Skip to content

HuggingFace

Transformer pretrained models

HuggingFace Transformers

HFExperiment dataclass

Source code in molfeat/trans/pretrained/hf_transformers.py 
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
@dataclass
class HFExperiment:
    model: PreTrainedModel
    tokenizer: Union[PreTrainedTokenizer, PreTrainedTokenizerFast]
    notation: str = "smiles"

    @classmethod
    def save(cls, model: HFExperiment, path: str, clean_up: bool = False):
        """Save a hugging face model to a specific path

        Args:
            model: model to save
            path: path to the folder root where to save the model
            clean_up: whether to clean up the local path after saving
        """
        local_path = tempfile.mkdtemp()
        # we can save both the tokenizer and the model to the same path
        model.model.save_pretrained(local_path)
        model.tokenizer.save_pretrained(local_path)

        # With transformers>=4.35.0, models are by default saved as safetensors.
        # For backwards compatibility, we also save the model as the older pickle-based format.
        model.model.save_pretrained(local_path, safe_serialization=False)

        dm.fs.copy_dir(local_path, path, force=True, progress=True, leave_progress=False)
        logger.info(f"Model saved to {path}")
        # clean up now
        if clean_up:
            mapper = dm.fs.get_mapper(local_path)
            mapper.fs.delete(local_path, recursive=True)
        return path

    @classmethod
    def load(cls, path: str, model_class=None, device: str = "cpu"):
        """Load a model from the given path
        Args:
            path: Path to the model to load
            model_class: optional model class to provide if the model should be loaded with a specific class
            device: the device to load the model on ("cpu" or "cuda")
        """
        if not dm.fs.is_local_path(path):
            local_path = tempfile.mkdtemp()
            dm.fs.copy_dir(path, local_path, force=True, progress=True, leave_progress=False)
        else:
            local_path = path

        if model_class is None:
            model_config = AutoConfig.from_pretrained(local_path)
            architectures = getattr(model_config, "architectures", [])
            if len(architectures) > 0:
                model_class = MODEL_MAPPING._load_attr_from_module(
                    model_config.model_type, architectures[0]
                )
            else:
                model_class = AutoModel
        model = model_class.from_pretrained(local_path).to(device)
        tokenizer = AutoTokenizer.from_pretrained(local_path)
        return cls(model, tokenizer)

load(path, model_class=None, device='cpu') classmethod

Load a model from the given path Args: path: Path to the model to load model_class: optional model class to provide if the model should be loaded with a specific class device: the device to load the model on ("cpu" or "cuda")

Source code in molfeat/trans/pretrained/hf_transformers.py 
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
@classmethod
def load(cls, path: str, model_class=None, device: str = "cpu"):
    """Load a model from the given path
    Args:
        path: Path to the model to load
        model_class: optional model class to provide if the model should be loaded with a specific class
        device: the device to load the model on ("cpu" or "cuda")
    """
    if not dm.fs.is_local_path(path):
        local_path = tempfile.mkdtemp()
        dm.fs.copy_dir(path, local_path, force=True, progress=True, leave_progress=False)
    else:
        local_path = path

    if model_class is None:
        model_config = AutoConfig.from_pretrained(local_path)
        architectures = getattr(model_config, "architectures", [])
        if len(architectures) > 0:
            model_class = MODEL_MAPPING._load_attr_from_module(
                model_config.model_type, architectures[0]
            )
        else:
            model_class = AutoModel
    model = model_class.from_pretrained(local_path).to(device)
    tokenizer = AutoTokenizer.from_pretrained(local_path)
    return cls(model, tokenizer)

save(model, path, clean_up=False) classmethod

Save a hugging face model to a specific path

Parameters:

Name Type Description Default
model HFExperiment

model to save

 required
path str

path to the folder root where to save the model

 required
clean_up bool

whether to clean up the local path after saving

 False
Source code in molfeat/trans/pretrained/hf_transformers.py 
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
@classmethod
def save(cls, model: HFExperiment, path: str, clean_up: bool = False):
    """Save a hugging face model to a specific path

