Skip to content

Dataset

InMemoryPerTokenValueDataset

Bases: InMemoryProteinDataset

An in-memory dataset of labeled strings, which are tokenized on demand.

Source code in bionemo/esm2/model/finetune/dataset.py 
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
class InMemoryPerTokenValueDataset(InMemoryProteinDataset):
    """An in-memory dataset of labeled strings, which are tokenized on demand."""

    def __init__(
        self,
        sequences: pd.Series,
        labels: pd.Series | None = None,
        tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
        seed: int = np.random.SeedSequence().entropy,  # type: ignore
    ):
        """Initializes a dataset for per-token classification fine-tuning.

        This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
        dataset sequences provided.

        Args:
            sequences (pd.Series): A pandas Series containing protein sequences.
            labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
            tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
            seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
                that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
                generated.
        """
        super().__init__(sequences, labels, tokenizer, seed)
        label_tokenizer = Label2IDTokenizer()
        self.label_tokenizer = label_tokenizer.build_vocab("CHE")
        self.label_cls_eos_id = MLM_LOSS_IGNORE_INDEX

    def transform_label(self, label: str) -> Tensor:
        """Transform the sequence label by tokenizing them.

        This method tokenizes the secondary structure token sequences.

        Args:
            label: secondary structure token sequences to be transformed

        Returns:
            tokenized label
        """
        label_ids = torch.tensor(self.label_tokenizer.text_to_ids(label))

        # # for multi-label classification with BCEWithLogitsLoss
        # tokenized_labels = torch.nn.functional.one_hot(label_ids, num_classes=self.label_tokenizer.vocab_size)
        # cls_eos = torch.full((1, self.label_tokenizer.vocab_size), self.label_cls_eos_id, dtype=tokenized_labels.dtype)

        # for multi-class (mutually exclusive) classification with CrossEntropyLoss
        tokenized_labels = label_ids
        cls_eos = torch.tensor([self.label_cls_eos_id], dtype=tokenized_labels.dtype)

        # add cls / eos label ids with padding value -100 to have the same shape as tokenized_sequence
        labels = torch.cat((cls_eos, tokenized_labels, cls_eos))
        return labels

__init__(sequences, labels=None, tokenizer=tokenizer.get_tokenizer(), seed=np.random.SeedSequence().entropy)

Initializes a dataset for per-token classification fine-tuning.

This is an in-memory dataset that does not apply masking to the sequence. But keeps track of in the dataset sequences provided.

Parameters:

Name Type Description Default
sequences Series

A pandas Series containing protein sequences.

 required
labels Series

A pandas Series containing labels. Defaults to None.

 None
tokenizer BioNeMoESMTokenizer

The tokenizer to use. Defaults to tokenizer.get_tokenizer().

 get_tokenizer()
seed int

Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure that getitem is deterministic, but can be random across different runs. If None, a random seed is generated.

 entropy
Source code in bionemo/esm2/model/finetune/dataset.py 
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
def __init__(
    self,
    sequences: pd.Series,
    labels: pd.Series | None = None,
    tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
    seed: int = np.random.SeedSequence().entropy,  # type: ignore
):
    """Initializes a dataset for per-token classification fine-tuning.

    This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
    dataset sequences provided.

    Args:
        sequences (pd.Series): A pandas Series containing protein sequences.
        labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
        tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
        seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
            that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
            generated.
    """
    super().__init__(sequences, labels, tokenizer, seed)
    label_tokenizer = Label2IDTokenizer()
    self.label_tokenizer = label_tokenizer.build_vocab("CHE")
    self.label_cls_eos_id = MLM_LOSS_IGNORE_INDEX

transform_label(label)

Transform the sequence label by tokenizing them.

This method tokenizes the secondary structure token sequences.

Parameters:

Name Type Description Default
label str

secondary structure token sequences to be transformed

 required

Returns:

Type Description
Tensor

tokenized label
Source code in bionemo/esm2/model/finetune/dataset.py 
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
def transform_label(self, label: str) -> Tensor:
    """Transform the sequence label by tokenizing them.

    This method tokenizes the secondary structure token sequences.

    Args:
        label: secondary structure token sequences to be transformed

    Returns:
        tokenized label
    """
    label_ids = torch.tensor(self.label_tokenizer.text_to_ids(label))

    # # for multi-label classification with BCEWithLogitsLoss
    # tokenized_labels = torch.nn.functional.one_hot(label_ids, num_classes=self.label_tokenizer.vocab_size)
    # cls_eos = torch.full((1, self.label_tokenizer.vocab_size), self.label_cls_eos_id, dtype=tokenized_labels.dtype)

    # for multi-class (mutually exclusive) classification with CrossEntropyLoss
    tokenized_labels = label_ids
    cls_eos = torch.tensor([self.label_cls_eos_id], dtype=tokenized_labels.dtype)

    # add cls / eos label ids with padding value -100 to have the same shape as tokenized_sequence
    labels = torch.cat((cls_eos, tokenized_labels, cls_eos))
    return labels

InMemoryProteinDataset

Bases: Dataset

An in-memory dataset that tokenize strings into BertSample instances.

