News for hackers. Stuff that matters.
Re:You could have designed state of the art positional encoding by jmmcd on Monday November 18, 2024 @08:02AM (#42171962)
Well, yes, I deserved that reply! And yes the code is printing True. It's not that I disbelieved you... but something is wrong here. Investigation below, thanks to Claude.ai for walking me through it!

    In [10]: o1[0, :, :3]
    Out[10]:
    tensor([[ 0.0053,  0.0017, -0.0012],
        [ 0.0053,  0.0017, -0.0012],
        [ 0.0053,  0.0017, -0.0012],
        [ 0.0053,  0.0017, -0.0012],
        [ 0.0053,  0.0017, -0.0012],
        [ 0.0053,  0.0017, -0.0012]],       grad_fn=<SliceBackward0>)
Every token has the same attention values. I expect attention(cat, everything) to differ from attention(dog, everything), even without positional encoding.

Further, the attention weights are uniform and identical for both sentences:

    In [46]: o1, aw1 = mha(W_q(e1), W_k(e1), W_v(e1))
    In [47]: o2, aw2 = mha(W_q(e2), W_k(e2), W_v(e2))
    In [48]: aw1.shape
    Out[48]: torch.Size([1, 6, 6])
    In [49]: aw2.shape
    Out[49]: torch.Size([1, 6, 6])
    In [50]: aw1
    Out[50]:
    tensor([[[0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667]]],
       grad_fn=<MeanBackward1>)

    In [51]: aw2
    Out[51]:
    tensor([[[0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667],
         [0.1667, 0.1667, 0.1667, 0.1667, 0.1667, 0.1667]]],
       grad_fn=<MeanBackward1>)
That is not expected. It's because the Linear layers are initialised with such small values. And the softmax causes a collapse.

Trying random weights on a larger scale:

    In [52]: W_q.weight.data *= 100
         W_k.weight.data *= 100
         W_v.weight.data *= 100

    In [55]: o1, aw1 = mha(W_q(e1), W_k(e1), W_v(e1))
    In [56]: o2, aw2 = mha(W_q(e2), W_k(e2), W_v(e2))
    In [57]: aw1
    Out[57]:
    tensor([[[0.2049, 0.1606, 0.1256, 0.1095, 0.1723, 0.2270],
         [0.0883, 0.2047, 0.1544, 0.2776, 0.1405, 0.1345],
         [0.1196, 0.1719, 0.1831, 0.1541, 0.1374, 0.2339],
         [0.1413, 0.2399, 0.1617, 0.2056, 0.1634, 0.0880],
         [0.1455, 0.1432, 0.2432, 0.1239, 0.1494, 0.1948],
         [0.1897, 0.1817, 0.1920, 0.1478, 0.1618, 0.1270]]],
       grad_fn=<MeanBackward1>)

    In [58]: aw2
    Out[58]:
    tensor([[[0.2049, 0.1606, 0.2270, 0.1095, 0.1723, 0.1256],
         [0.0883, 0.2047, 0.1345, 0.2776, 0.1405, 0.1544],
         [0.1897, 0.1817, 0.1270, 0.1478, 0.1618, 0.1920],
         [0.1413, 0.2399, 0.0880, 0.2056, 0.1634, 0.1617],
         [0.1455, 0.1432, 0.1948, 0.1239, 0.1494, 0.2432],
         [0.1196, 0.1719, 0.2339, 0.1541, 0.1374, 0.1831]]],
       grad_fn=<MeanBackward1>)

    In [60]: o1[:, :, :5]
    Out[60]:
    tensor([[[ 0.0145,  0.3128, -0.3659, -0.1884,  0.1724],
         [-0.2319,  0.1407, -0.6010, -0.4064,  0.4259],
         [-0.3231,  0.1622, -0.6351, -0.1711,  0.4014],
         [-0.0596,  0.2610, -0.7388, -0.2987,  0.3214],
         [-0.2750,  0.0676, -0.4140, -0.2024,  0.3383],
         [-0.1434,  0.0871, -0.3154, -0.0755,  0.3314]]],
       grad_fn=<SliceBackward0>)

    In [61]: o2[:, :, :5]
    Out[61]:
    tensor([[[ 0.0145,  0.3128, -0.3659, -0.1884,  0.1724],
         [-0.2319,  0.1407, -0.6010, -0.4064,  0.4259],
         [-0.1434,  0.0871, -0.3154, -0.0755,  0.3314],
         [-0.0596,  0.2610, -0.7388, -0.2987,  0.3214],
         [-0.2750,  0.0676, -0.4140, -0.2024,  0.3383],
         [-0.3231,  0.1622, -0.6351, -0.1711,  0.4014]]],
       grad_fn=<SliceBackward0>)

    In [62]: print("Matches: ", torch.allclose(o1, o2, atol=1e-6))
    Matches:  False
Re:You could have designed state of the art positional encoding by FL33TW00D on Monday November 18, 2024 @08:25AM (#42172097)
Hm! Very interesting! Thank you for taking the time to debug that.

I'm going to have to think hard about how to rewrite the motivating example to explain this best.

Edit: updated the post, thanks for pointing out the pernicious init values!