    Args:
        model: model to save
        path: path to the folder root where to save the model
        clean_up: whether to clean up the local path after saving
    """
    local_path = tempfile.mkdtemp()
    # we can save both the tokenizer and the model to the same path
    model.model.save_pretrained(local_path)
    model.tokenizer.save_pretrained(local_path)

    # With transformers>=4.35.0, models are by default saved as safetensors.
    # For backwards compatibility, we also save the model as the older pickle-based format.
    model.model.save_pretrained(local_path, safe_serialization=False)

    dm.fs.copy_dir(local_path, path, force=True, progress=True, leave_progress=False)
    logger.info(f"Model saved to {path}")
    # clean up now
    if clean_up:
        mapper = dm.fs.get_mapper(local_path)
        mapper.fs.delete(local_path, recursive=True)
    return path

HFModel

Bases: PretrainedStoreModel

Transformer model loading model loading

Source code in molfeat/trans/pretrained/hf_transformers.py 
 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
class HFModel(PretrainedStoreModel):
    """Transformer model loading model loading"""

    def __init__(
        self,
        name: str,
        cache_path: Optional[os.PathLike] = None,
        store: Optional[ModelStore] = None,
    ):
        """Model loader initializer

        Args:
            name (str, optional): Name of the model for ada.
            cache_path (os.PathLike, optional): Local cache path for faster loading. This is the cache_path parameter for ADA loading !
        """

        super().__init__(name, cache_path=cache_path, store=store)
        self._model = None

    @classmethod
    def _ensure_local(cls, object_path: Union[str, os.PathLike]):
        """Make sure the input path is a local path otherwise download it

        Args:
            object_path: Path to the object

        """
        if dm.fs.is_local_path(object_path):
            return object_path
        local_path = tempfile.mkdtemp()
        if dm.fs.is_file(object_path):
            local_path = os.path.join(local_path, os.path.basename(object_path))
            dm.fs.copy_file(object_path, local_path)
        else:
            dm.fs.copy_dir(object_path, local_path)
        return local_path

    @classmethod
    def from_pretrained(
        cls,
        model: Union[str, PreTrainedModel],
        tokenizer: Union[str, PreTrainedTokenizer, PreTrainedTokenizerFast],
        model_class=None,
        model_name: Optional[str] = None,
    ):
        """Load model using huggingface pretrained model loader hook

        Args:
            model: Model to load. Can also be the name on the hub or the path to the model
            tokenizer: Tokenizer to load. Can also be the name on the hub or the path to the tokenizer
            model_class: optional model class to provide if the model should be loaded with a specific class
            model_name: optional model name to give to this model.
        """

        # load the model
        if isinstance(model, PreTrainedModel):
            model_obj = model
        else:
            if dm.fs.exists(model):
                model = cls._ensure_local(model)
            if model_class is None:
                model_config = AutoConfig.from_pretrained(model)
                architectures = getattr(model_config, "architectures", [])
                if len(architectures) > 0:
                    model_class = MODEL_MAPPING._load_attr_from_module(
                        model_config.model_type, architectures[0]
                    )
                else:
                    model_class = AutoModel
            model_obj = model_class.from_pretrained(model)

        if isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)):
            tokenizer_obj = tokenizer
        else:
            if dm.fs.exists(tokenizer):
                tokenizer = cls._ensure_local(tokenizer)
            tokenizer_obj = AutoTokenizer.from_pretrained(tokenizer)
        name = model_name or f"hf_model_{uuid.uuid4().hex[:8]}"
        model = HFModel(name=name, store=ModelStore())
        model._model = HFExperiment(model=model_obj, tokenizer=tokenizer_obj)
        return model

