Merge branch 'sdavis_trn' into 'master'

transformer

See merge request machine-learning/bonito!166

README.md

@@ -1,11 +1,11 @@
 # Bonito
 
-[![PyPI version](https://badge.fury.io/py/ont-bonito.svg)](https://badge.fury.io/py/ont-bonito) 
+[![PyPI version](https://badge.fury.io/py/ont-bonito.svg)](https://badge.fury.io/py/ont-bonito)
 [![py38](https://img.shields.io/badge/python-3.8-brightgreen.svg)](https://img.shields.io/badge/python-3.8-brightgreen.svg)
 [![py39](https://img.shields.io/badge/python-3.9-brightgreen.svg)](https://img.shields.io/badge/python-3.9-brightgreen.svg)
 [![py310](https://img.shields.io/badge/python-3.10-brightgreen.svg)](https://img.shields.io/badge/python-3.10-brightgreen.svg)
 [![py311](https://img.shields.io/badge/python-3.11-brightgreen.svg)](https://img.shields.io/badge/python-3.11-brightgreen.svg)
-[![cu117](https://img.shields.io/badge/cuda-11.7-blue.svg)](https://img.shields.io/badge/cuda-11.7-blue.svg)
+[![cu118](https://img.shields.io/badge/cuda-11.8-blue.svg)](https://img.shields.io/badge/cuda-11.8-blue.svg)
 
 Bonito is an open source research basecaller for Oxford Nanopore reads.
 
@@ -30,6 +30,15 @@ $ bonito download --models --show  # show all available models
 $ bonito download --models         # download all available models
 ```
 
+## Transformer Models
+
+The `bonito.transformer` package requires
+[flash-attn](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#installation-and-features).
+
+This must be manually installed as the `flash-attn` packaging system prevents it from being listed as a normal dependency.
+
+Setting `CUDA_HOME` to the relevant library directory will help avoid CUDA version mismatches between packages.
+
 ## Modified Bases
 
 Modified base calling is handled by [Remora](https://github.com/nanoporetech/remora).
@@ -49,7 +58,7 @@ $ bonito basecaller dna_r9.4.1 --save-ctc --reference reference.mmi /data/reads
 $ bonito train --directory /data/training/ctc-data /data/training/model-dir
 ```
 
-In addition to training a new model from scratch you can also easily fine tune one of the pretrained models.  
+In addition to training a new model from scratch you can also easily fine tune one of the pretrained models.
 
 ```bash
 bonito train --epochs 1 --lr 5e-4 --pretrained dna_r10.4.1_e8.2_400bps_hac@v4.0.0 --directory /data/training/ctc-data /data/training/fine-tuned-model
@@ -62,7 +71,7 @@ $ bonito download --training
 $ bonito train /data/training/model-dir
 ```
 
-All training calls use Automatic Mixed Precision to speed up training. To disable this, set the `--no-amp` flag to True. 
+All training calls use Automatic Mixed Precision to speed up training. To disable this, set the `--no-amp` flag to True.
 
