Skip to content

Model

MassSpectrumTransFusion(cfg, max_transcript_len=200)

Bases: TransFusion

Diffusion reconstruction model conditioned on mass spectra.

Source code in instanovo/diffusion/model.py 
 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
def __init__(self, cfg: MassSpectrumModelConfig, max_transcript_len: int = 200) -> None:
    super().__init__(cfg, max_transcript_len)
    layers = []
    for i in range(cfg.layers):
        add_cond_cross_attn = i in list(self.cfg.cond_cross_attn_layers)
        layer = MassSpectrumTransformer(
            self.cfg.dim,
            self.cfg.t_emb_dim,
            self.cfg.cond_emb_dim,
            self.cfg.nheads,
            add_cond_seq=add_cond_cross_attn,
            dropout=self.cfg.dropout,
            use_wavlm_attn=cfg.attention_type == "wavlm" and not add_cond_cross_attn,
            wavlm_num_bucket=cfg.wavlm_num_bucket,
            wavlm_max_dist=cfg.wavlm_max_dist,
            has_rel_attn_bias=(cfg.attention_type == "wavlm" and i == 1),
        )
        # add relative attn bias at i=1 as that is first attn where we do not use
        # cross attention.
        layers.append(layer)
    self.layers = nn.ModuleList(layers)

    self.conditioning_pos_emb = None
    if cfg.localised_attn:
        self.encoder = LocalisedSpectrumEncoder(
            dim_model=cfg.dim,
            n_head=cfg.nheads,
            dim_feedforward=cfg.dim,
            n_layers=cfg.layers,
            dropout=cfg.dropout,
            window_size=cfg.window_size,
            mass_encoding=cfg.mass_encoding,
        )
    else:
        self.encoder = CustomSpectrumEncoder(
            dim_model=cfg.dim,
            n_head=cfg.nheads,
            dim_feedforward=cfg.dim_feedforward,
            n_layers=cfg.layers,
            dropout=cfg.dropout,
            mass_encoding=cfg.mass_encoding,
        )

    # precursor embedding
    self.charge_encoder = torch.nn.Embedding(cfg.max_charge, cfg.dim)
    self.mass_encoder = MassEncoder(cfg.dim)

    self.cache_spectra = None
    self.cache_cond_emb = None
    self.cache_cond_padding_mask = None

forward(x, t, spectra, spectra_padding_mask, precursors, x_padding_mask=None)

Transformer with conditioning cross attention.

  • x: (bs, seq_len) long tensor of character indices or (bs, seq_len, vocab_size) if cfg.diffusion_type == 'continuous'
  • t: (bs, ) long tensor of timestep indices
  • cond_emb: (bs, seq_len2, cond_emb_dim) if using wavlm encoder, else (bs, T)
  • x_padding_mask: (bs, seq_len) if using wavlm encoder, else (bs, T)
  • cond_padding_mask: (bs, seq_len2)

Returns logits (bs, seq_len, vocab_size)

Source code in instanovo/diffusion/model.py 
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
def forward(
    self,
    x: Tensor,
    t: Tensor,
    spectra: Tensor,
    spectra_padding_mask: Tensor,
    precursors: Tensor,
    x_padding_mask: Tensor | None = None,
) -> Tensor:
    """Transformer with conditioning cross attention.

    - `x`: (bs, seq_len) long tensor of character indices
        or (bs, seq_len, vocab_size) if cfg.diffusion_type == 'continuous'
    - `t`: (bs, ) long tensor of timestep indices
    - `cond_emb`: (bs, seq_len2, cond_emb_dim) if using wavlm encoder, else (bs, T)
    - `x_padding_mask`: (bs, seq_len) if using wavlm encoder, else (bs, T)
    - `cond_padding_mask`: (bs, seq_len2)

    Returns logits (bs, seq_len, vocab_size)
    """
    # 1. Base: character, timestep embeddings and zeroing
    bs = x.shape[0]
    x = self.char_embedding(x)  # (bs, seq_len, dim)

    if self.cfg.pos_encoding == "relative":
        x = self.pos_embedding(x)
    else:
        pos_emb = self.pos_embedding.weight[None].expand(
            bs, -1, -1
        )  # (seq_len, dim) --> (bs, seq_len, dim)
        x = x + pos_emb

    t_emb = timestep_embedding(
        t, self.cfg.t_emb_dim, self.cfg.t_emb_max_period, dtype=spectra.dtype
    )  # (bs, t_dim)
    # 2. Classifier-free guidance: with prob cfg.drop_cond_prob, zero out and drop conditional probability
    if self.training:
        zero_cond_inds = torch.rand_like(t, dtype=spectra.dtype) < self.cfg.drop_cond_prob
    else:
        # never randomly zero when in eval mode
        zero_cond_inds = torch.zeros_like(t, dtype=torch.bool)
        if spectra_padding_mask.all():
            # BUT, if all cond information is padded then we are obviously doing unconditional synthesis,
            # so, force zero_cond_inds to be all ones
            zero_cond_inds = ~zero_cond_inds

