Upload folder using huggingface_hub

added_tokens.json

@@ -0,0 +1,3 @@
+{
+  "<PAD>": 32000
+}

config.json

@@ -0,0 +1,475 @@
+{
+  "arch_specifier": "no-align+fused-gelu-mlp",
+  "architectures": [
+    "OpenVLAForActionPrediction"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_prismatic.OpenVLAConfig",
+    "AutoModelForVision2Seq": "modeling_prismatic.OpenVLAForActionPrediction"
+  },
+  "hf_llm_id": "meta-llama/Llama-2-7b-hf",
+  "image_resize_strategy": "resize-naive",
+  "image_sizes": [
+    224,
+    224
+  ],
+  "llm_backbone_id": "llama2-7b-pure",
+  "llm_max_length": 2048,
+  "model_type": "openvla",
+  "n_action_bins": 256,
+  "norm_stats": {
+    "libero_spatial": {
+      "action": {
+        "mask": [
+          true,
+          true,
+          true,
+          true,
+          true,
+          true,
+          false
+        ],
+        "max": [
+          0.9375,
+          0.9375,
+          0.9375,
+          0.1971428543329239,
+          0.33642858266830444,
+          0.375,
+          1.0
+        ],
+        "mean": [
+          0.15312479436397552,
+          0.13707277178764343,
+          -0.15526802837848663,
+          -0.005176450591534376,
+          -0.01120874285697937,
+          -0.020194264128804207,
+          0.4578818082809448
+        ],
+        "min": [
+          -0.9375,
+          -0.9375,
+          -0.9375,
+          -0.1875,
+          -0.3675000071525574,
+          -0.36000001430511475,
+          0.0
+        ],
+        "q01": [
+          -0.7454732114076613,
+          -0.6616071462631226,
+          -0.9375,
+          -0.1071428582072258,
+          -0.20678570866584778,
+          -0.1842857152223587,
+          0.0
+        ],
+        "q99": [
+          0.9375,
+          0.8758928775787354,
+          0.9321428537368774,
+          0.1039285734295845,
+          0.17678570747375488,
+          0.14571428298950195,
+          1.0
+        ],
+        "std": [
+          0.41272708773612976,
+          0.34724321961402893,
+          0.50869220495224,
+          0.037266165018081665,
+          0.07244449853897095,
+          0.05762382969260216,
+          0.49827873706817627
+        ]
+      },
+      "num_trajectories": 432,
+      "num_transitions": 52970,
+      "proprio": {
+        "max": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "mean": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "min": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "q01": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "q99": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ],
+        "std": [
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0,
+          0.0
+        ]
+      }
+    }
+  },
+  "output_projector_states": false,
+  "pad_to_multiple_of": 64,
+  "pad_token_id": 32000,
+  "quantization_config": {
+    "backend": null,
+    "batch_size": 1,
+    "bits": 4,
+    "block_name_to_quantize": null,
+    "cache_block_outputs": true,
+    "checkpoint_format": "gptq",
+    "damp_percent": 0.1,
+    "dataset": [
+      "pick up blue pen and put into drawer",
+      "Pick up the pot and place it on the napkin.",
+      "put potato on plate",
+      "move the spoon to the bottom left of the table",
+      "Put the banana on the green cloth",
+      "Move the silver pot so that it's directly in front of the canned goods.",
+      "Place the metal pot on the purple cloth.",
+      "Pick the blue spoon and place it on blue towel",
+      "Put the mushroom into the pot.",
+      "Put the banana to the left hand side of the table towards the bottom.",
+      "Move the egg onto the towel",
+      "Move the pot from edge to corner of the table.",
+      "put sweet potato in pot which is in sink distractors",
+      "Place the orange cloth directly in front of the tall can.",
+      "Place the pot on top of the blue cloth",
+      "Move the orange cloth to the far edge of the table",
+      "flip pot upright which is in sink",
+      "Move the spoon behind the metal pot.",
+      "close microwave",
+      "Place the corn inside of the silver pot.",
+      "Place the spoon to the left of the pot.",
+      "upright basil bottle cardboard fence",
+      "put pear on plate and blueberry in pot or pan in sink",