    @classmethod
    def register_pretrained(
        cls,
        model: Union[str, PreTrainedModel],
        tokenizer: Union[str, PreTrainedTokenizer, PreTrainedTokenizerFast],
        model_card: ModelInfo,
        model_class=None,
    ):
        """Register a pretrained huggingface model to the model store
        Args:
            model: Model to load. Can also be the name on the hub or the path to the model
            tokenizer: Tokenizer to load. Can also be the name on the hub or the path to the tokenizer
            model_class: optional model class to provide if the model should be loaded with a specific class
            model_card: optional model card to provide for registering this model
        """
        model = cls.from_pretrained(model, tokenizer, model_class, model_name=model_card.name)
        model.store.register(model_card, model._model, save_fn=HFExperiment.save)
        return model

    def get_notation(self, default_notation: Optional[str] = None):
        """Get the notation of the model"""
        notation = default_notation
        try:
            modelcard = self.store.search(name=self.name)[0]
            notation = modelcard.inputs
        except Exception:
            pass
        return notation

    def load(self):
        """Load Transformer Pretrained featurizer model"""
        if self._model is not None:
            return self._model
        download_output_dir = self._artifact_load(
            name=self.name, download_path=self.cache_path, store=self.store
        )
        model_path = dm.fs.join(download_output_dir, self.store.MODEL_PATH_NAME)
        self._model = HFExperiment.load(model_path)
        return self._model

__init__(name, cache_path=None, store=None)

Model loader initializer

Parameters:

Name Type Description Default
name str

Name of the model for ada.

 required
cache_path PathLike

Local cache path for faster loading. This is the cache_path parameter for ADA loading !

 None
Source code in molfeat/trans/pretrained/hf_transformers.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
def __init__(
    self,
    name: str,
    cache_path: Optional[os.PathLike] = None,
    store: Optional[ModelStore] = None,
):
    """Model loader initializer

    Args:
        name (str, optional): Name of the model for ada.
        cache_path (os.PathLike, optional): Local cache path for faster loading. This is the cache_path parameter for ADA loading !
    """

    super().__init__(name, cache_path=cache_path, store=store)
    self._model = None

from_pretrained(model, tokenizer, model_class=None, model_name=None) classmethod

Load model using huggingface pretrained model loader hook

Parameters:

Name Type Description Default
model Union[str, PreTrainedModel]

Model to load. Can also be the name on the hub or the path to the model

 required
tokenizer Union[str, PreTrainedTokenizer, PreTrainedTokenizerFast]

Tokenizer to load. Can also be the name on the hub or the path to the tokenizer

 required
model_class

optional model class to provide if the model should be loaded with a specific class

 None
model_name Optional[str]

optional model name to give to this model.

 None
Source code in molfeat/trans/pretrained/hf_transformers.py 
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
@classmethod
def from_pretrained(
    cls,
    model: Union[str, PreTrainedModel],
    tokenizer: Union[str, PreTrainedTokenizer, PreTrainedTokenizerFast],
    model_class=None,
    model_name: Optional[str] = None,
):
    """Load model using huggingface pretrained model loader hook

    Args:
        model: Model to load. Can also be the name on the hub or the path to the model
        tokenizer: Tokenizer to load. Can also be the name on the hub or the path to the tokenizer
        model_class: optional model class to provide if the model should be loaded with a specific class
        model_name: optional model name to give to this model.
    """

    # load the model
    if isinstance(model, PreTrainedModel):
        model_obj = model
    else:
        if dm.fs.exists(model):
            model = cls._ensure_local(model)
        if model_class is None:
            model_config = AutoConfig.from_pretrained(model)
            architectures = getattr(model_config, "architectures", [])
            if len(architectures) > 0:
                model_class = MODEL_MAPPING._load_attr_from_module(
                    model_config.model_type, architectures[0]
                )
            else:
                model_class = AutoModel
        model_obj = model_class.from_pretrained(model)

    if isinstance(tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast)):
        tokenizer_obj = tokenizer
    else:
        if dm.fs.exists(tokenizer):
            tokenizer = cls._ensure_local(tokenizer)
        tokenizer_obj = AutoTokenizer.from_pretrained(tokenizer)
    name = model_name or f"hf_model_{uuid.uuid4().hex[:8]}"
    model = HFModel(name=name, store=ModelStore())
    model._model = HFExperiment(model=model_obj, tokenizer=tokenizer_obj)
    return model

get_notation(default_notation=None)