    # 3. DENOVO calculate spectrum embedding here
    if self.training:
        cond_emb, cond_padding_mask = self.encoder(spectra, spectra_padding_mask)
    else:
        if self.cache_spectra is not None and torch.equal(self.cache_spectra, spectra):
            cond_emb, cond_padding_mask = self.cache_cond_emb, self.cache_cond_padding_mask
        else:
            cond_emb, cond_padding_mask = self.encoder(spectra, spectra_padding_mask)
            self.cache_spectra = spectra
            self.cache_cond_emb = cond_emb
            self.cache_cond_padding_mask = cond_padding_mask

    # set mask for these conditional entries to true everywhere (i.e. mask them out)
    masses = self.mass_encoder(precursors[:, None, [0]])
    charges = self.charge_encoder(precursors[:, 1].int() - 1)
    precursor_emb = masses + charges[:, None, :]

    cond_padding_mask[zero_cond_inds] = True
    cond_emb[zero_cond_inds] = 0

    # 4. Iterate through layers
    pos_bias = None
    for layer in self.layers:
        x, pos_bias = layer(
            x,
            t_emb,
            precursor_emb,
            cond_emb,
            x_padding_mask,
            cond_padding_mask,
            pos_bias=pos_bias,
        )
    # 5. Pass through head to get logits
    x = self.head(x)  # (bs, seq_len, vocab size)

    return x

MassSpectrumTransformer

Bases: Pogfuse

A transformer model specialised for encoding mass spectra.

forward(x, t_emb, precursor_emb, cond_emb=None, x_padding_mask=None, cond_padding_mask=None, pos_bias=None)

Compute encodings with the model.

Forward with x (bs, seq_len, dim), summing t_emb (bs, dim) before the transformer layer, and appending conditioning_emb (bs, seq_len2, dim) to the key/value pairs of the attention. Also pooled_conv_emb (bs, dim) is summed with the timestep embeddings

Optionally specify key/value padding for input x with x_padding_mask (bs, seq_len), and optionally specify key/value padding mask for conditional embedding with cond_padding_mask (bs, seq_len2). By default no padding is used. Good idea to use cond padding but not x padding.

pos_bias is positional bias for wavlm-style attention gated relative position bias.

Returns x of same shape (bs, seq_len, dim)

Source code in instanovo/diffusion/model.py 
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
def forward(
    self,
    x: Tensor,
    t_emb: Tensor,
    precursor_emb: Tensor,
    cond_emb: Tensor | None = None,
    x_padding_mask: Tensor | None = None,
    cond_padding_mask: Tensor | None = None,
    pos_bias: Tensor | None = None,
) -> Tensor:
    """Compute encodings with the model.

    Forward with `x` (bs, seq_len, dim), summing `t_emb` (bs, dim) before the transformer layer,
    and appending `conditioning_emb` (bs, seq_len2, dim) to the key/value pairs of the attention.
    Also `pooled_conv_emb` (bs, dim) is summed with the timestep embeddings

    Optionally specify key/value padding for input `x` with `x_padding_mask` (bs, seq_len), and optionally
    specify key/value padding mask for conditional embedding with `cond_padding_mask` (bs, seq_len2).
    By default no padding is used. Good idea to use cond padding but not x padding.

    `pos_bias` is positional bias for wavlm-style attention gated relative position bias.

    Returns `x` of same shape (bs, seq_len, dim)
    """
    # -----------------------
    # 1. Get and add timestep embedding
    t = self.t_layers(t_emb)[:, None]  # (bs, 1, dim)
    p = self.cond_pooled_layers(precursor_emb)  # (bs, 1, dim)
    x += t + p  # (bs, seq_len, dim)
    # -----------------------
    # 2. Get and append conditioning embeddings
    if self.add_cond_seq:
        c = self.cond_layers(cond_emb)  # (bs, seq_len2, dim)
    else:
        c = None
    # -----------------------
    # 3. Do transformer layer
    # -- Self-attention block
    x1, pos_bias = self._sa_block(
        x, c, x_padding_mask=x_padding_mask, c_padding_mask=cond_padding_mask, pos_bias=pos_bias
    )

    # -- Layer-norm with residual connection
    x = self.norm1(x + x1)

    # -- Layer-norm with feedfoward block and residual connection
    x = self.norm2(x + self._ff_block(x))

    return x, pos_bias