+      "open large4fbox flaps",
+      "flip pot upright in sink distractors",
+      "Put the strawberry into the metal bowl.",
+      "Move the pineapple closer to the blue spoon against the wall.",
+      "turn lever vertical to front distractors",
+      "move the green cloth to the front edge next to orange",
+      "move red spoon to middle lower part of table",
+      "turn lever vertical to front distractors",
+      "Slide the green rag into the corner near the blue handle.",
+      "open fridge",
+      "put pan in sink",
+      "close fridge",
+      "upright basil bottle cardboard fence",
+      "Put the spatula to the top-left of the cloth.",
+      "wipe pot with sponge",
+      "close fridge",
+      "put green squash in pot or pan",
+      "put spatula on plate sink",
+      "Put the knife on the left edge of the counter.",
+      "put pear in bowl",
+      "put spatula in pan",
+      "put pot in sink",
+      "open microwave",
+      "Place the salt next to the brush above the napkin.",
+      "Put the cheese towards the right upper corner of the table.",
+      "Move the silver pan off of the purple rag and next to the green broccoli.",
+      "Move red pot closer to right side of banana.",
+      "put cauliflower in pot or pan an put pot or pan on stove",
+      "Slide the cloth diagonally to the front of the spoon",
+      "Move the can to the back of the counter",
+      "Move the cucumber to the front right of the table.",
+      "open fridge",
+      "Pull the red rectangle near the edge of the table.",
+      "put pear in bowl cardboardfence",
+      "Place the red pepper on the bottom right corner of the table.",
+      "take sushi off plate",
+      "pour almonds in pot",
+      "open microwave",
+      "Move the pot to the left of the spoon, in front of the microwave.",
+      "Pick up the corn from the steel pan and place it on the table.",
+      "put sweet potato in pot",
+      "fold cloth in half",
+      "put pear in bowl cardboardfence",
+      "put sushi on plate",
+      "Turn the yellow spoon horizontally in front of the microwave.",
+      "move towel up to just left of bowl",
+      "Place wok on top of purple cloth.",
+      "close fridge",
+      "Place the corn to the right of the blue brush.",
+      "Move the strawberry and place on orange towel",
+      "take lid off pot cardboardfence",
+      "turn lever vertical to-front",
+      "put cup from counter to sink",
+      "Pick up the red spoon on to of orange cloth and place it on its left",
+      "put clothes in laundry machine",
+      "Moves the green tool off the napkin to the bottom right.",
+      "move green peper into pot",
+      "Slide the yellow rag forward so that it touches the front edge of the table.",
+      "Move the pot to the inner edge of the table.",
+      "Move the cheese slice in between pot and blue spoon",
+      "put knife in pot cardboard fence",
+      "MOVE THE BLENDER INSIDE THE VESSEL",
+      "Put the orange cloth to the top left of the table.",
+      "move the blue cloth from front to behind on the table",
+      "put banana in pot cardboard fence",
+      "Place the silver pan on the purple cloth.",
+      "put cup from anywhere into sink",
+      "open white1fbox flap",
+      "close white1fbox flap",
+      "take lid off pot or pan",
+      "Put the metal pot in front of the mushroom.",
+      "move the scoop to the far left in front of the pepper",
+      "Take the wok off the blue cloth and move it behind the cloth.",
+      "lift bowl",
+      "Move the bowl to the top right of table",
+      "put banana in pot or pan",
+      "Move the blue towel to the right of the spoon",
+      "take lid off pot cardboardfence",
+      "open microwave",
+      "put corn in pot which is in sink distractors",
+      "put banana on plate",
+      "Place the mushroom in front of the microwave.",
+      "Put the mushroom in the pot.",
+      "put big spoon from basket to tray",
+      "Move the orange cloth to the lower right of the table",
+      "flip pot upright in sink distractors",
+      "Move the purple cloth to the left of the grey toy",
+      "take corn out of bowl sink",