Get the notation of the model

Source code in molfeat/trans/pretrained/hf_transformers.py 
195
196
197
198
199
200
201
202
203
def get_notation(self, default_notation: Optional[str] = None):
    """Get the notation of the model"""
    notation = default_notation
    try:
        modelcard = self.store.search(name=self.name)[0]
        notation = modelcard.inputs
    except Exception:
        pass
    return notation

load()

Load Transformer Pretrained featurizer model

Source code in molfeat/trans/pretrained/hf_transformers.py 
205
206
207
208
209
210
211
212
213
214
def load(self):
    """Load Transformer Pretrained featurizer model"""
    if self._model is not None:
        return self._model
    download_output_dir = self._artifact_load(
        name=self.name, download_path=self.cache_path, store=self.store
    )
    model_path = dm.fs.join(download_output_dir, self.store.MODEL_PATH_NAME)
    self._model = HFExperiment.load(model_path)
    return self._model

register_pretrained(model, tokenizer, model_card, model_class=None) classmethod

Register a pretrained huggingface model to the model store Args: model: Model to load. Can also be the name on the hub or the path to the model tokenizer: Tokenizer to load. Can also be the name on the hub or the path to the tokenizer model_class: optional model class to provide if the model should be loaded with a specific class model_card: optional model card to provide for registering this model

Source code in molfeat/trans/pretrained/hf_transformers.py 
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
@classmethod
def register_pretrained(
    cls,
    model: Union[str, PreTrainedModel],
    tokenizer: Union[str, PreTrainedTokenizer, PreTrainedTokenizerFast],
    model_card: ModelInfo,
    model_class=None,
):
    """Register a pretrained huggingface model to the model store
    Args:
        model: Model to load. Can also be the name on the hub or the path to the model
        tokenizer: Tokenizer to load. Can also be the name on the hub or the path to the tokenizer
        model_class: optional model class to provide if the model should be loaded with a specific class
        model_card: optional model card to provide for registering this model
    """
    model = cls.from_pretrained(model, tokenizer, model_class, model_name=model_card.name)
    model.store.register(model_card, model._model, save_fn=HFExperiment.save)
    return model

PretrainedHFTransformer

Bases: PretrainedMolTransformer

HuggingFace Transformer for feature extraction.

Note

For convenience and consistency, this featurizer only accepts as inputs smiles and molecules, then perform the internal conversion, based on the notation provided.

Source code in molfeat/trans/pretrained/hf_transformers.py 
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
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
class PretrainedHFTransformer(PretrainedMolTransformer):
    """
    HuggingFace Transformer for feature extraction.

    !!! note
        For convenience and consistency, this featurizer only accepts as inputs
        smiles and molecules, then perform the internal conversion, based on the notation provided.
    """

    NEEDS_RANDOM_SEED = ["bert", "roberta"]

    def __init__(
        self,
        kind: Union[str, HFModel] = "ChemBERTa-77M-MLM",
        notation: Optional[str] = "none",
        pooling: str = "mean",
        concat_layers: Union[List[int], int] = -1,
        prefer_encoder: bool = True,
        dtype=np.float32,
        device="cpu",
        max_length: int = 128,
        ignore_padding: bool = True,
        preload: bool = False,
        n_jobs: int = 0,
        random_seed: Optional[int] = None,
        **params,
    ):
        """
        HuggingFace Transformer for featurizer extraction
        The default behaviour of this feature extractor is to return the last hidden state of the encoder
        similar to what is performed by the pipeline 'feature-extraction' in hugging face.

