mask is not None:
+                projected_patch_attention_mask = torch.full(
+                    (projected_patch_embeddings.shape[0], projected_patch_embeddings.shape[1]),
+                    fill_value=True,
+                    dtype=attention_mask.dtype,
+                    device=attention_mask.device,
+                )
+
+            # Get Input Embeddings (from Language Model Embeddings)
+            input_embeddings = self.get_input_embeddings()(input_ids)
+
+            # Build Multimodal Embeddings & Attention Mask =>> Prismatic defaults to inserting after <BOS> token (1:)
+            multimodal_embeddings = torch.cat(
+                [input_embeddings[:, :1, :], projected_patch_embeddings, input_embeddings[:, 1:, :]], dim=1
+            )
+            multimodal_attention_mask = None
+            if attention_mask is not None:
+                multimodal_attention_mask = torch.cat(
+                    [attention_mask[:, :1], projected_patch_attention_mask, attention_mask[:, 1:]], dim=1
+                )
+
+            # Build Labels (if specified) =>> Ignore Labels for Patch Embeddings
+            multimodal_labels = None
+            if labels is not None:
+                projected_patch_labels = torch.full(
+                    (projected_patch_embeddings.shape[0], projected_patch_embeddings.shape[1]),
+                    fill_value=IGNORE_INDEX,
+                    dtype=labels.dtype,
+                    device=labels.device,
+                )
+                multimodal_labels = torch.cat([labels[:, :1], projected_patch_labels, labels[:, 1:]], dim=1)
+
+            # Dispatch to Language Model
+            language_model_output = self.language_model(
+                input_ids=None,
+                attention_mask=multimodal_attention_mask,
+                position_ids=None,
+                past_key_values=None,
+                inputs_embeds=multimodal_embeddings,
+                labels=multimodal_labels,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+
+        # === Otherwise =>> Assume Invalid! ===
+        elif (input_ids.shape[0] != pixel_values.shape[0]) or (inputs_embeds.shape[0] != pixel_values.shape[0]):
+            raise ValueError("Non-homogenous batch of (text, image) input -- forward() does not support mixed batches!")
+
+        else:
+            raise ValueError(
+                "Invalid PrismaticForConditionalGeneration `forward()` call with provided arguments:\n"
+                f"=> `input_ids` = {input_ids is not None}\n"
+                f"=> `attention_mask` = {attention_mask is not None}\n"
+                f"=> `pixel_values` = {pixel_values is not None}\n"
+                f"=> `labels` = {labels is not None}\n"
+                f"=> `input_embeds` = {inputs_embeds is not None}\n"
+                f"=> `past_key_values` = {past_key_values is not None}\n"
+                f"=> `use_cache` = {use_cache}"
+            )
+
+        # Unpack `language_model_output` and return PrismaticCausalLMOutputWithPast (or tuple if not `return_dict`)
+        if not return_dict:
+            if output_projector_features and (projected_patch_embeddings is not None):
+                return *language_model_output, projected_patch_embeddings
+
+            return language_model_output
+
+        return PrismaticCausalLMOutputWithPast(
+            loss=language_model_output.loss,
+            logits=language_model_output.logits,
+            past_key_values=language_model_output.past_key_values,
+            hidden_states=language_model_output.hidden_states,
+            attentions=language_model_output.attentions,
+            projector_features=projected_patch_embeddings,
+        )
+
+    # === GenerationMixin Methods ===
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **kwargs: str,
+    ) -> Dict[str, torch.Tensor]:
+        """Borrowed from `LlamaForCausalLM` and simplified for batch size = 1; mirrors original PrismaticVLM logic."""
+        if ((input_ids is not None) and (input_ids.shape[0] > 1)) or (
+            (inputs_embeds is not None) and (inputs_embeds.shape[0] > 1)
+        ):
+            raise ValueError("Generation with batch size > 1 is not currently supported!")
+
+        # Handle `past_key_values` (cache) =>> assume `input_ids` just has unprocessed tokens
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+
+        # If `input_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"input_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        # Make sure `pixel_values` are preserved in `model_inputs`
+        model_inputs.update(
+            {
+                "attention_mask": attention_mask,
+                "pixel_values": pixel_values,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+            }
+        )
+
+        return model_inputs
+
+    # Defer to Language Model (all handle this differently, with different return types)
+    def _reorder_cache(self, *args, **kwargs) -> Any:
+        return self.language_model._reorder_cache(*args, **kwargs)
+
+
+class EmmaxForActionPrediction(PrismaticForConditionalGeneration):
+    config_class: PretrainedConfig = OpenVLAConfig
+
+    def __init__(self, config: OpenVLAConfig) -> None:
+        super().__init__(config)
+        self.norm_stats = config.norm_stats
+
+        # Compute action bins
+        self.bins = np.linspace(-1, 1, config.n_action_bins)
+        self.bin_centers = (self.bins[:-1] + self.bins[1:]) / 2.0
+
+        # Compute vocab size for de-tokenization -- revert added "multiple of"
+        self.vocab_size = self.config.text_config.vocab_size - self.config.pad_to_multiple_of
+
+    def predict_action(
+        self, input_ids: Optional[torch.LongTensor] = None, unnorm_key: Optional[str] = None, **kwargs: str
+    ) -> np.ndarray:
+        """Thin wrapper around super().generate() that decodes predicted actions and de-normalizes them."""