+      "Hold the mushroom which is infront of the oven and drop it under the tiffin box, which is on the right edge of the table.",
+      "take cucumber out of cup",
+      "flip pot upright in sink distractors",
+      "put spoon in pot",
+      "Move the pot directly right of the towel",
+      "flip pot upright in sink distractors",
+      "unzip zipper bag",
+      "Place the spoon on top of the orange cloth.",
+      "take cucumber out of cup",
+      "put pot in sink",
+      "Move the chicken thigh to the top left corner of the table.",
+      "put pepper in pot or pan",
+      "Move the mushroom to the left of the yellow spoon",
+      "take carrot off plate cardboardfence",
+      "Move blue and gray item to the left of the pot.",
+      "Put the spoon to the left of the pot.",
+      "pick ball and put it on the vessel",
+      "Put the blue fork on top of the napkin",
+      "take sushi out of pot cardboard fence",
+      "Move the purple cloth to the far left edge of the table",
+      "zip zipper bag",
+      "take bowl off plate cardboard fence",
+      "take carrot out of pot cardboard fence",
+      "close fridge",
+      "Put the pot in the bottom left hand corner of the table.",
+      "open microwave",
+      "Pick up the spoon and put it on the left hand corner of the table.",
+      "take broccoli out of pan cardboardfence",
+      "Place the brush behind the pot.",
+      "Place the mushroom behind the spatula.",
+      "Move the brush to the middle bottom between burners",
+      "pick up blue pen and put into drawer",
+      "Move the bowl to the left adjacent to the spoon",
+      "Pick up the metal strainer and place it on the left eye.",
+      "Put the spoon under the pot.",
+      "turn lever vertical to front distractors",
+      "Place the pot next to the knife.",
+      "push the blue cloth a few inches back",
+      "Move the cloth to the inner front edge of the table.",
+      "Move the pot over the green cloth",
+      "pick up pan from stove distractors",
+      "Grab the bell pepper outside of the pot and place it towards the bottom of the table to the right side of the yellow towel.",
+      "put eggplant into pot or pan",
+      "Pick the cup from the vessel and place it on the blue towel in the middle of the table",
+      "put sweet potato in pan which is on stove distractors",
+      "Remove the pear from the pot and place it on the right back corner of the counter.",
+      "put knife on cutting board cardboard fence",
+      "wipe pot with sponge",
+      "pour almonds in pot",
+      "open oven",
+      "Move the mushroom in front of the blue cloth",
+      "Move the pot to the left of the brush.",
+      "Move the mushroom and put it in the pot.",
+      "Move the carrot from the pot.",
+      "topple basil bottle cardboard fence",
+      "Put the metal pot on top of the yellow cloth.",
+      "close large4fbox flaps",
+      "end effector reaching knife",
+      "unzip zipper bag",
+      "Move the food item to the lower left side of the table",
+      "Put the red spoon in between the pan and the green cloth",
+      "Place the spatula to the left of the pot.",
+      "put pan on towel",
+      "Move the red spoon so that it's in front of the silver pan, near the edge of the table.",
+      "turn faucet front to left",
+      "Move the purple cloth and place it at the center of the table.",
+      "take carrot off plate cardboardfence",
+      "right pepper shaker",
+      "turn faucet front to left",
+      "Place the container in the pan",
+      "Move the cloth directly between the pot and mushroom.",
+      "put the mushroom in the pot",
+      "end effector reaching pot or pan",
+      "Put orange and green object in silver pan",
+      "Place the blue spoon next to the metal pot",
+      "Put the green spoon beside the metal bowl.",
+      "put eggplant in pot or pan",
+      "Place the spoon on the front right corner of the counter near the bowl.",
+      "pick up glass cup",
+      "NOT MOMENT THE ANY THING",
+      "put cucumber in cup",
+      "Move the white item to the front of the microwave",
+      "Put the fork onto the cloth.",
+      "take bowl off plate",
+      "pick appl from vessel and place it on the right bottom of the table",
