dgenerate-ultralytics-headless 8.3.236__py3-none-any.whl → 8.3.237__py3-none-any.whl

ultralytics/models/sam/sam3/text_encoder_ve.py

@@ -0,0 +1,307 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+from __future__ import annotations
+
+from collections import OrderedDict
+from typing import Callable
+
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+
+from .model_misc import LayerScale
+
+
+class ResidualAttentionBlock(nn.Module):
+    """Transformer block with multi-head attention, layer normalization, and MLP feed-forward network."""
+
+    def __init__(
+        self,
+        d_model: int,
+        n_head: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: float | None = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+    ):
+        """Initialize residual attention block with configurable dimensions and normalization."""
+        super().__init__()
+        # Attention
+        self.attn = nn.MultiheadAttention(d_model, n_head, batch_first=True)
+
+        # LayerNorm, LayerScale
+        self.ln_1 = norm_layer(d_model)
+        self.ln_2 = norm_layer(d_model)
+
+        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
+
+        # MLP
+        mlp_width = int(d_model * mlp_ratio)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, mlp_width)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(mlp_width, d_model)),
+                ]
+            )
+        )
+
+    def attention(
+        self, q_x: torch.Tensor, k_x: torch.Tensor = None, v_x: torch.Tensor = None, attn_mask: torch.Tensor = None
+    ) -> torch.Tensor:
+        """Compute multi-head attention with optional cross-attention support and masking."""
+        k_x = k_x if k_x is not None else q_x
+        v_x = v_x if v_x is not None else q_x
+        if attn_mask is not None:
+            # Leave boolean masks as is
+            if not attn_mask.dtype == torch.bool:
+                attn_mask = attn_mask.to(q_x.dtype)
+
+        return self.attn(q_x, k_x, v_x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(
+        self, q_x: torch.Tensor, k_x: torch.Tensor = None, v_x: torch.Tensor = None, attn_mask: torch.Tensor = None
+    ) -> torch.Tensor:
+        """Apply residual attention with layer normalization and MLP, supporting optional cross-attention."""
+        k_x = self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
+        v_x = self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None
+        x = q_x + self.ls_1(self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask))
+        x = x + self.ls_2(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Transformer(nn.Module):
+    """Stack of residual attention blocks forming a transformer encoder with optional gradient checkpointing."""
+
+    def __init__(
+        self,
+        width: int,
+        layers: int,
+        heads: int,
+        mlp_ratio: float = 4.0,
+        ls_init_value: float | None = None,
+        act_layer: Callable[[], nn.Module] = nn.GELU,
+        norm_layer: Callable[[int], nn.Module] = nn.LayerNorm,
+        compile_mode: str | None = None,
+        use_act_checkpoint: bool = False,
+    ):
+        """Initialize transformer with configurable depth, width, and optional compilation/checkpointing."""
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.grad_checkpointing = use_act_checkpoint
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width,
+                    heads,
+                    mlp_ratio,
+                    ls_init_value=ls_init_value,
+                    act_layer=act_layer,
+                    norm_layer=norm_layer,
+                )
+                for _ in range(layers)
+            ]
+        )
+
+        if compile_mode is not None:
+            self.forward = torch.compile(self.forward, mode=compile_mode, fullgraph=True)
+            if self.grad_checkpointing:
+                torch._dynamo.config.optimize_ddp = False
+
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor = None) -> torch.Tensor:
+        """Process input through all transformer blocks with optional gradient checkpointing during training."""
+        for _, r in enumerate(self.resblocks):
+            if self.grad_checkpointing and not torch.jit.is_scripting() and self.training:
+                x = checkpoint(r, x, None, None, attn_mask, use_reentrant=False)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+
+def text_global_pool(
+    x: torch.Tensor, text: torch.Tensor = None, pool_type: str = "argmax"
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Extract pooled representation and tokens from text embeddings using specified pooling strategy
+    (first/last/argmax/none).