        !!! warning
            For bert models, the default pooling layers is a neural network. Therefore, do not use the default
            Or provide a random seed for reproducibility (in this case pooling will act as random projection to the same manifold)

        !!! note
            The pooling module of this featurizer is accessible through the `_pooling_obj` attribute.

        Args:
            kind: name of the featurizer as available in the model store
            notation: optional line notation to use. Only use if it cannot be found from the model card.
            pooling: type of pooling to use. One of ['default', 'mean', 'max', 'sum', 'clf', ]. The value "default" corresponds to the default litterature pooling for each model type.
                See `molfeat.utils.pooler.get_default_hf_pooler` for more details.
            concat_layers: Layer to concat to get the representation. By default the last hidden layer is returned.
            prefer_encoder: For an encoder-decoder model, prefer the embeddings provided by the encoder.
            dtype: Data type to output
            device: Torch device on which to run the featurizer.
            max_length: Maximum length of the input sequence to consider. Please update this for large sequences
            ignore_padding: Whether to ignore padding in the representation (default: True) to avoid effect of batching
            preload: Whether to preload the model into memory or not
            n_jobs: number of jobs to use
            random_seed: random seed to use for reproducibility whenever a DNN pooler is used (e.g bert/roberta)
        """

        if not requires.check("transformers"):
            raise ValueError(
                "Cannot find transformers and/or tokenizers. It's required for this featurizer !"
            )

        super().__init__(
            dtype=dtype,
            device=device,
            n_jobs=n_jobs,
            **params,
        )
        if concat_layers is None:
            concat_layers = -1
        if not isinstance(concat_layers, list):
            concat_layers = [concat_layers]
        self.concat_layers = concat_layers
        self.max_length = max_length
        self.ignore_padding = ignore_padding
        self._require_mols = False
        self.random_seed = random_seed
        self.preload = preload
        self.pooling = pooling
        self.prefer_encoder = prefer_encoder
        self.device = torch.device(device)
        self._pooling_obj = None
        if isinstance(kind, HFModel):
            self.kind = kind.name
            self.featurizer = kind
        else:
            self.kind = kind
            self.featurizer = HFModel(name=self.kind)
        self.notation = self.featurizer.get_notation(notation) or "none"
        self.converter = SmilesConverter(self.notation)
        if self.preload:
            self._preload()

    def _update_params(self):
        """Update the parameters of this model"""
        # pylint: disable=no-member
        super()._update_params()

        hf_model = HFModel(
            name=self.kind,
        )
        self.featurizer = hf_model.load()
        config = self.featurizer.model.config.to_dict()
        self._pooling_obj = self._pooling_obj = (
            get_default_hgf_pooler(self.pooling, config, random_seed=self.random_seed)
            if self._pooling_obj is None
            else self._pooling_obj
        )

    def _preload(self):
        """Perform preloading of the model from the store"""
        super()._preload()
        self.featurizer.model.to(self.device)
        self.featurizer.max_length = self.max_length

        # we can be confident that the model has been loaded here
        if self._pooling_obj is not None and self.preload:
            return
        config = self.featurizer.model.config.to_dict()
        cur_tokenizer = self.featurizer.tokenizer
        for special_token_id_name in [
            "pad_token_id",
            "bos_token_id",
            "eos_token_id",
            "unk_token_id",
            "sep_token_id",
            "mask_token_id",
        ]:
            token_id = getattr(cur_tokenizer, special_token_id_name)
            if token_id is not None:
                config[special_token_id_name] = token_id

        self._pooling_obj = (
            get_default_hgf_pooler(self.pooling, config, random_seed=self.random_seed)
            if self._pooling_obj is None
            else self._pooling_obj
        )
        # pooling layer is still none, that means we could not fetch it properly
        if self._pooling_obj is None:
            logger.warning(
                "Cannot confidently find the pooling layer and therefore will not apply pooling"
            )

    def _convert(self, inputs: list, **kwargs):
        """Convert the list of molecules to the right format for embedding