+      "Place the silver pot on the blue cloth.",
+      "Place the knife next to the potato.",
+      "put pot or pan on stove and put mushroom in pot or pan",
+      "take carrot out of pot cardboard fence",
+      "turn lever vertical to-front",
+      "Move the red pepper to the front left corner of the table",
+      "take cup off plate",
+      "take bowl off plate",
+      "pick up pan from stove distractors",
+      "Place the baster in the pot.",
+      "Move the green spatula behind the yellow cloth.",
+      "Place the spatula behind the vessel.",
+      "close fridge",
+      "open microwave",
+      "Move the orange object to the left of the pot.",
+      "put lid on pot or pan",
+      "Move the mushroom into the pot.",
+      "Move and place the spoon to the right side of the silver vessel.",
+      "move the fish behind the towel across from the pot",
+      "Put the shaker into the pot.",
+      "Place the green cloth next to the red chili",
+      "put knife on cutting board",
+      "Video format not showing.",
+      "flip pot upright in sink distractors",
+      "pick up scissors and put into drawer",
+      "turn lever vertical to front",
+      "open fridge",
+      "Move the knife under the towel.",
+      "put saltcontainer on plate and ketchupbottle in pot or pan in sink",
+      "PICK UP THE FRUIT AND PUT NEAR PURPLE TOWEL",
+      "Move the pan to the right of the bottle and banana.",
+      "Put the pear to the right of the pot.",
+      "put detergent in sink",
+      "put blueberry on plate and spoon in pot or pan in sink",
+      "Move the pan to the middle left edge of the table.",
+      "Move the hot dog towards the top of the table.",
+      "Pick up the yellow flower and place it behind the pot",
+      "Pick up pot and drop onto the blue towel",
+      "put cup from sink to drying rack",
+      "moved fork to the back of the counter",
+      "Move the silver pot to the center of the table.",
+      "put the Spoon on the towel",
+      "Place the red spoon on middle of the table",
+      "open drawer of box",
+      "Move the brush nearby the lower right corner.",
+      "turn lever vertical to front distractors",
+      "Move pot to the right of the spoon.",
+      "fold cloth in half",
+      "put cup from counter to sink",
+      "move yellow knife to left of blue object",
+      "Pick up the green spatula and place it on the green cloth",
+      "pick up green mug",
+      "put cup from counter to sink",
+      "Put the green spatula in the upper left corner of the table",
+      "put potato on plate",
+      "take sushi out of pot cardboard fence",
+      "put eggplant into pot or pan",
+      "Move the blue cloth to the left of the silver bowl",
+      "put sushi on plate",
+      "Move the spoon between the cans and the mushroom"
+    ],
+    "desc_act": false,
+    "exllama_config": {
+      "version": 1
+    },
+    "group_size": 128,
+    "max_input_length": null,
+    "meta": {
+      "quantizer": [
+        "optimum:1.26.1",
+        "auto_gptq:0.7.1"
+      ]
+    },
+    "model_seqlen": null,
+    "module_name_preceding_first_block": null,
+    "modules_in_block_to_quantize": null,
+    "pad_token_id": null,
+    "quant_method": "gptq",
+    "sym": true,
+    "tokenizer": null,
+    "true_sequential": true,
+    "use_cuda_fp16": false,
+    "use_exllama": true
+  },
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 4096,
+    "initializer_range": 0.02,
+    "intermediate_size": 11008,
+    "max_position_embeddings": 2048,
+    "mlp_bias": false,
+    "model_type": "llama",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 32,
+    "num_key_value_heads": 32,
+    "pad_token_id": 32000,
+    "pretraining_tp": 1,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 10000.0,
+    "torch_dtype": "bfloat16",
+    "use_cache": true,
+    "vocab_size": 32064
+  },
+  "timm_model_ids": [
+    "vit_large_patch14_reg4_dinov2.lvd142m",
+    "vit_so400m_patch14_siglip_224"
+  ],
+  "timm_override_act_layers": [
+    null,
+    null
+  ],
+  "torch_dtype": "float16",
+  "transformers_version": "4.53.1",
+  "use_cache": false,
+  "use_fused_vision_backbone": true,
+  "vision_backbone_id": "dinosiglip-vit-so-224px"
+}

