File size: 10,216 Bytes
238fe38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
# Copyright 2026 InstaDeep Ltd and CentraleSupélec
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""CellScribe: scFoundation cell encoder + Qwen2.5 conditioned through AdaLN-Zero.

counts -> scFoundation -> pooled embedding [B, 3072]
       -> Linear + LayerNorm -> conditioning vector [B, 896]
prompt -> Qwen2.5-0.5B-Instruct whose layer norms are modulated by the
          conditioning vector -> "The cell type is ..."

Load with ``AutoModel.from_pretrained("InstaDeepAI/CellScribe",
trust_remote_code=True)``.
"""

from typing import Optional, Sequence

import torch
import torch.nn as nn
import transformers
from transformers import PreTrainedModel

from .ada_ln_zero import attach_ada_ln_zero, create_conditional_model_class
from .configuration_cellscribe import CellScribeConfig
from .postprocessing import extract_predicted_cell_type, parse_cell_state_prediction
from .preprocessing import (
    CELL_TYPES_NAME,
    GENE_LIST_NAME,
    GeneMapper,
    ScFoundationPreprocessor,
    load_gene_list,
)
from .prompts import build_in_context_prompt, build_prompt
from .scfoundation import SCFoundationWrapper


class LinearHead(nn.Module):
    def __init__(self, input_size: int, output_size: int) -> None:
        super().__init__()
        self.network = nn.Sequential(nn.Linear(input_size, output_size))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.network(x)


class CellScribe(PreTrainedModel):
    config_class = CellScribeConfig
    # The output embeddings are tied to the input embeddings (transformers 5 maps
    # each tied weight to its source, transformers 4 lists the tied weights).
    _tied_weights_keys = (
        {"llm.lm_head.weight": "llm.model.embed_tokens.weight"}
        if int(transformers.__version__.split(".")[0]) >= 5
        else ["llm.lm_head.weight"]
    )

    def __init__(self, config: CellScribeConfig) -> None:
        super().__init__(config)
        self.scrna_encoder = SCFoundationWrapper(config.scfoundation)

        llm_config = transformers.AutoConfig.for_model(**config.llm)
        text_embedding_dim = llm_config.hidden_size
        self.scrna_projector = LinearHead(
            config.scrnaseq_embedding_dim, text_embedding_dim
        )
        self.conditioning_norm = nn.LayerNorm(
            text_embedding_dim, elementwise_affine=False, eps=1e-6
        )

        base_llm_class = getattr(transformers, llm_config.architectures[0])
        conditional_llm = create_conditional_model_class(base_llm_class)(llm_config)
        self.llm = attach_ada_ln_zero(conditional_llm, text_embedding_dim)

        # Set by from_pretrained (files next to the weights).
        self.gene_list: Optional[list[str]] = None
        self.cell_types: Optional[list[str]] = None
        # Input helpers, kept out of the module tree (and of the state dict).
        self._input_helpers: dict = {}
        self.post_init()

    def _init_weights(self, module: nn.Module) -> None:
        """Every weight comes from the checkpoint: nothing to initialise."""

    def get_input_embeddings(self) -> nn.Module:
        return self.llm.get_input_embeddings()

    def get_output_embeddings(self) -> nn.Module:
        return self.llm.get_output_embeddings()

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
        model = super().from_pretrained(pretrained_model_name_or_path, *args, **kwargs)
        from transformers.utils import cached_file

        hub_kwargs = {
            k: kwargs[k]
            for k in ("cache_dir", "revision", "token", "local_files_only")
            if k in kwargs
        }
        model.gene_list = load_gene_list(
            cached_file(pretrained_model_name_or_path, GENE_LIST_NAME, **hub_kwargs)
        )
        cell_types = cached_file(
            pretrained_model_name_or_path,
            CELL_TYPES_NAME,
            _raise_exceptions_for_missing_entries=False,
            **hub_kwargs,
        )
        model.cell_types = load_gene_list(cell_types) if cell_types else None
        return model

    # ------------------------------------------------------------ inputs

    def preprocess(self, counts, gene_names: Sequence[str]) -> dict:
        """Raw counts ``(cells x genes)`` named by gene symbols -> the
        ``scrnaseq_preprocessed_input`` of :meth:`forward` / :meth:`generate`.