 ## Developer Quickstart
 
@@ -72,9 +81,13 @@ $ cd bonito
 $ python3 -m venv venv3
 $ source venv3/bin/activate
 (venv3) $ pip install --upgrade pip
-(venv3) $ pip install -e .
+(venv3) $ pip install -e .[cu118] --extra-index-url https://download.pytorch.org/whl/cu118
 ```
 
+The `ont-bonito[cu118]` and `ont-bonito[cu121]` optional dependencies can be used, along
+with the corresponding `--extra-index-url`, to ensure the PyTorch package matches the
+local CUDA setup.
+
 ## Interface
 
  - `bonito view` - view a model architecture for a given `.toml` file and the number of parameters in the network.

bonito/nn.py

@@ -2,6 +2,8 @@
 Bonito nn modules.
 """
 
+from collections import OrderedDict
+
 import torch
 from torch.nn import Module
 from torch.nn.init import orthogonal_
@@ -94,6 +96,81 @@ def to_dict(self, include_weights=False):
     def __repr__(self):
         return torch.nn.ModuleList.__repr__(self)
 
+
+@register
+class Stack(Serial):
+    @classmethod
+    def from_dict(cls, model_dict, layer_types=None):
+        return cls([from_dict(model_dict["layer"], layer_types) for _ in range(model_dict["depth"])])
+
+    def to_dict(self, include_weights=False):
+        if include_weights:
+            raise NotImplementedError
+        layer_dicts = [to_dict(layer) for layer in self]
+        for layer_dict in layer_dicts[1:]:
+            assert layer_dict == layer_dicts[0], "all layers should be the same"
+        return {"layer": layer_dicts[0], "depth": len(self)}
+
+
+@register
+class NamedSerial(torch.nn.Sequential):
+    @classmethod
+    def from_dict(cls, model_dict, layer_types=None):
+        return cls({k: from_dict(v, layer_types) for k, v in model_dict.items()})
+
+    def __init__(self, layers):
+        # Sequential throws error if given dict
+        super().__init__(OrderedDict(layers.items()))
+
+    def to_dict(self, include_weights=False):
+        if include_weights:
+            raise NotImplementedError
+        return {k: to_dict(v) for k, v in self.named_children()}
+
+
+class MakeContiguous(Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x):
+        return x.contiguous()
+
+
+@register
+class LinearUpsample(Module):
+    """
+    Applies a linear transformation to upsample the sequence length by ``scale_factor``.
+    """
+
+    def __init__(self, d_model, scale_factor, batch_first=True):
+        super().__init__()
+        self.d_model = d_model
+        self.scale_factor = scale_factor
+        self.batch_first = batch_first
+        self.linear = torch.nn.Linear(d_model, self.scale_factor * d_model)
+
+    def forward(self, src):
+        if not self.batch_first:
+            src = src.permute([1, 0, 2])
+        N, L, E = src.shape
+        h = self.linear(src).reshape(N, self.scale_factor * L, E)
+        if not self.batch_first:
+            h = h.permute([1, 0, 2])
+        return h
+
+    def output_stride(self, input_stride):
+        return input_stride // self.scale_factor
+
+    def to_dict(self, include_weights=False):
+        if include_weights:
+            raise NotImplementedError
+        return {
+            "d_model": self.d_model,
+            "scale_factor": self.scale_factor,
+            "batch_first": self.batch_first
+        }
+
+
 @register
 class Reverse(Module):
 
@@ -248,7 +325,7 @@ def extra_repr(self):
         if self.permute:
             rep += f', permute={self.permute}'
         return rep
-            
+
 
 @register
 class Permute(Module):

bonito/transformer/__init__.py

@@ -0,0 +1,2 @@
+from .model import Model
+from .basecall import basecall

bonito/transformer/basecall.py

@@ -0,0 +1 @@
+from bonito.crf import basecall

bonito/transformer/model.py

@@ -0,0 +1,154 @@
+import logging
+import types
+from functools import lru_cache
+
+logger = logging.getLogger(__name__)
+
+import torch
+import torch.nn.functional as F
+try:
+    from flash_attn import flash_attn_qkvpacked_func
+    from flash_attn.layers.rotary import RotaryEmbedding
+    from flash_attn.modules.mlp import GatedMlp
+    from flash_attn.ops.triton.layer_norm import RMSNorm
+except ImportError:
+    logger.warning(
+        "please install flash-attn to use the transformer module: "
+        "`pip install flash-attn --no-build-isolation`"
+    )
+
+from bonito.crf.model import SeqdistModel
+from bonito.nn import from_dict, register, LinearCRFEncoder, MakeContiguous, Module, Permute, Serial
+
+
+def deepnorm_params(depth):
+    """
+    Returns the DeepNorm (https://arxiv.org/abs/2203.00555) alpha and beta parameters.
+    """
+    alpha = round((2*depth)**0.25, 7)
+    beta = round((8*depth)**(-1/4), 7)
+    return alpha, beta
+
+
+@lru_cache(maxsize=2)
+def sliding_window_mask(seq_len, window, device):
+    band = torch.full((seq_len, seq_len), fill_value=1.0)
+    band = torch.triu(band, diagonal=-window[0])
+    band = band * torch.tril(band, diagonal=window[1])
+    band = band.to(torch.bool).to(device)
+    return band
+
+
+class MultiHeadAttention(Module):
+    def __init__(self, d_model, nhead, qkv_bias=False, out_bias=True, rotary_dim=None, attn_window=None):
+        super().__init__()