+    """
+    if pool_type == "first":
+        pooled, tokens = x[:, 0], x[:, 1:]
+    elif pool_type == "last":
+        pooled, tokens = x[:, -1], x[:, :-1]
+    elif pool_type == "argmax":
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        assert text is not None
+        pooled, tokens = x[torch.arange(x.shape[0]), text.argmax(dim=-1)], x
+    else:
+        pooled = tokens = x
+    return pooled, tokens
+
+
+class TextTransformer(nn.Module):
+    """Text transformer encoder with causal masking and flexible pooling strategies."""
+
+    def __init__(
+        self,
+        context_length: int = 77,
+        vocab_size: int = 49408,
+        width: int = 512,
+        heads: int = 8,
+        layers: int = 12,
+        mlp_ratio: float = 4.0,
+        ls_init_value: float | None = None,
+        output_dim: int = 512,
+        no_causal_mask: bool = False,
+        pool_type: str = "none",  # no pooling
+        proj_bias: bool = False,
+        act_layer: Callable = nn.GELU,
+        norm_layer: Callable = nn.LayerNorm,
+        output_tokens: bool = False,
+        use_ln_post: bool = True,
+        compile_mode: str | None = None,
+        use_act_checkpoint: bool = False,
+    ):
+        """Initialize text transformer with embedding layers, transformer blocks, and pooling options."""
+        super().__init__()
+        assert pool_type in ("first", "last", "argmax", "none")
+        self.output_tokens = output_tokens
+        self.num_pos = self.context_length = context_length
+        self.vocab_size = vocab_size
+        self.width = width
+        self.output_dim = output_dim
+        self.heads = heads
+        self.pool_type = pool_type
+
+        self.token_embedding = nn.Embedding(self.vocab_size, width)
+        self.positional_embedding = nn.Parameter(torch.empty(self.num_pos, width))
+        self.transformer = Transformer(
+            width=width,
+            layers=layers,
+            heads=heads,
+            mlp_ratio=mlp_ratio,
+            ls_init_value=ls_init_value,
+            act_layer=act_layer,
+            norm_layer=norm_layer,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.ln_final = norm_layer(width) if use_ln_post else nn.Identity()
+        if no_causal_mask:
+            self.attn_mask = None
+        else:
+            self.register_buffer("attn_mask", self.build_causal_mask(), persistent=False)
+        if proj_bias:
+            self.text_projection = nn.Linear(width, output_dim)
+        else:
+            self.text_projection = nn.Parameter(torch.empty(width, output_dim))
+
+    def build_causal_mask(self) -> torch.Tensor:
+        """Create a causal attention mask to prevent attention to future tokens."""
+        # lazily create causal attention mask, with full attention between the tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.num_pos, self.num_pos)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def forward(self, text: torch.Tensor) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
+        """Forward pass through the text transformer, returning pooled output and optionally token embeddings."""
+        seq_len = text.shape[1]
+        x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+        attn_mask = self.attn_mask
+        if attn_mask is not None:
+            attn_mask = attn_mask[:seq_len, :seq_len]
+
+        x = x + self.positional_embedding[:seq_len]
+        x = self.transformer(x, attn_mask=attn_mask)
+
+        x = self.ln_final(x)
+        pooled, tokens = text_global_pool(x, text, pool_type=self.pool_type)
+        if self.text_projection is not None:
+            if isinstance(self.text_projection, nn.Linear):
+                pooled = self.text_projection(pooled)
+            else:
+                pooled = pooled @ self.text_projection
+        if self.output_tokens:
+            return pooled, tokens
+        return pooled
+
+
+class VETextEncoder(nn.Module):
+    """Text encoder for Vision Encoder (VE) models, combining a text transformer and a linear resizer."""
+
+    def __init__(
+        self,
+        d_model: int,
+        tokenizer: Callable,
+        width: int = 1024,
+        heads: int = 16,
+        layers: int = 24,
+        context_length: int = 32,
+        vocab_size: int = 49408,
+        use_ln_post: bool = True,
+        compile_mode: str | None = None,
+        use_act_checkpoint: bool = True,
+    ):
+        """Initialize VE text encoder with a text transformer and a linear resizer to match decoder dimensions."""