        Args:
            inputs: inputs to preprocess

        Returns:
            processed: pre-processed input list
        """
        self._preload()

        if isinstance(inputs, (str, dm.Mol)):
            inputs = [inputs]

        def _to_smiles(x):
            return dm.to_smiles(x) if not isinstance(x, str) else x

        parallel_kwargs = getattr(self, "parallel_kwargs", {})

        if len(inputs) > 1:
            smiles = dm.utils.parallelized(
                _to_smiles,
                inputs,
                n_jobs=self.n_jobs,
                **parallel_kwargs,
            )
            inputs = dm.utils.parallelized(
                self.converter.encode,
                smiles,
                n_jobs=self.n_jobs,
                **parallel_kwargs,
            )
        else:
            inputs = self.converter.encode(_to_smiles(inputs[0]))
        # this check is necessary for some tokenizers
        if isinstance(inputs, str):
            inputs = [inputs]
        encoded = self.featurizer.tokenizer(
            list(inputs),
            truncation=True,
            padding=True,
            max_length=self.max_length,
            return_tensors="pt",
        )
        return encoded

    def _embed(self, inputs, **kwargs):
        """
        Perform embedding of inputs using the pretrained model

        Args:
            inputs: smiles or seqs
            kwargs: any additional parameters
        """
        self._preload()

        # Move inputs to the correct device
        inputs = {key: value.to(self.device) for key, value in inputs.items()}

        attention_mask = inputs.get("attention_mask", None)
        if attention_mask is not None and self.ignore_padding:
            attention_mask = attention_mask.unsqueeze(-1).to(self.device)  # B, S, 1
        else:
            attention_mask = None
        with torch.no_grad():
            if (
                isinstance(self.featurizer.model, EncoderDecoderModel)
                or hasattr(self.featurizer.model, "encoder")
            ) and self.prefer_encoder:
                out_dict = self.featurizer.model.encoder(output_hidden_states=True, **inputs)
            else:
                out_dict = self.featurizer.model(output_hidden_states=True, **inputs)
            hidden_state = out_dict["hidden_states"]
            emb_layers = []
            for layer in self.concat_layers:
                emb = hidden_state[layer].detach()  # B, S, D
                emb = self._pooling_obj(
                    emb,
                    inputs["input_ids"],
                    mask=attention_mask,
                    ignore_padding=self.ignore_padding,
                )
                emb_layers.append(emb)
            emb = torch.cat(emb_layers, dim=1)
        return emb.cpu().numpy()  # Move the final tensor to CPU before converting to numpy array

    def set_max_length(self, max_length: int):
        """Set the maximum length for this featurizer"""
        self.max_length = max_length
        self._preload()

__init__(kind='ChemBERTa-77M-MLM', notation='none', pooling='mean', concat_layers=-1, prefer_encoder=True, dtype=np.float32, device='cpu', max_length=128, ignore_padding=True, preload=False, n_jobs=0, random_seed=None, **params)

HuggingFace Transformer for featurizer extraction The default behaviour of this feature extractor is to return the last hidden state of the encoder similar to what is performed by the pipeline 'feature-extraction' in hugging face.

Warning

For bert models, the default pooling layers is a neural network. Therefore, do not use the default Or provide a random seed for reproducibility (in this case pooling will act as random projection to the same manifold)

Note

The pooling module of this featurizer is accessible through the _pooling_obj attribute.

Parameters:

Name Type Description Default
kind Union[str, HFModel]

name of the featurizer as available in the model store

 'ChemBERTa-77M-MLM'
notation Optional[str]

optional line notation to use. Only use if it cannot be found from the model card.

 'none'
pooling str

type of pooling to use. One of ['default', 'mean', 'max', 'sum', 'clf', ]. The value "default" corresponds to the default litterature pooling for each model type. See molfeat.utils.pooler.get_default_hf_pooler for more details.

 'mean'
concat_layers Union[List[int], int]

Layer to concat to get the representation. By default the last hidden layer is returned.

