不同的 Pyro Paramstore 访问方法给出不同的结果

Question

我正在学习 forecasting 中的 Pyro 入门教程，并在训练模型后尝试访问学习的参数，我对其中一些使用不同的访问方法得到了不同的结果（而对其他人则得到相同的结果） ).

这是教程中精简的可重现代码：

import torch
import pyro
import pyro.distributions as dist
from pyro.contrib.examples.bart import load_bart_od
from pyro.contrib.forecast import ForecastingModel, Forecaster

pyro.enable_validation(True)
pyro.clear_param_store()

pyro.__version__
# '1.3.1'
torch.__version__
# '1.5.0+cu101'

# import & prepare the data
dataset = load_bart_od()
T, O, D = dataset["counts"].shape
data = dataset["counts"][:T // (24 * 7) * 24 * 7].reshape(T // (24 * 7), -1).sum(-1).log()
data = data.unsqueeze(-1)
T0 = 0              # begining
T2 = data.size(-2)  # end
T1 = T2 - 52        # train/test split

# define the model class
class Model1(ForecastingModel):

    def model(self, zero_data, covariates):
        data_dim = zero_data.size(-1)  
        feature_dim = covariates.size(-1)

        bias = pyro.sample("bias", dist.Normal(0, 10).expand([data_dim]).to_event(1))
        weight = pyro.sample("weight", dist.Normal(0, 0.1).expand([feature_dim]).to_event(1))
        prediction = bias + (weight * covariates).sum(-1, keepdim=True)
        assert prediction.shape[-2:] == zero_data.shape

        noise_scale = pyro.sample("noise_scale", dist.LogNormal(-5, 5).expand([1]).to_event(1))
        noise_dist = dist.Normal(0, noise_scale)

        self.predict(noise_dist, prediction)

# fit the model
pyro.set_rng_seed(1)
pyro.clear_param_store()
time = torch.arange(float(T2)) / 365
covariates = torch.stack([time], dim=-1)
forecaster = Forecaster(Model1(), data[:T1], covariates[:T1], learning_rate=0.1)

到目前为止一切顺利；现在，我想检查存储在 Paramstore 中的学习到的潜在参数。似乎有不止一种方法可以做到这一点；使用 get_all_param_names() 方法：

for name in pyro.get_param_store().get_all_param_names():
    print(name, pyro.param(name).data.numpy())

我明白了

AutoNormal.locs.bias [14.585433]
AutoNormal.scales.bias [0.00631594]
AutoNormal.locs.weight [0.11947815]
AutoNormal.scales.weight [0.00922901]
AutoNormal.locs.noise_scale [-2.0719821]
AutoNormal.scales.noise_scale [0.03469057]

但是使用named_parameters()方法：

pyro.get_param_store().named_parameters()

为位置 (locs) 参数提供相同的值，但 为所有 scales 个参数提供不同的值：

dict_items([
('AutoNormal.locs.bias', Parameter containing: tensor([14.5854], requires_grad=True)), 
('AutoNormal.scales.bias', Parameter containing: tensor([-5.0647], requires_grad=True)), 
('AutoNormal.locs.weight', Parameter containing: tensor([0.1195], requires_grad=True)), 
('AutoNormal.scales.weight', Parameter containing: tensor([-4.6854], requires_grad=True)),
('AutoNormal.locs.noise_scale', Parameter containing: tensor([-2.0720], requires_grad=True)), 
('AutoNormal.scales.noise_scale', Parameter containing: tensor([-3.3613], requires_grad=True))
])

这怎么可能？根据documentation，Paramstore是一个简单的键值存储；里面只有这六个键：

pyro.get_param_store().get_all_param_names() # .keys() method gives identical result
# result
dict_keys([
'AutoNormal.locs.bias',
'AutoNormal.scales.bias', 
'AutoNormal.locs.weight', 
'AutoNormal.scales.weight', 
'AutoNormal.locs.noise_scale', 
'AutoNormal.scales.noise_scale'])

因此，不可能一种方法访问一组项目而另一种方法访问另一组项目。

我是不是漏掉了什么？

Answer 1

pyro.param() returns transformed parameters 在这种情况下为 scales.

的正实数

Answer 2

情况是这样的，在Github thread我打开这个问题的同时打开...

Paramstore 不再是只是一个简单的键值存储——它还执行约束转换；引用上面的 Pyro 开发人员 link:

here's some historical background. The ParamStore was originally just a key-value store. Then we added support for constrained parameters; this introduced a new layer of separation between user-facing constrained values and internal unconstrained values. We created a new dict-like user-facing interface that exposed only constrained values, but to keep backwards compatibility with old code we kept the old interface around. The two interfaces are distinguished in the source files [...] but as you observe it looks like we forgot to mark the old interface as DEPRECATED.

I guess in clarifying docs we should:

clarify that the ParamStore is no longer a simple key-value store but also performs constraint transforms;

mark all "old" style interface methods as DEPRECATED;

remove "old" style interface usage from examples and tutorials.

因此，事实证明，虽然 pyro.param() returns 约束（面向用户）space 的结果，但较旧的方法 named_parameters() returns 不受约束（即仅供内部使用）的值，因此存在明显的差异。

不难验证以上两种方法返回的scales值确实存在对数关系：

import numpy as np
items = list(pyro.get_param_store().named_parameters())  # unconstrained space

i = 0
for name in pyro.get_param_store().keys():  
  if 'scales' in name:
    temp = np.log(
                  pyro.param(name).item()  # constrained space
                 )
    print(temp, items[i][1][0].item() , np.allclose(temp, items[i][1][0].item()))
  i+=1

# result:
-5.027793402915326 -5.0277934074401855 True
-4.600319371162187 -4.6003193855285645 True
-3.3920585732532835 -3.3920586109161377 True

为什么这种差异只影响 scales 个参数？这是因为 scales（即本质上方差）根据定义被限制为正数；这不适用于不受约束的 locs（即均值），因此这两种表示对它们来说是一致的。

作为result of the question above, a new bullet has now been added in the Paramstore documentation，给出相关提示：

in general parameters are associated with both constrained and unconstrained values. for example, under the hood a parameter that is constrained to be positive is represented as an unconstrained tensor in log space.

以及旧接口named_parameters()方法的documentation中：

Note that, in the event the parameter is constrained, unconstrained_value is in the unconstrained space implicitly used by the constraint.

不同的 Pyro Paramstore 访问方法给出不同的结果

Different access methods to Pyro Paramstore give different results

probabilistic-programming

pyro.ai