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)

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 32000,
+  "transformers_version": "4.53.1"
+}

model-00001-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09689720229c37f98dd41ffcdf4121c390e7714bf1ed56c840967cf853fa8592
+size 4979533016

model-00002-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:65a6500969c155fcf62b2a6e5c6b32241b940e988ff6d71890747cbbef86615e
+size 520182600

modeling_prismatic.py

@@ -0,0 +1,562 @@
+"""
+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 .configuration_prismatic import OpenVLAConfig, PrismaticConfig
+
+# 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 OpenVLAForActionPrediction(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 .generate() that decodes predicted actions and unnormalizes them."""
+        # If the special empty token ('') does not already appear after the colon (':') token in the prompt
+        # (after "OUT:" or "ASSISTANT:"), insert it to match the inputs seen at training time
+        if not torch.all(input_ids[:, -1] == 29871):
+            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, max_new_tokens=self.get_action_dim(unnorm_key), **kwargs)
+
+        # Extract predicted action tokens and translate into (normalized) continuous actions
+        predicted_action_token_ids = generated_ids[0, -self.get_action_dim(unnorm_key) :].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
+
+    @staticmethod
+    def _check_unnorm_key(norm_stats: Dict[str, Dict[str, Any]], unnorm_key: Optional[str]) -> str:
+        if unnorm_key is None:
+            assert len(norm_stats) == 1, (
+                f"Your model was trained on more than one dataset, "
+                f"please pass a `unnorm_key` from the following options to choose the statistics "
+                f"used for un-normalizing actions: {norm_stats.keys()}"
+            )
+            unnorm_key = next(iter(norm_stats.keys()))
+
+        assert unnorm_key in norm_stats, (
+            f"The `unnorm_key` you chose is not in the set of available dataset 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"]

preprocessor_config.json

@@ -0,0 +1,114 @@
+{
+  "auto_map": {
+    "AutoImageProcessor": "processing_prismatic.PrismaticImageProcessor",
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "image_processor_type": "PrismaticImageProcessor",
+  "image_resize_strategy": "resize-naive",
+  "input_sizes": [
+    [
+      3,
+      224,
+      224
+    ],
+    [
+      3,
+      224,
+      224
+    ]
+  ],
+  "interpolations": [
+    "bicubic",
+    "bicubic"
+  ],
+  "means": [
+    [
+      0.485,
+      0.456,
+      0.406
+    ],
+    [
+      0.5,
+      0.5,
+      0.5
+    ]
+  ],
+  "processor_class": "PrismaticProcessor",
+  "stds": [
+    [
+      0.229,
+      0.224,
+      0.225
+    ],
+    [
+      0.5,
+      0.5,
+      0.5
+    ]
+  ],
+  "tvf_crop_params": [
+    {
+      "output_size": [
+        224,
+        224
+      ]
+    },
+    {
+      "output_size": [
+        224,
+        224
+      ]
+    }
+  ],
+  "tvf_do_letterbox": false,
+  "tvf_letterbox_fill": null,
+  "tvf_normalize_params": [
+    {
+      "inplace": false,
+      "mean": [
+        0.484375,
+        0.455078125,
+        0.40625
+      ],
+      "std": [
+        0.228515625,
+        0.2236328125,
+        0.224609375
+      ]
+    },
+    {
+      "inplace": false,
+      "mean": [
+        0.5,
+        0.5,
+        0.5
+      ],
+      "std": [
+        0.5,
+        0.5,
+        0.5
+      ]
+    }
+  ],
+  "tvf_resize_params": [
+    {
+      "antialias": true,
+      "interpolation": 3,
+      "max_size": null,
+      "size": [
+        224,
+        224
+      ]
+    },
+    {
+      "antialias": true,
+      "interpolation": 3,
+      "max_size": null,
+      "size": [
+        224,
+        224
+      ]
+    }
+  ],
+  "use_fused_vision_backbone": true
+}