Source code in bionemo/esm2/model/finetune/dataset.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
 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
class InMemoryProteinDataset(Dataset):
    """An in-memory dataset that tokenize strings into BertSample instances."""

    def __init__(
        self,
        sequences: pd.Series,
        labels: pd.Series | None = None,
        tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
        seed: int = np.random.SeedSequence().entropy,  # type: ignore
    ):
        """Initializes a dataset of protein sequences.

        This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
        dataset sequences provided.

        Args:
            sequences (pd.Series): A pandas Series containing protein sequences.
            labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
            tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
            seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
                that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
                generated.
        """
        self.sequences = sequences
        self.labels = labels

        self.seed = seed
        self._len = len(self.sequences)
        self.tokenizer = tokenizer

    @classmethod
    def from_csv(
        cls, csv_path: str | os.PathLike, tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer()
    ):
        """Class method to create a ProteinDataset instance from a CSV file."""
        df = pd.read_csv(csv_path)

        # Validate presence of required columns
        if "sequences" not in df.columns:
            raise KeyError("The CSV must contain a 'sequences' column.")

        sequences = df["sequences"]
        labels = df["labels"] if "labels" in df.columns else None
        return cls(sequences, labels, tokenizer)

    def __len__(self) -> int:
        """The size of the dataset."""
        return self._len

    def __getitem__(self, index: int) -> BertSample:
        """Obtains the BertSample at the given index."""
        sequence = self.sequences[index]
        tokenized_sequence = self._tokenize(sequence)

        label = tokenized_sequence if self.labels is None else self.transform_label(self.labels.iloc[index])
        # Overall mask for a token being masked in some capacity - either mask token, random token, or left as-is
        loss_mask = ~torch.isin(tokenized_sequence, Tensor(self.tokenizer.all_special_ids))

        return {
            "text": tokenized_sequence,
            "types": torch.zeros_like(tokenized_sequence, dtype=torch.int64),
            "attention_mask": torch.ones_like(tokenized_sequence, dtype=torch.int64),
            "labels": label,
            "loss_mask": loss_mask,
            "is_random": torch.zeros_like(tokenized_sequence, dtype=torch.int64),
        }

    def _tokenize(self, sequence: str) -> Tensor:
        """Tokenize a protein sequence.

        Args:
            sequence: The protein sequence.

        Returns:
            The tokenized sequence.
        """
        tensor = self.tokenizer.encode(sequence, add_special_tokens=True, return_tensors="pt")
        return tensor.flatten()  # type: ignore

    def transform_label(self, label):
        """Transform the label.

        This method should be implemented by subclass if label needs additional transformation.

        Args:
            label: label to be transformed

        Returns:
            transformed_label
        """
        return label

__getitem__(index)

Obtains the BertSample at the given index.

Source code in bionemo/esm2/model/finetune/dataset.py 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
def __getitem__(self, index: int) -> BertSample:
    """Obtains the BertSample at the given index."""
    sequence = self.sequences[index]
    tokenized_sequence = self._tokenize(sequence)

    label = tokenized_sequence if self.labels is None else self.transform_label(self.labels.iloc[index])
    # Overall mask for a token being masked in some capacity - either mask token, random token, or left as-is
    loss_mask = ~torch.isin(tokenized_sequence, Tensor(self.tokenizer.all_special_ids))

    return {
        "text": tokenized_sequence,
        "types": torch.zeros_like(tokenized_sequence, dtype=torch.int64),
        "attention_mask": torch.ones_like(tokenized_sequence, dtype=torch.int64),
        "labels": label,
        "loss_mask": loss_mask,
        "is_random": torch.zeros_like(tokenized_sequence, dtype=torch.int64),
    }

__init__(sequences, labels=None, tokenizer=tokenizer.get_tokenizer(), seed=np.random.SeedSequence().entropy)

Initializes a dataset of protein sequences.

This is an in-memory dataset that does not apply masking to the sequence. But keeps track of in the dataset sequences provided.

Parameters:

Name Type Description Default
sequences Series

A pandas Series containing protein sequences.

 required
labels Series

A pandas Series containing labels. Defaults to None.

 None
tokenizer BioNeMoESMTokenizer

The tokenizer to use. Defaults to tokenizer.get_tokenizer().

 get_tokenizer()
seed int

Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure that getitem is deterministic, but can be random across different runs. If None, a random seed is generated.

 entropy
Source code in bionemo/esm2/model/finetune/dataset.py 
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
def __init__(
    self,
    sequences: pd.Series,
    labels: pd.Series | None = None,
    tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
    seed: int = np.random.SeedSequence().entropy,  # type: ignore
):
    """Initializes a dataset of protein sequences.

    This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
    dataset sequences provided.