+        super().__init__()
+        self.context_length = context_length
+        self.use_ln_post = use_ln_post
+        self.tokenizer = tokenizer
+
+        self.encoder = TextTransformer(
+            context_length=self.context_length,
+            vocab_size=vocab_size,
+            width=width,
+            heads=heads,
+            layers=layers,
+            # we want the tokens, not just the pooled output
+            output_tokens=True,
+            use_ln_post=use_ln_post,
+            compile_mode=compile_mode,
+            use_act_checkpoint=use_act_checkpoint,
+        )
+        self.resizer = nn.Linear(self.encoder.width, d_model)
+
+    def forward(
+        self, text: list[str] | tuple[torch.Tensor, torch.Tensor, dict], input_boxes: list | None = None
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Encode text input, either raw strings or pre-encoded tensors, and resize to match decoder dimensions."""
+        if isinstance(text[0], str):
+            # no use case for this
+            assert input_boxes is None or len(input_boxes) == 0, "not supported"
+
+            # Encode the text
+            tokenized = self.tokenizer(text, context_length=self.context_length).to(
+                self.resizer.weight.device
+            )  # [b, seq_len]
+            text_attention_mask = (tokenized != 0).bool()
+
+            # manually embed the tokens
+            inputs_embeds = self.encoder.token_embedding(tokenized)  # [b, seq_len, d=1024]
+            _, text_memory = self.encoder(tokenized)  # [b, seq_len, d=1024]
+
+            assert text_memory.shape[1] == inputs_embeds.shape[1]
+            # Invert attention mask because its the opposite in pytorch transformer
+            text_attention_mask = text_attention_mask.ne(1)
+            # Transpose memory because pytorch's attention expects sequence first
+            text_memory = text_memory.transpose(0, 1)
+            # Resize the encoder hidden states to be of the same d_model as the decoder
+            text_memory_resized = self.resizer(text_memory)
+        else:
+            # The text is already encoded, use as is.
+            text_attention_mask, text_memory_resized, tokenized = text
+            inputs_embeds = tokenized["inputs_embeds"]
+            assert input_boxes is None or len(input_boxes) == 0, "Can't replace boxes in text if it's already encoded"
+
+        # Note that the input_embeds are returned in pytorch's convention (sequence first)
+        return (
+            text_attention_mask,
+            text_memory_resized,
+            inputs_embeds.transpose(0, 1),
+        )

ultralytics/models/sam/sam3/tokenizer_ve.py

@@ -0,0 +1,242 @@
+# Ultralytics 🚀 AGPL-3.0 License - https://ultralytics.com/license
+
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+"""
+Text Tokenizer.
+
+Copied and lightly adapted from VE repo, which in turn copied
+from open_clip and openAI CLIP.
+"""
+
+from __future__ import annotations
+
+import gzip
+import html
+import io
+import os
+import string
+from functools import lru_cache
+
+import ftfy
+import regex as re
+import torch
+from iopath.common.file_io import g_pathmgr
+
+
+@lru_cache
+def bytes_to_unicode():
+    """Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode
+    strings. This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. When
+    you're at something like a 10B token dataset you end up needing around 5K for decent coverage. This is a
+    significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup tables between utf-8
+    bytes and unicode strings. And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length
+    strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    """Basic text cleaning: fix unicode and unescape HTML entities."""
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    """Remove redundant whitespace."""
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+def _clean_canonicalize(x):
+    """Clean text and canonicalize it."""
+    # basic, remove whitespace, remove punctuation, lower case
+    return canonicalize_text(basic_clean(x))
+
+
+def _clean_lower(x):
+    """Clean text and return lowercase."""
+    # basic, remove whitespace, lower case
+    return whitespace_clean(basic_clean(x)).lower()
+
+
+def _clean_whitespace(x):
+    """Clean text and remove redundant whitespace."""
+    # basic, remove whitespace
+    return whitespace_clean(basic_clean(x))
+
+
+def get_clean_fn(type: str):
+    """Get text cleaning function by name."""