processing_prismatic.py

@@ -0,0 +1,257 @@
+"""
+processing_prismatic.py
+
+HuggingFace-style preprocessor definitions for Prismatic VLMs, inheriting from `ProcessorMixin`. Default configuration
+specifies `siglip-224px+7b`.
+"""
+
+from typing import Any, ClassVar, List, Optional, Tuple, Union
+
+import timm.data
+import torch
+import torchvision.transforms.functional as TVF
+from PIL import Image
+from torchvision.transforms import CenterCrop, Compose, Normalize, Resize, ToTensor
+from transformers import PreTrainedTokenizerBase
+from transformers.image_processing_utils import BatchFeature, ImageProcessingMixin
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from transformers.utils import TensorType
+
+
+# === Image Processing ===
+def letterbox_pad_transform(image: Image.Image, padding_fill_value: Tuple[int, int, int]) -> Image.Image:
+    """Given a PIL.Image, pad to square by adding a symmetric border around the height/width."""
+    (w, h), max_wh = image.size, max(image.size)
+    horizontal_pad, vertical_pad = int((max_wh - w) / 2), int((max_wh - h) / 2)
+    padding = (horizontal_pad, vertical_pad, horizontal_pad, vertical_pad)
+
+    return TVF.pad(image, padding, fill=padding_fill_value, padding_mode="constant")
+
+
+class PrismaticImageProcessor(ImageProcessingMixin):
+    model_input_names: ClassVar[List[str]] = ["pixel_values"]
+
+    def __init__(
+        self,
+        use_fused_vision_backbone: bool = False,
+        image_resize_strategy: str = "letterbox",
+        input_sizes: Optional[List[Tuple[int, int, int]]] = None,
+        interpolations: Optional[List[str]] = None,
+        means: Optional[List[Tuple[float, float, float]]] = None,
+        stds: Optional[List[Tuple[float, float, float]]] = None,
+        **kwargs: str,
+    ) -> None:
+        """
+        Initialize a PrismaticImageProcessor as a wrapper around a torchvision transform; this transform will be
+        created by TIMM, and edited to follow our custom `image_resize_strategy` logic.
+
+        @param use_fused_vision_backbone: Boolean indicating single or fused (dual) vision backbone
+        @param image_resize_strategy: Prismatic image resize strategy in < resize-naive | resize-crop | letterbox >
+        @param input_size: [TIMM :: `data_cfg`] Input image size as tuple (channels, width, height)
+        @param interpolation: [TIMM :: `data_cfg`] Interpolation as string (default: "bicubic")
+        @param mean: [TIMM :: `data_cfg`] Normalization mean as float tuple (or two-tuple if `fused_backbone`)
+        @param std: [TIMM :: `data_cfg`] Normalization std as float tuple (or two-tuple if `fused_backbone`)
+        """
+        self.use_fused_vision_backbone = use_fused_vision_backbone
+        self.image_resize_strategy = image_resize_strategy
+
+        # Handle `None` default values
+        input_sizes = [(3, 224, 224)] if input_sizes is None else input_sizes
+        means = [(0.5, 0.5, 0.5)] if means is None else means
+        stds = [(0.5, 0.5, 0.5)] if stds is None else stds
+
+        # TIMM `data_cfg` Parameters
+        self.input_sizes, self.interpolations, self.means, self.stds = input_sizes, interpolations, means, stds
+
+        # Grab torchvision transforms via TIMM =>> need to parse for specific "functional" transform values!
+        self.tvf_resize_params, self.tvf_crop_params, self.tvf_normalize_params = [], [], []
+        self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
+
+        for idx in range(len(input_sizes)):
+            transform = timm.data.create_transform(
+                input_size=self.input_sizes[idx],
+                interpolation=self.interpolations[idx],
+                mean=self.means[idx],
+                std=self.stds[idx],
+                crop_pct=1.0,  # Set to 1.0 to ignore cropping (initial Resize sets `input_size`)
+                crop_mode="center",  # Default crop mode -- no-op when `crop_pct == 1.0`
+                is_training=False,  # No image augmentations when loading the transform!
+            )
+
+            # [Validation] Ensure appropriate transform structure, expected sizes
+            if not (
+                isinstance(transform, Compose)
+                and (len(transform.transforms) == 4)