+
+        # We need to add this special empty token ('') after the colon (':') token in "ASSISTANT:"
+        # in order for the predictions to match the training configuration and be accurate.
+        # NOTE: This is NOT needed for ECoT
+        # input_ids = torch.cat(
+        #     (input_ids, torch.unsqueeze(torch.Tensor([29871]).long(), dim=0).to(input_ids.device)), dim=1
+        # )
+
+        # Run VLA inference
+        generated_ids = self.generate(input_ids, **kwargs)
+
+        # Extract predicted action tokens and translate into (normalized) continuous actions
+        predicted_action_token_ids = generated_ids[0, -(self.get_action_dim(unnorm_key) + 1) : -1].cpu().numpy()
+        discretized_actions = self.vocab_size - predicted_action_token_ids
+        discretized_actions = np.clip(discretized_actions - 1, a_min=0, a_max=self.bin_centers.shape[0] - 1)
+        normalized_actions = self.bin_centers[discretized_actions]
+
+        # Unnormalize actions
+        action_norm_stats = self.get_action_stats(unnorm_key)
+        mask = action_norm_stats.get("mask", np.ones_like(action_norm_stats["q01"], dtype=bool))
+        action_high, action_low = np.array(action_norm_stats["q99"]), np.array(action_norm_stats["q01"])
+        actions = np.where(
+            mask,
+            0.5 * (normalized_actions + 1) * (action_high - action_low) + action_low,
+            normalized_actions,
+        )
+
+
+        return actions, generated_ids
+
+    @torch.inference_mode()
+    def generate_actions(self, image: Image, prompt_text: str, type: str, **kwargs: str) -> str:
+        # For now, only support generation with a batch size of 1 for simplicity
+        image_transform, tokenizer = self.vision_backbone.image_transform, self.llm_backbone.tokenizer
+
+        # Prepare Inputs
+        input_ids = tokenizer(prompt_text, truncation=True, return_tensors="pt").input_ids.to(self.device)
+
+        pixel_values = image_transform(image)
+        if isinstance(pixel_values, torch.Tensor):
+            pixel_values = pixel_values[None, ...].to(self.device)
+        elif isinstance(pixel_values, dict):
+            pixel_values = {k: v[None, ...].to(self.device) for k, v in pixel_values.items()}
+        else:
+            raise ValueError(f"Unsupported `pixel_values` type = {type(pixel_values)}")
+
+        # Invoke super().generate --> taps into `GenerationMixin` which (redirects) to `forward()`
+        autocast_dtype = self.llm_backbone.half_precision_dtype
+        # with torch.autocast("cuda", dtype=autocast_dtype, enabled=self.enable_mixed_precision_training):
+        with torch.autocast("cuda", dtype=torch.float16):
+            # fmt: off
+            generated_ids = super().generate(
+                input_ids=input_ids,            # Shape: [1, seq]
+                pixel_values=pixel_values,      # Shape: [1, 3, res, res] or Dict[str, Shape[1, 3, res, res]]
+                **kwargs
+            )
+            # fmt: on
+
+        generated_text = tokenizer.decode(generated_ids[0, input_ids.shape[1] :], skip_special_tokens=True).strip()
+
+        s = solver
+        actions, reasoning = s.extract_action_policies(generated_text)
+        # unnorm_key = "bridge_orig"
+
+        # unnormalize
+        unnorm_key = None
+        action_norm_stats = self.get_action_stats(unnorm_key)
+        mask = action_norm_stats.get("mask", np.ones_like(action_norm_stats["q01"], dtype=bool))
+        action_high, action_low = np.array(action_norm_stats["q99"]), np.array(action_norm_stats["q01"])
+        _actions = []
+        for action in actions:
+            action_norm = np.where(
+                mask, 0.5 * (np.array(action) + 1) * (action_high - action_low) + action_low, action
+            )
+            _actions.append(action_norm)
+
+        return _actions, generated_text
+
+
+    @staticmethod
+    def _check_unnorm_key(norm_stats: Dict[str, Dict[str, Any]], unnorm_key: Optional[str]) -> str:
+        if unnorm_key is None and len(norm_stats) != 1:
+            raise ValueError(
+                f"Your model was trained on more than one dataset. "
+                f"Please pass a `unnorm_key` from the following options to choose the statistics used for "
+                f"de-normalizing actions: {norm_stats.keys()}"
+            )
+
+        # If None, grab the (singular) dataset in `norm_stats` to use as `unnorm_key`
+        unnorm_key = unnorm_key if unnorm_key is not None else next(iter(norm_stats.keys()))
+        if unnorm_key not in norm_stats:
+            raise ValueError(
+                f"The `unnorm_key` you chose ({unnorm_key = }) is not in the available statistics. "
+                f"Please choose from: {norm_stats.keys()}"
+            )
+
+        return unnorm_key
+
+    def get_action_dim(self, unnorm_key: Optional[str] = None) -> int:
+        """Get the dimensionality of the policy's action space."""
