Datasets:
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
reinforcement-learning
policy-optimization
klpo
exact-moment-replay
score-centering
stratified-sampling
License:
File size: 4,632 Bytes
f95cbd4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 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 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 | import torch
import pytest
from celestis_rl.streaming import centered_linear_scores,linear_token_loss
from celestis_rl.estimators import draw_correction
from celestis_rl.losses import token_loss
@pytest.mark.parametrize('mode',['exact','sampled'])
@pytest.mark.parametrize('chunk',[1,4,20])
@pytest.mark.parametrize('temperature',[.3,1.,2.])
@pytest.mark.parametrize('has_bias',[False,True])
def test_streamed_gradient_oracle(mode,chunk,temperature,has_bias):
torch.manual_seed(31)
h=torch.randn(5,4,dtype=torch.float64,requires_grad=True)
w=torch.randn(11,4,dtype=torch.float64,requires_grad=True)
bias=torch.randn(11,dtype=torch.float64,requires_grad=True) if has_bias else None
q=torch.randn(5,11,dtype=torch.float64).softmax(-1)
a=torch.tensor([2,6,2,0,10]);g=torch.randn(5,dtype=torch.float64)
opts={'q':q} if mode=='exact' else {}
c=draw_correction(q[:,None],torch.ones(5,1,dtype=torch.bool),4,'iid',generator=torch.Generator().manual_seed(13))
ids=torch.cat((a[:,None],c.ids[:,0]),-1)
coefficients=torch.cat((torch.ones(5,1,dtype=torch.float64),-c.weights[:,0]),-1)
if mode=='sampled':opts={'ids':ids,'coefficients':coefficients}
res=centered_linear_scores(h,w,a,bias=bias,temperature=temperature,chunk_size=chunk,row_chunk_size=2,**opts)
params=[h,w]+([bias] if has_bias else [])
custom=torch.autograd.grad((res.centered_logits*g).sum(),params)
z=(h@w.T+(bias if has_bias else 0))/temperature
target=z.gather(-1,a[:,None]).squeeze(-1)-(q*z).sum(-1) if mode=='exact' else (z.gather(-1,ids)*coefficients).sum(-1)
dense=torch.autograd.grad((target*g).sum(),params)
assert torch.allclose(res.centered_logits,target,atol=2e-12,rtol=2e-12)
assert torch.allclose(res.action_log_probs,z.log_softmax(-1).gather(-1,a[:,None]).squeeze(-1),atol=2e-12)
assert not res.action_log_probs.requires_grad
for x,y in zip(custom,dense):assert torch.allclose(x,y,atol=2e-12,rtol=2e-12)
@pytest.mark.parametrize('sampled',[False,True])
def test_token_loss_masked_equivalence(sampled):
torch.manual_seed(17)
h=torch.randn(2,3,4,dtype=torch.float64,requires_grad=True)
w=torch.randn(9,4,dtype=torch.float64,requires_grad=True)
q=torch.randn(2,3,9,dtype=torch.float64).softmax(-1)
a=torch.tensor([[1,3,6],[4,2,0]]);mask=torch.tensor([[True,True,True],[True,True,False]])
r=torch.tensor([1.,.5],dtype=torch.float64);weights=torch.tensor([.75,1.25],dtype=torch.float64)
with torch.no_grad():h[1,2]=float('nan');q[1,2]=float('nan')
rec=draw_correction(q,mask,3,'iid') if sampled else None
L,_=linear_token_loss(h,w,a,q,mask,r,correction=rec,beta=.2,outer_weights=weights,normalizer=4.,chunk_size=4)
gs=torch.autograd.grad(L,(h,w))
z=h.masked_fill(~mask[...,None],0)@w.T
oracle,_=token_loss(z,a,q,mask,r,correction=rec,beta=.2,outer_weights=weights,normalizer=4.)
gd=torch.autograd.grad(oracle,(h,w))
for x,y in zip(gs,gd):assert torch.allclose(x,y,atol=1e-12)
assert torch.allclose(L,oracle,atol=1e-12)
@pytest.mark.parametrize('sampled',[False,True])
def test_gradcheck(sampled):
torch.manual_seed(40)
h=torch.randn(2,3,dtype=torch.float64,requires_grad=True)
w=torch.randn(5,3,dtype=torch.float64,requires_grad=True)
a=torch.tensor([1,3]);q=torch.ones(2,5,dtype=torch.float64)/5
opts={'q':q}
if sampled:opts={'ids':torch.tensor([[1,2],[3,1]]),'coefficients':torch.tensor([[1.,-1.],[1.,-1.]],dtype=torch.float64)}
assert torch.autograd.gradcheck(lambda h,w:centered_linear_scores(h,w,a,chunk_size=2,**opts).centered_logits,(h,w))
def test_measure_and_nan_validation():
h=torch.ones(1,2);w=torch.ones(3,2);a=torch.tensor([0])
with pytest.raises(ValueError):centered_linear_scores(h,w,a,q=torch.ones(1,3))
with pytest.raises(ValueError):centered_linear_scores(h,w,a,ids=a[:,None],coefficients=torch.ones(1,1))
with pytest.raises(ValueError):centered_linear_scores(h,w,a,q=torch.ones(1,3)/3,row_chunk_size=0)
with pytest.raises(ValueError):centered_linear_scores(h*float('nan'),w,a,q=torch.ones(1,3)/3)
def test_tied_embedding_gradient():
torch.manual_seed(12)
w=torch.randn(8,4,dtype=torch.float64,requires_grad=True)
input_ids=torch.tensor([1,3,6]);a=torch.tensor([2,1,4]);q=torch.ones(3,8,dtype=torch.float64)/8
h=w[input_ids]
out=centered_linear_scores(h,w,a,q=q,chunk_size=3).centered_logits.sum()
gc=torch.autograd.grad(out,w)[0]
logits=w[input_ids]@w.T
reference=(logits.gather(-1,a[:,None]).squeeze(-1)-(q*logits).sum(-1)).sum()
gd=torch.autograd.grad(reference,w)[0]
assert torch.allclose(gc,gd,atol=1e-12)
|