+                and isinstance(transform.transforms[0], Resize)
+                and isinstance(transform.transforms[1], CenterCrop)
+                and isinstance(transform.transforms[2], ToTensor)
+                and isinstance(transform.transforms[3], Normalize)
+                and (transform.transforms[0].size == self.input_sizes[idx][-1])
+                and (transform.transforms[1].size == self.input_sizes[idx][-2:])
+            ):
+                raise ValueError(f"Unexpected TIMM image transformation structure/sizes: `{transform}`")
+
+            # HF Image Processors *must* be JSON-serializable; as such, cannot have torchvision. as an attribute.
+            #   => Instead, we're going to parse the transform and call "torchvision.transforms.functional" (`tvf`)
+            resize_t, crop_t, norm_t = transform.transforms[0], transform.transforms[1], transform.transforms[3]
+            self.tvf_resize_params.append(
+                {
+                    "size": resize_t.size,
+                    "interpolation": TVF.pil_modes_mapping[resize_t.interpolation],
+                    "max_size": None,
+                    "antialias": True,
+                }
+            )
+            self.tvf_crop_params.append({"output_size": crop_t.size})
+            self.tvf_normalize_params.append(
+                {
+                    "mean": norm_t.mean.float().numpy().tolist(),
+                    "std": norm_t.std.float().numpy().tolist(),
+                    "inplace": False,
+                }
+            )
+            self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
+
+            # Handle Prismatic `image_resize_strategy`
+            if self.image_resize_strategy == "resize-naive":
+                self.tvf_resize_params[idx]["size"] = (resize_t.size, resize_t.size)
+            elif self.image_resize_strategy == "letterbox":
+                self.tvf_do_letterbox, self.tvf_letterbox_fill = True, tuple([int(x * 255) for x in self.means[idx]])
+            elif self.image_resize_strategy == "resize-crop":
+                pass
+            else:
+                raise ValueError(f"Image resize strategy `{self.image_resize_strategy}` is not supported!")
+
+        # Dispatch **kwargs to super()
+        super().__init__(**kwargs)
+
+    def apply_transform(self, img: Image.Image) -> torch.Tensor:
+        """Apply `functional` variant of TIMM's Transform = Compose([Resize -> CenterCrop -> ToTensor -> Normalize])"""
+        if self.tvf_do_letterbox:
+            img = letterbox_pad_transform(img, self.tvf_letterbox_fill)
+
+        # [Contract] Fused Backbones expect "channel-stacked" inputs; we'll unpack on the model side!
+        imgs_t = []
+        for idx in range(len(self.input_sizes)):
+            img_idx = TVF.resize(img, **self.tvf_resize_params[idx])
+            img_idx = TVF.center_crop(img_idx, **self.tvf_crop_params[idx])
+            img_idx_t = TVF.to_tensor(img_idx)
+            img_idx_t = TVF.normalize(img_idx_t, **self.tvf_normalize_params[idx])
+            imgs_t.append(img_idx_t)
+
+        # [Contract] `imgs_t` is a list of Tensors of shape [3, input_size, input_size]; stack along dim = 0
+        img_t = torch.vstack(imgs_t)
+
+        return img_t
+
+    def preprocess(
+        self,
+        images: Union[Image.Image, List[Image.Image]],
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **_: str,
+    ) -> BatchFeature:
+        """
+        Preprocess an image (or batch of images); note that unlike the `transformers :: BaseImageProcessor` we
+        explicitly only handle PIL.Image.Image instances for simplicity.
+
+        @param images: A (batch of) PIL.Image.Image instance(s) to preprocess.
+        @param return_tensors: BatchFeature default Tensor format (e.g., "pt" for torch); if None, returns np.ndarray
+
+        @return: Instance of `transformers :: BatchFeature` with a single key "pixel_values"
+        """
+        if not isinstance(images, list):
+            images = [images]
+
+        # Apply `self.img_transform` to each image (will return list of torch.Tensors); stack into "batched" Tensor
+        pixel_values = torch.stack([self.apply_transform(img.convert("RGB")) for img in images])
+