+        unnorm_key = self._check_unnorm_key(self.norm_stats, unnorm_key)
+        return len(self.norm_stats[unnorm_key]["action"]["q01"])
+
+    def get_action_stats(self, unnorm_key: Optional[str] = None) -> Dict[str, Any]:
+        """Get all the logged statistics for the given dataset."""
+        unnorm_key = self._check_unnorm_key(self.norm_stats, unnorm_key)
+        return self.norm_stats[unnorm_key]["action"]

solver.py

@@ -0,0 +1,191 @@
+from collections import defaultdict
+
+import numpy as np
+from prismatic.vla.action_tokenizer import ActionTokenizer
+from transformers import AutoTokenizer
+
+
+class Solver:
+    def __init__(self, action_tokenizer=None, verbose=True) -> None:
+        self.verbose = verbose
+        self.action_tokenizer = action_tokenizer
+        self.coordinates_key = "NEXT GRIPPER:"
+        self.movement_key = "MOVEMENT:"
+        self.policy_key = "POLICIES:"
+
+    def compare_movement(self, pred_pos, label_pos):
+
+        dist = np.sum(np.abs(pred_pos - label_pos))
+        relative_dist = np.sum(np.abs(dist / label_pos))
+        return dist, relative_dist, dist == 0
+
+    def compare_policy(self, pred_pol, label_pol):
+        dist = 0
+        cnt = 0
+        for i in range(min(len(label_pol), len(pred_pol))):
+            for j in range(len(label_pol[0])):
+                dist += label_pol[i][j] == pred_pol[i][j]
+                cnt += 1
+        assert cnt % 7 == 0
+        return dist / cnt
+
+    def extract_2d_coordinates(self, text):
+        try:
+            coordinates_index = text.index(self.coordinates_key) + len(self.coordinates_key)
+            coord = text[coordinates_index:]
+            coord = [o for o in coord.split("\n") if len(o.strip()) != 0]
+            coord = eval(coord[0].strip())
+        except Exception:
+            coord = [0, 0]
+        return coord
+
+    def extract_movement_plan(self, text):
+        require_unorm = None
+        try:
+            # text after key word
+            movement_index = text.index(self.movement_key) + len(self.movement_key)
+            movement_level = text[movement_index:]
+            movement_level = [o for o in movement_level.split("\n") if len(o.strip()) != 0]
+            movement_level = movement_level[0].strip()
+
+            if "gripper" not in movement_level:  # for normalized tokenized version
+                require_unorm = True
+                movement_token_ids = self.action_tokenizer.tokenizer(movement_level, add_special_tokens=False).input_ids
+                movement_norm = self.action_tokenizer.decode_token_ids_to_actions(np.array(movement_token_ids))
+                movement_norm = movement_norm[1:8]
+                assert len(movement_norm) == 7
+            else:  # for unnormalized text version
+                require_unorm = False
+                movement_level = [o for o in movement_level.split(";") if len(o) > 0]
+                movement_level = movement_level[:7]
+
+                position = defaultdict(int)
+                movement_to_pos = dict(
+                    move_backward=(-1, "y"),
+                    move_forward=(1, "y"),
+                    move_right=(-1, "x"),
+                    move_left=(1, "x"),
+                    move_downward=(-1, "z"),
+                    move_upward=(1, "z"),
+                    roll_downward=(-1, "ox"),
+                    roll_upward=(1, "ox"),
+                    swing_downward=(-1, "ox"),
+                    swing_upward=(1, "ox"),
+                    pitch_downward=(-1, "oy"),
+                    pitch_upward=(1, "oy"),
+                    yaw_downward=(-1, "oz"),
+                    yaw_upward=(1, "oz"),
+                    rotate_clockwise=(-1, "oz"),
+                    rotate_counterclockwise=(1, "oz"),
+                    close_gripper=(-1, "grip"),
+                    open_gripper=(1, "grip"),
+                )
+
+                for ml in movement_level:
+                    direction = "_".join(ml.split()[:2])
+                    sign, axis = movement_to_pos[direction]
+                    scale = 1
+                    if "o" in axis:  # for orientation
+                        scale = scale * 1e-3
+                    elif "grip" in axis:  # for gripper