 -1
prefer_encoder bool

For an encoder-decoder model, prefer the embeddings provided by the encoder.

 True
dtype

Data type to output

 float32
device

Torch device on which to run the featurizer.

 'cpu'
max_length int

Maximum length of the input sequence to consider. Please update this for large sequences

 128
ignore_padding bool

Whether to ignore padding in the representation (default: True) to avoid effect of batching

 True
preload bool

Whether to preload the model into memory or not

 False
n_jobs int

number of jobs to use

 0
random_seed Optional[int]

random seed to use for reproducibility whenever a DNN pooler is used (e.g bert/roberta)

 None
Source code in molfeat/trans/pretrained/hf_transformers.py 
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
def __init__(
    self,
    kind: Union[str, HFModel] = "ChemBERTa-77M-MLM",
    notation: Optional[str] = "none",
    pooling: str = "mean",
    concat_layers: Union[List[int], int] = -1,
    prefer_encoder: bool = True,
    dtype=np.float32,
    device="cpu",
    max_length: int = 128,
    ignore_padding: bool = True,
    preload: bool = False,
    n_jobs: int = 0,
    random_seed: Optional[int] = None,
    **params,
):
    """
    HuggingFace Transformer for featurizer extraction
    The default behaviour of this feature extractor is to return the last hidden state of the encoder
    similar to what is performed by the pipeline 'feature-extraction' in hugging face.

    !!! warning
        For bert models, the default pooling layers is a neural network. Therefore, do not use the default
        Or provide a random seed for reproducibility (in this case pooling will act as random projection to the same manifold)

    !!! note
        The pooling module of this featurizer is accessible through the `_pooling_obj` attribute.

    Args:
        kind: name of the featurizer as available in the model store
        notation: optional line notation to use. Only use if it cannot be found from the model card.
        pooling: type of pooling to use. One of ['default', 'mean', 'max', 'sum', 'clf', ]. The value "default" corresponds to the default litterature pooling for each model type.
            See `molfeat.utils.pooler.get_default_hf_pooler` for more details.
        concat_layers: Layer to concat to get the representation. By default the last hidden layer is returned.
        prefer_encoder: For an encoder-decoder model, prefer the embeddings provided by the encoder.
        dtype: Data type to output
        device: Torch device on which to run the featurizer.
        max_length: Maximum length of the input sequence to consider. Please update this for large sequences
        ignore_padding: Whether to ignore padding in the representation (default: True) to avoid effect of batching
        preload: Whether to preload the model into memory or not
        n_jobs: number of jobs to use
        random_seed: random seed to use for reproducibility whenever a DNN pooler is used (e.g bert/roberta)
    """

    if not requires.check("transformers"):
        raise ValueError(
            "Cannot find transformers and/or tokenizers. It's required for this featurizer !"
        )

    super().__init__(
        dtype=dtype,
        device=device,
        n_jobs=n_jobs,
        **params,
    )
    if concat_layers is None:
        concat_layers = -1
    if not isinstance(concat_layers, list):
        concat_layers = [concat_layers]
    self.concat_layers = concat_layers
    self.max_length = max_length
    self.ignore_padding = ignore_padding
    self._require_mols = False
    self.random_seed = random_seed
    self.preload = preload
    self.pooling = pooling
    self.prefer_encoder = prefer_encoder
    self.device = torch.device(device)
    self._pooling_obj = None
    if isinstance(kind, HFModel):
        self.kind = kind.name
        self.featurizer = kind
    else:
        self.kind = kind
        self.featurizer = HFModel(name=self.kind)
    self.notation = self.featurizer.get_notation(notation) or "none"
    self.converter = SmilesConverter(self.notation)
    if self.preload:
        self._preload()

set_max_length(max_length)

Set the maximum length for this featurizer

Source code in molfeat/trans/pretrained/hf_transformers.py 
444
445
446
447
def set_max_length(self, max_length: int):
    """Set the maximum length for this featurizer"""
    self.max_length = max_length
    self._preload()