+        # Return BatchFeature =>> note that for compatibility, constructor expects Dict[str, np.ndarray], so we convert
+        return BatchFeature(data={"pixel_values": pixel_values.float().numpy()}, tensor_type=return_tensors)
+
+    def __call__(self, images: Union[Image.Image, List[Image.Image]], **kwargs) -> BatchFeature:
+        return self.preprocess(images, **kwargs)
+
+
+# === PrismaticProcessor =>> Wraps both ImageProcessor and Tokenizer ===
+#   =>> https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava/processing_llava.py
+class PrismaticProcessor(ProcessorMixin):
+    attributes: ClassVar[List[str]] = ["image_processor", "tokenizer"]
+    image_processor_class: str = "AutoImageProcessor"
+    tokenizer_class: str = "AutoTokenizer"
+
+    def __init__(
+        self,
+        image_processor: Optional[ImageProcessingMixin] = None,
+        tokenizer: Optional[PreTrainedTokenizerBase] = None,
+    ) -> None:
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
+        images: Union[Image.Image, List[Image.Image]],
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Optional[Union[bool, str, TruncationStrategy]] = None,
+        max_length: Optional[int] = None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+    ) -> BatchFeature:
+        """
+        Preprocess a given (batch) of text/images for a Prismatic VLM; forwards text to the underlying LLM's tokenizer,
+        forwards images to PrismaticImageProcessor.
+
+        @param text: The (batch) of text to encode; must be a string or list of strings.
+        @param images: A (batch of) PIL.Image.Image instance(s) to preprocess.
+        @param padding: Sequence padding strategy (if multiple specified) in < True = "longest" | "max_length" | False >
+        @param truncation: Truncation strategy for the output sequences; requires `max_length` to be specified
+        @param max_length: Maximum length (in tokens) to truncate
+        @param return_tensors: Type of return tensors (usually "pt" or TensorType.PYTORCH)
+
+        @return: BatchFeature with keys for `input_ids`, `attention_mask` and `pixel_values`.
+        """
+        pixel_values = self.image_processor(images, return_tensors=return_tensors)["pixel_values"]
+        text_inputs = self.tokenizer(
+            text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
+        )
+
+        # [Validate] Need same number of images and text inputs!
+        if pixel_values.shape[0] != text_inputs.input_ids.shape[0]:
+            raise ValueError("Batch is malformed; expected same number of images and text inputs!")
+
+        return BatchFeature(data={**text_inputs, "pixel_values": pixel_values})
+
+    # === Tokenizer Dispatch Utilities =>> check `PreTrainedTokenizerBase` for documentation ===
+    def batch_decode(
+        self,
+        sequences: Union[List[int], List[List[int]], torch.Tensor, Any],  # `Any` = np.ndarray | tf.Tensor
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        **kwargs: str,
+    ) -> List[str]:
+        return self.tokenizer.batch_decode(
+            sequences=sequences,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    def decode(
+        self,
+        token_ids: Union[int, List[int], torch.Tensor, Any],  # `Any` = np.ndarray | tf.Tensor
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = None,
+        **kwargs: str,
+    ) -> str:
+        return self.tokenizer.decode(
+            token_ids=token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def model_input_names(self) -> List[str]:
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "processor_class": "PrismaticProcessor"
+}

special_tokens_map.json

@@ -0,0 +1,30 @@
+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<PAD>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
+size 499723

tokenizer_config.json

@@ -0,0 +1,55 @@
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32000": {
+      "content": "<PAD>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoProcessor": "processing_prismatic.PrismaticProcessor"
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "extra_special_tokens": {},
+  "legacy": false,
+  "model_max_length": 2048,
+  "pad_token": "<PAD>",
+  "padding_side": "right",
+  "processor_class": "PrismaticProcessor",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}