+                        scale = scale
+                    else:  # for xyz
+                        scale = scale / 180 * np.pi
+
+                    if "grip" in axis:
+                        level = round("open" in ml)
+                    else:
+                        level = int(ml.split()[2])
+
+                    position[axis] += sign * scale * level
+                movement_norm = [position[idx] for idx in ["x", "y", "z", "ox", "oy", "oz", "grip"]]
+
+        except:
+            movement_norm = [-100] * 7
+
+        return require_unorm, np.array(movement_norm)
+
+    def extract_action_policies(self, text):
+        try:
+            if self.policy_key in text:
+
+                policy_index = text.index(self.policy_key) + len(self.policy_key)
+                policy = text[policy_index:]
+                remain_text = text[: text.index(self.policy_key)]
+                policies = [o for o in policy.split("\n") if len(o.strip()) != 0]
+                policies = policies[0].strip()
+            else:
+                policies = text.strip()
+                remain_text = ""
+
+            policies_num = []
+            for policy_text in policies.split(";"):
+                policy_token = self.action_tokenizer.tokenizer(policy_text, add_special_tokens=False).input_ids
+                action_policy = self.action_tokenizer.decode_token_ids_to_actions(np.array(policy_token))
+                # The first token is meaningless
+                action_policy = action_policy[1:]
+                action_policy = action_policy[:7]
+                # assert len(action_policy) == 7
+                if len(action_policy) != 7:
+                    action_policy = [0] * 7
+                policies_num.append(action_policy.tolist())
+
+        except:
+            policies_num = [[0] * 7]
+            remain_text = text
+
+        return policies_num, remain_text
+
+    def evaluate_single(self, ground_truth, prediction, verbose=False):
+        gt_policies, ground_truth = self.extract_action_policies(ground_truth)
+        pred_policies, prediction = self.extract_action_policies(prediction)
+
+        _, pred_movement = self.extract_movement_plan(prediction)
+        _, gt_movement = self.extract_movement_plan(ground_truth)
+
+        dist, relative_dist, _ = self.compare_movement(label_pos=gt_movement, pred_pos=pred_movement)
+
+        # pred_2d = self.extract_2d_coordinates(prediction)
+        # gt_2d = self.extract_2d_coordinates(ground_truth)
+
+        next_state_score = 0
+
+        acc = self.compare_policy(label_pol=gt_policies, pred_pol=pred_policies)
+
+        return next_state_score, acc, dist, relative_dist, pred_policies, gt_policies
+
+    def evaluate_batch(self, batch_gt, batch_pred, verbose=False):
+        state_acc_ls = []
+        action_acc_ls = []
+        L1_loss_ls = []
+        relative_L1_loss_ls = []
+        pred_policies_ls = []
+        gt_policies_ls = []
+        for i in range(len(batch_gt)):
+            ground_truth = batch_gt[i]
+            prediction = batch_pred[i]
+            next_state_score, action_policy_score, L1_dist, relative_L1_dist, pred_policies, gt_policies = (
+                self.evaluate_single(ground_truth, prediction)
+            )
+            state_acc_ls.append(next_state_score)
+            action_acc_ls.append(action_policy_score)
+            L1_loss_ls.append(L1_dist)
+            relative_L1_loss_ls.append(relative_L1_dist)
+            pred_policies_ls.append(pred_policies)
+            gt_policies_ls.append(gt_policies)
+            if verbose:
+                print(f"Ground Truth:\n\n {ground_truth}")
+                print()
+                print(f"prediction:\n\n {prediction}")
+                print()
+                print(f"Ground Truth Policies:\n\n {gt_policies}")
+                print(f"prediction policies:\n\n {pred_policies}")
+                print("*" * 40)
+
+        return state_acc_ls, action_acc_ls, L1_loss_ls, relative_L1_loss_ls, pred_policies_ls, gt_policies_ls
+
+
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf", model_max_length=2048, padding_side="right")
+action_tokenizer = ActionTokenizer(tokenizer)
+solver = Solver(action_tokenizer)
+