Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

TST parameter fixup #897

Open
wants to merge 3 commits into
base: main
Choose a base branch
from

Conversation

andersgb
Copy link

@andersgb andersgb commented Apr 23, 2024

First commit should be trivially correct (docstring only)
Second commit fixes a bug, which was my original intention for opening this PR
Third commit reintroduces the original behavior because I don't think the assertion is necessary. I tried to understand and compare to how the pytorch implementation of multi head attention handles these dimensions.

So the net change of this PR is only docstring changes. See commit messages for further details.

Fix documentation on default value of d_k and d_v. If they are not provided, and
d_model and n_heads are kept at default values, they will be set to 128//16
which is 8.

Also, specify the usual values range to 8-64 which should more or less correspond
to a d_model/n_heads range.
Follow the 050_models.TSTPlus.ipynb implementation and make sure we actually
assert what we say we assert on this line. This commit breaks the inline test
where d_model is 128 and n_heads is 3, see below.

```
AssertionError in /home/anders/dev/ml/tsai/nbs/049_models.TST.ipynb:
===========================================================================

While Executing Cell timeseriesAI#13:
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
Cell In[1], line 2
      1 t = torch.rand(16, 50, 128)
----> 2 output = _TSTEncoderLayer(q_len=50, d_model=128, n_heads=3, d_k=None, d_v=None, d_ff=512, dropout=0.1, activation='gelu')(t)
      3 output.shape

File ~/anaconda3/envs/tsai_dev/lib/python3.9/site-packages/fastcore/meta.py:40, in PrePostInitMeta.__call__(cls, *args, **kwargs)
     38 if type(res)==cls:
     39     if hasattr(res,'__pre_init__'): res.__pre_init__(*args,**kwargs)
---> 40     res.__init__(*args,**kwargs)
     41     if hasattr(res,'__post_init__'): res.__post_init__(*args,**kwargs)
     42 return res

Cell In[1], line 11, in _TSTEncoderLayer.__init__(self, q_len, d_model, n_heads, d_k, d_v, d_ff, dropout, activation)
      8 def __init__(self, q_len:int, d_model:int, n_heads:int, d_k:Optional[int]=None, d_v:Optional[int]=None, d_ff:int=256, dropout:float=0.1,
      9              activation:str="gelu"):
---> 11     assert not d_model%n_heads, f"d_model ({d_model}) must be divisible by n_heads ({n_heads})"
     12     d_k = ifnone(d_k, d_model // n_heads)
     13     d_v = ifnone(d_v, d_model // n_heads)

AssertionError: d_model (128) must be divisible by n_heads (3)
```
I believe this assertion is unnecessary because the dimensions will actually
work out even if d_model is not divisible by n_heads. See a model printout of
d_model=128, n_heads=3, d_k=11, d_v=9 below. Not saying a parameter change like
this is a good idea, but it will happily run and learn on a sample dataset I'm
using at least. The in_features and out_features of the entire MHA block will be
d_model in any case.

Comparing with torch.nn.MultiheadAttention [1] (which is used in the original
paper implementation [2]), I think our `d_k*n_heads` corresponds to the `kdim`
optional parameter. And, similarly, our `d_v*n_heads` corresponds to the `vdim`
parameter.

[1] https://pytorch.org/docs/stable/generated/torch.nn.MultiheadAttention.html
[2] https://github.com/gzerveas/mvts_transformer/blob/3f2e378bc77d02e82a44671f20cf15bc7761671a/src/models/ts_transformer.py#L152

```
In [2]: clf.model
Out[2]:
TST(
  (W_P): Linear(in_features=600, out_features=128, bias=True)
  (dropout): Dropout(p=0.1, inplace=False)
  (encoder): _TSTEncoder(
    (layers): ModuleList(
      (0-2): 3 x _TSTEncoderLayer(
        (self_attn): _MultiHeadAttention(
          (W_Q): Linear(in_features=128, out_features=33, bias=False)
          (W_K): Linear(in_features=128, out_features=33, bias=False)
          (W_V): Linear(in_features=128, out_features=27, bias=False)
          (W_O): Linear(in_features=27, out_features=128, bias=False)
        )
        (dropout_attn): Dropout(p=0.1, inplace=False)
        (batchnorm_attn): Sequential(
          (0): Transpose(1, 2)
          (1): BatchNorm1d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (2): Transpose(1, 2)
        )
        (ff): Sequential(
          (0): Linear(in_features=128, out_features=256, bias=True)
          (1): GELU(approximate='none')
          (2): Dropout(p=0.1, inplace=False)
          (3): Linear(in_features=256, out_features=128, bias=True)
        )
        (dropout_ffn): Dropout(p=0.1, inplace=False)
        (batchnorm_ffn): Sequential(
          (0): Transpose(1, 2)
          (1): BatchNorm1d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
          (2): Transpose(1, 2)
        )
      )
    )
  )
  (flatten): fastai.layers.Flatten(full=False)
  (head): Sequential(
    (0): GELU(approximate='none')
    (1): fastai.layers.Flatten(full=False)
    (2): Linear(in_features=2304, out_features=2, bias=True)
  )
)
```
Copy link

Check out this pull request on  ReviewNB

See visual diffs & provide feedback on Jupyter Notebooks.


Powered by ReviewNB

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

1 participant