+    if type == "canonicalize":
+        return _clean_canonicalize
+    elif type == "lower":
+        return _clean_lower
+    elif type == "whitespace":
+        return _clean_whitespace
+    else:
+        assert False, f"Invalid clean function ({type})."
+
+
+def canonicalize_text(text, *, keep_punctuation_exact_string=None):
+    """Returns canonicalized `text` (lowercase and punctuation removed). From:
+    https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf08d3388b07671616d3d/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94.
+
+    Args:
+        text: string to be canonicalized.
+        keep_punctuation_exact_string: If provided, then this exact string kept. For example providing '{}' will keep
+            any occurrences of '{}' (but will still remove '{' and '}' that appear separately).
+    """
+    text = text.replace("_", " ")
+    if keep_punctuation_exact_string:
+        text = keep_punctuation_exact_string.join(
+            part.translate(str.maketrans("", "", string.punctuation))
+            for part in text.split(keep_punctuation_exact_string)
+        )
+    else:
+        text = text.translate(str.maketrans("", "", string.punctuation))
+    text = text.lower()
+    text = re.sub(r"\s+", " ", text)
+    return text.strip()
+
+
+class SimpleTokenizer:
+    """A simple tokenizer for text inputs."""
+
+    def __init__(
+        self,
+        bpe_path: str | os.PathLike,
+        additional_special_tokens: list[str] | None = None,
+        context_length: int = 77,
+        clean: str = "lower",
+    ):
+        """The tokenizer for text inputs."""
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        with g_pathmgr.open(bpe_path, "rb") as fh:
+            bpe_bytes = io.BytesIO(fh.read())
+            merges = gzip.open(bpe_bytes).read().decode("utf-8").split("\n")
+        # merges = gzip.open(bpe_path).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        special_tokens = ["<start_of_text>", "<end_of_text>"]
+        if additional_special_tokens:
+            special_tokens += additional_special_tokens
+        vocab.extend(special_tokens)
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {t: t for t in special_tokens}
+        special = "|".join(special_tokens)
+        self.pat = re.compile(
+            special + r"""|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+        self.vocab_size = len(self.encoder)
+        self.all_special_ids = [self.encoder[t] for t in special_tokens]
+        self.sot_token_id = self.all_special_ids[0]
+        self.eot_token_id = self.all_special_ids[1]
+        self.context_length = context_length
+        self.clean_fn = get_clean_fn(clean)
+
+    def bpe(self, token):
+        """Byte Pair Encoding."""
+        if token in self.cache:
+            return self.cache[token]
+        word = (*tuple(token[:-1]), token[-1] + "</w>")
+        pairs = get_pairs(word)
+        if not pairs:
+            return token + "</w>"
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except Exception:
+                    new_word.extend(word[i:])
+                    break
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        """Encode text to a sequence of BPE tokens."""
+        bpe_tokens = []
+        text = self.clean_fn(text)
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    def decode(self, tokens):
+        """Decodes a sequence of tokens back into a text string."""
+        text = "".join([self.decoder[token] for token in tokens])
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors="replace").replace("</w>", " ")
+        return text
+
+    def __call__(self, texts: str | list[str], context_length: int | None = None) -> torch.LongTensor:
+        """Returns the tokenized representation of given input string(s) Parameters. ---------- texts : Union[str,
+        list[str]] An input string or a list of input strings to tokenize context_length : int The context
+        length to use; all CLIP models use 77 as the context length.
+
+        Returns:
+            -------: A two-dimensional tensor containing the resulting tokens, shape = [number of input strings,
+                context_length]
+        """
+        if isinstance(texts, str):
+            texts = [texts]
+        context_length = context_length or self.context_length
+        assert context_length, "Please set a valid context length"
+        all_tokens = [[self.sot_token_id, *self.encode(text), self.eot_token_id] for text in texts]
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+        for i, tokens in enumerate(all_tokens):
+            if len(tokens) > context_length:
+                tokens = tokens[:context_length]  # Truncate
+                tokens[-1] = self.eot_token_id
+            result[i, : len(tokens)] = torch.tensor(tokens)
+        return result