+        assert d_model % nhead == 0, "d_model must be divisible by nhead"
+
+        self.d_model = d_model
+        self.nhead = nhead
+        self.head_dim = d_model // nhead
+        self.rotary_dim = self.head_dim if rotary_dim is None else rotary_dim
+
+        self.Wqkv = torch.nn.Linear(d_model, 3 * d_model, bias=qkv_bias)
+        self.out_proj = torch.nn.Linear(d_model, d_model, bias=out_bias)
+
+        self.rotary_emb = RotaryEmbedding(self.rotary_dim, interleaved=False)
+        self.attn_window = (-1, -1) if attn_window is None else tuple(attn_window)
+
+    def attn_func(self, qkv):
+        if torch.cuda.get_device_capability(qkv.device)[0] >= 8 and (torch.is_autocast_enabled() or qkv.dtype == torch.half):
+            attn_output = flash_attn_qkvpacked_func(qkv, window_size=self.attn_window)
+        else:
+            q, k, v = torch.chunk(qkv.permute(0, 2, 3, 1, 4), chunks=3, dim=1)
+            mask = sliding_window_mask(qkv.shape[1], self.attn_window, q.device)
+            attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
+            attn_output = attn_output.permute(0, 1, 3, 2, 4)
+        return attn_output
+
+    def forward(self, x):
+        N, T, _ = x.shape
+
+        qkv = self.Wqkv(x).view(N, T, 3, self.nhead, self.head_dim)
+
+        qkv = self.rotary_emb(qkv)
+
+        attn_output = self.attn_func(qkv).reshape(N, T, self.d_model)
+
+        out = self.out_proj(attn_output)
+
+        return out
+
+
+@register
+class TransformerEncoderLayer(Module):
+    def __init__(self, d_model, nhead, dim_feedforward, deepnorm_alpha, deepnorm_beta, attn_window=None):
+        super().__init__()
+        self.kwargs = {
+            "d_model": d_model,
+            "nhead": nhead,
+            "dim_feedforward": dim_feedforward,
+            "deepnorm_alpha": deepnorm_alpha,
+            "deepnorm_beta": deepnorm_beta,
+            "attn_window": attn_window
+        }
+
+        self.self_attn = MultiHeadAttention(
+            d_model=d_model,
+            nhead=nhead,
+            qkv_bias=False,
+            out_bias=True,
+            attn_window=attn_window
+        )
+        self.ff = GatedMlp(
+            d_model,
+            hidden_features=dim_feedforward,
+            activation=F.silu,
+            bias1=False,
+            bias2=False,
+            multiple_of=1,
+        )
+        self.norm1 = RMSNorm(d_model)
+        self.norm2 = RMSNorm(d_model)
+
+        self.register_buffer("deepnorm_alpha", torch.tensor(deepnorm_alpha))
+        self.reset_parameters()
+
+    def reset_parameters(self):
+        db = self.kwargs["deepnorm_beta"]
+        d_model = self.kwargs["d_model"]
+        torch.nn.init.xavier_normal_(self.ff.fc1.weight, gain=db)
+        torch.nn.init.xavier_normal_(self.ff.fc2.weight, gain=db)
+        torch.nn.init.xavier_normal_(self.self_attn.out_proj.weight, gain=db)
+        torch.nn.init.xavier_normal_(self.self_attn.Wqkv.weight[2*d_model:], gain=db)
+        torch.nn.init.xavier_normal_(self.self_attn.Wqkv.weight[:2*d_model], gain=1)
+
+    def forward(self, x):
+        x = self.norm1(self.self_attn(x), self.deepnorm_alpha*x)
+        x = self.norm2(self.ff(x), self.deepnorm_alpha*x)
+        return x
+
+    def to_dict(self, include_weights=False):
+        if include_weights:
+            raise NotImplementedError
+        return self.kwargs
+
+
+def use_koi(self, **kwargs):
+    # koi needs modified LinearCRFLayer settings
+    def _expand_blanks(m):
+        if isinstance(m, LinearCRFEncoder):
+            m.expand_blanks = False
+    self.encoder.apply(_expand_blanks)
+    self.encoder = Serial([
+        self.encoder,
+        Permute([1, 0, 2]),
+        MakeContiguous(),
+    ])
+
+
+def Model(config):
+    model_config = {k: v for k, v in config["model"].items() if k != "package"}
+    model = from_dict(model_config)
+    model.config = config
+    model.use_koi = types.MethodType(use_koi, model)
+    return model

requirements.txt

@@ -3,17 +3,18 @@ mappy==2.24
 toml==0.10.2
 tqdm>4,<5
 scipy==1.10.1
+networkx~=3.1.0
 numpy~=1.24.2
 pysam==0.21.0
 parasail==1.3.4
 pandas>1,<2
-requests~=2.28.2
+requests~=2.31.0
 ont-koi==0.3.0
 ont-remora==2.1.3
 ont-fast5-api==3.3.0
 pod5==0.3.6
 fast-ctc-decode==0.3.5
 python-dateutil==2.8.2
 edlib==1.3.9
-# cuda requirements
-torch~=2.0.0
+torch~=2.2.1
+wheel

setup.py

@@ -1,7 +1,6 @@
 import os
 import re
 from setuptools import setup, find_packages
-from setuptools.command.install import install
 
 
 __pkg_name__ = 'bonito'
@@ -35,12 +34,15 @@
     author='Oxford Nanopore Technologies, Ltd',
     author_email='support@nanoporetech.com',
     url='https://github.com/nanoporetech/bonito',
-    entry_points = {
+    extras_require={
+        # --extra-index-url https://download.pytorch.org/whl/cu118
+        "cu118": ["torch==2.2.1+cu118"],
+        # --extra-index-url https://download.pytorch.org/whl/cu121
+        "cu121": ["torch==2.2.1+cu121"],
+    },
+    entry_points={
         'console_scripts': [
             '{0} = {0}:main'.format(__pkg_name__)
         ]
     },
-    dependency_links=[
-        'https://download.pytorch.org/whl/cu113',
-    ]
 )