    Args:
        sequences (pd.Series): A pandas Series containing protein sequences.
        labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
        tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
        seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
            that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
            generated.
    """
    self.sequences = sequences
    self.labels = labels

    self.seed = seed
    self._len = len(self.sequences)
    self.tokenizer = tokenizer

__len__()

The size of the dataset.

Source code in bionemo/esm2/model/finetune/dataset.py 
85
86
87
def __len__(self) -> int:
    """The size of the dataset."""
    return self._len

from_csv(csv_path, tokenizer=tokenizer.get_tokenizer()) classmethod

Class method to create a ProteinDataset instance from a CSV file.

Source code in bionemo/esm2/model/finetune/dataset.py 
70
71
72
73
74
75
76
77
78
79
80
81
82
83
@classmethod
def from_csv(
    cls, csv_path: str | os.PathLike, tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer()
):
    """Class method to create a ProteinDataset instance from a CSV file."""
    df = pd.read_csv(csv_path)

    # Validate presence of required columns
    if "sequences" not in df.columns:
        raise KeyError("The CSV must contain a 'sequences' column.")

    sequences = df["sequences"]
    labels = df["labels"] if "labels" in df.columns else None
    return cls(sequences, labels, tokenizer)

transform_label(label)

Transform the label.

This method should be implemented by subclass if label needs additional transformation.

Parameters:

Name Type Description Default
label

label to be transformed

 required

Returns:

Type Description

transformed_label
Source code in bionemo/esm2/model/finetune/dataset.py 
119
120
121
122
123
124
125
126
127
128
129
130
def transform_label(self, label):
    """Transform the label.

    This method should be implemented by subclass if label needs additional transformation.

    Args:
        label: label to be transformed

    Returns:
        transformed_label
    """
    return label

InMemorySingleValueDataset

Bases: InMemoryProteinDataset

An in-memory dataset that tokenizes strings into BertSample instances.

Source code in bionemo/esm2/model/finetune/dataset.py 
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
class InMemorySingleValueDataset(InMemoryProteinDataset):
    """An in-memory dataset that tokenizes strings into BertSample instances."""

    def __init__(
        self,
        sequences: pd.Series,
        labels: pd.Series | None = None,
        tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
        seed: int = np.random.SeedSequence().entropy,  # type: ignore
    ):
        """Initializes a dataset for single-value regression fine-tuning.

        This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
        dataset sequences provided.

        Args:
            sequences (pd.Series): A pandas Series containing protein sequences.
            labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
            tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
            seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
                that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
                generated.
        """
        super().__init__(sequences, labels, tokenizer, seed)

    def transform_label(self, label: float) -> Tensor:
        """Transform the regression label.

        Args:
            label: regression value

        Returns:
            tokenized label
        """
        return torch.tensor([label], dtype=torch.float)

__init__(sequences, labels=None, tokenizer=tokenizer.get_tokenizer(), seed=np.random.SeedSequence().entropy)

Initializes a dataset for single-value regression fine-tuning.

This is an in-memory dataset that does not apply masking to the sequence. But keeps track of in the dataset sequences provided.

Parameters:

Name Type Description Default
sequences Series

A pandas Series containing protein sequences.

 required
labels Series

A pandas Series containing labels. Defaults to None.

 None
tokenizer BioNeMoESMTokenizer

The tokenizer to use. Defaults to tokenizer.get_tokenizer().

 get_tokenizer()
seed int

Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure that getitem is deterministic, but can be random across different runs. If None, a random seed is generated.

 entropy
Source code in bionemo/esm2/model/finetune/dataset.py 
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def __init__(
    self,
    sequences: pd.Series,
    labels: pd.Series | None = None,
    tokenizer: tokenizer.BioNeMoESMTokenizer = tokenizer.get_tokenizer(),
    seed: int = np.random.SeedSequence().entropy,  # type: ignore
):
    """Initializes a dataset for single-value regression fine-tuning.

    This is an in-memory dataset that does not apply masking to the sequence. But keeps track of <mask> in the
    dataset sequences provided.

    Args:
        sequences (pd.Series): A pandas Series containing protein sequences.
        labels (pd.Series, optional): A pandas Series containing labels. Defaults to None.
        tokenizer (tokenizer.BioNeMoESMTokenizer, optional): The tokenizer to use. Defaults to tokenizer.get_tokenizer().
        seed: Random seed for reproducibility. This seed is mixed with the index of the sample to retrieve to ensure
            that __getitem__ is deterministic, but can be random across different runs. If None, a random seed is
            generated.
    """
    super().__init__(sequences, labels, tokenizer, seed)

transform_label(label)

Transform the regression label.

Parameters:

Name Type Description Default
label float

regression value

 required

Returns:

Type Description
Tensor

tokenized label
Source code in bionemo/esm2/model/finetune/dataset.py 
158
159
160
161
162
163
164
165
166
167
def transform_label(self, label: float) -> Tensor:
    """Transform the regression label.

    Args:
        label: regression value

    Returns:
        tokenized label
    """
    return torch.tensor([label], dtype=torch.float)