        Genes are mapped onto the model's gene list (``gene_list.txt``): listed
        genes absent from the input are set to zero, other genes are ignored.
        """
        if self.gene_list is None:
            raise ValueError("load the model with from_pretrained (gene list)")
        if not self._input_helpers:
            cfg = self.config.scfoundation
            self._input_helpers["mapper"] = GeneMapper(self.gene_list)
            self._input_helpers["preprocessor"] = ScFoundationPreprocessor(
                num_genes=len(self.gene_list) + 2,  # + 2 sequencing-depth tokens
                pad_token_id=cfg["pad_token_id"],
                target_res_str=cfg["tgthighres"],
            )
        preprocessor = self._input_helpers["preprocessor"].to(self.device)
        mapped = torch.from_numpy(self._input_helpers["mapper"](counts, gene_names))
        return preprocessor(mapped.to(self.device))

    @staticmethod
    def build_prompt(tokenizer, task: str) -> str:
        """Chat prompt of a task: ``"cell_typing"`` or ``"cell_state"``."""
        return build_prompt(tokenizer, task)

    @staticmethod
    def build_in_context_prompt(tokenizer, example_labels: Sequence[str]):
        """In-context prompt token ids and the token span of each example line."""
        return build_in_context_prompt(tokenizer, example_labels)

    # ------------------------------------------------------------ outputs

    def parse_cell_type(self, text: str) -> str:
        """Cell type of a generated cell-typing answer, read against the training
        vocabulary (``cell_types.txt``), as in the paper's evaluation."""
        return extract_predicted_cell_type(text, self.cell_types)

    @staticmethod
    def parse_cell_state(text: str):
        """``(programs, hallmark)`` of a generated cell-state answer: the
        ``(meta-program, score)`` pairs, strongest first, and the dominant
        hallmark."""
        return parse_cell_state_prediction(text)

    def predictor(self, tokenizer):
        """High-level :class:`CellScribePredictor`: raw counts in, parsed cell
        types / cell states out. Puts the model in eval mode."""
        from .inference import CellScribePredictor

        return CellScribePredictor(model=self, tokenizer=tokenizer)

    # ------------------------------------------------------------ model

    def _transform_conditioning(self, embeddings: torch.Tensor) -> torch.Tensor:
        return self.conditioning_norm(self.scrna_projector(embeddings))

    def embed_cells(self, scrnaseq_preprocessed_input: dict) -> torch.Tensor:
        """Preprocessed cells -> conditioning vectors ``[B, hidden_size]``."""
        return self._transform_conditioning(
            self.scrna_encoder(**scrnaseq_preprocessed_input)
        )

    def _get_conditioning(
        self,
        scrnaseq_preprocessed_input: Optional[dict],
        conditioning: Optional[torch.Tensor],
    ) -> torch.Tensor:
        if conditioning is not None:
            return conditioning
        if scrnaseq_preprocessed_input is None:
            raise ValueError(
                "Pass either 'scrnaseq_preprocessed_input' or 'conditioning'."
            )
        return self.embed_cells(scrnaseq_preprocessed_input)

    def forward(
        self,
        input_ids: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        scrnaseq_preprocessed_input: Optional[dict] = None,
        conditioning: Optional[torch.Tensor] = None,
        conditioning_default: Optional[torch.Tensor] = None,
        **kwargs,
    ):
        """Teacher-forced pass.

        Give either the preprocessed cells or a ``conditioning`` computed by
        :meth:`embed_cells`: ``[B, H]`` (one cell per sequence) or ``[B, T, H]``
        (one cell per token, for in-context learning; positions past ``T`` use
        ``conditioning_default``, ``[B, H]``).
        """
        conditioning = self._get_conditioning(scrnaseq_preprocessed_input, conditioning)
        return self.llm(
            input_ids=input_ids,
            attention_mask=attention_mask,
            conditioning_signal=conditioning,
            conditioning_default=conditioning_default,
            **kwargs,
        )

    @torch.no_grad()
    def generate(  # type: ignore[override]
        self,
        input_ids: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        scrnaseq_preprocessed_input: Optional[dict] = None,
        conditioning: Optional[torch.Tensor] = None,
        conditioning_default: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> torch.Tensor:
        """Greedy generation; returns only the generated tokens.

        ``conditioning`` / ``conditioning_default``: as in :meth:`forward`. With a
        per-token ``[B, T, H]`` plan covering the prompt, the generated tokens are
        conditioned by ``conditioning_default`` (the query cell).
        """
        conditioning = self._get_conditioning(scrnaseq_preprocessed_input, conditioning)
        kwargs.setdefault("max_new_tokens", self.config.max_new_tokens)
        output = self.llm.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            conditioning_signal=conditioning,
            conditioning_default=conditioning_default,
            **kwargs,
        )
        return output[:, input_ids.shape[1] :]