TypeError: __init__() got an unexpected keyword argument 'normalize'

[Y-Isaac]

HI,

When I use polyfun.py to re-estimate per-SNP heritabilities via S-LDSC, there have a error:

[INFO] Reading summary statistics from /public/home/P202306/polyfun_test/summary/pheno1_munged.parquet ...
[INFO] Read summary statistics for 13087844 SNPs.
[INFO] Reading reference panel LD Score from /public/home/P202306/polyfun_test/output/pheno1/pheno1.[1-22] ...
[INFO] Read reference panel LD Scores for 13156184 SNPs.
[INFO] Reading regression weight LD Score from /public/home/P202306/polyfun_test/ldscore/weight/[1-22] ...
[INFO] Read regression weight LD Scores for 13156184 SNPs.
[INFO] After merging with reference panel LD, 13087844 SNPs remain.
[INFO] After merging with regression SNP LD, 13087844 SNPs remain.
[INFO] Removed 183 SNPs with chi^2 > 431.334 (13087661 SNPs remain)
Traceback (most recent call last):
File "/public/home/P202306/software/polyfun/polyfun.py", line 849, in
polyfun_obj.polyfun_main(args)
File "/public/home/P202306/software/polyfun/polyfun.py", line 780, in polyfun_main
self.compute_h2_bins(args, constrain_range=True)
File "/public/home/P202306/software/polyfun/polyfun.py", line 757, in compute_h2_bins
self.run_ldsc(args, use_ridge=False, nn=True, evenodd_split=True, keep_large=False)
File "/public/home/P202306/software/polyfun/polyfun.py", line 217, in run_ldsc
hsqhat = regressions.Hsq(chisq,
File "/public/home/P202306/software/polyfun/ldsc_polyfun/regressions.py", line 401, in init
LD_Score_Regression.init(self, y, x, w, N, M, n_blocks, intercept=intercept,
File "/public/home/P202306/software/polyfun/ldsc_polyfun/regressions.py", line 243, in init
jknife = jk.LstsqJackknifeSlow(x, y, is_large_chi2, n_blocks, evenodd_split=evenodd_split, nn=True, chr_num=chr_num, nnls_exact=nnls_exact)
File "/public/home/P202306/software/polyfun/ldsc_polyfun/jackknife.py", line 267, in init
lasso = Lasso(alpha=1e-100, fit_intercept=False, normalize=False, precompute=xtx, positive=True, max_iter=10000, random_state=0)
TypeError: init() got an unexpected keyword argument 'normalize'

And this is my code, if it's helpful:

python ~/software/polyfun/polyfun.py
--compute-h2-bins
--output-prefix /public/home/P202306/polyfun_test/output/pheno1/pheno1
--sumstats /public/home/P202306/polyfun_test/summary/pheno1_munged.parquet
--w-ld-chr /public/home/P202306/polyfun_test/ldscore/weight/

[jdblischak]

lasso = Lasso(alpha=1e-100, fit_intercept=False, normalize=False, precompute=xtx, positive=True, max_iter=10000, random_state=0)
TypeError: init() got an unexpected keyword argument 'normalize'

It looks like the function Lasso no longer has the argument normalize. This is likely a version difference. Could you please try running your code in the locked conda env, polyfun.yml.lock, that is known to work with the polyfun scripts:

mamba create --name polyfun --file polyfun.yml.lock
conda activate polyfun

[Y-Isaac]

@jdblischak HI,

Following your advice, I configured the environment based on the file polyfun.yml.lock. Since mamba is not available on my server, I used the following command to create the environment, hoping it would serve the same purpose:

conda create --name polyfun-lock --file polyfun.yml.lock

Unfortunately, this method doesn't seem to have worked, as I encountered the same error message again. I tried checking the version of the sklearn package in both environments. In polyfun-lock, the version is 1.2.2, while in the polyfun environment, it's 1.3.2, yet they both produced the same error.

All in all, thank you very much for your help!

[Y-Isaac]

I attempted to remove the normalize=False parameter from lines 267 and 295 in the jackknife.py script (I'm not certain this was a reliable action, it was only a trial), and this time it worked. I reviewed the prior probability results for chromosome 22, which range from 7.19e-7 to 9.09e-9, from highest to lowest. Compared to the example file, this result seems to be normal.

I eagerly look forward to your guidance on what I should do next. I hope you have a pleasant day!

[jdblischak]

I investigated the argument normalize. Turns out it was deprecated in scikit-learn 1.0.0!

API Change : The parameter normalize of linear_model.LinearRegression is deprecated and will be removed in 1.2. Motivation for this deprecation: normalize parameter did not take any effect if fit_intercept was set to False and therefore was deemed confusing. The behavior of the deprecated LinearModel(normalize=True) can be reproduced with a Pipeline with LinearModel (where LinearModel is LinearRegression, Ridge, RidgeClassifier, RidgeCV or RidgeClassifierCV) as follows: make_pipeline(StandardScaler(with_mean=False), LinearModel()). The normalize parameter in LinearRegression was deprecated in #17743 by Maria Telenczuk and Alexandre Gramfort. Same for Ridge, RidgeClassifier, RidgeCV, and RidgeClassifierCV, in: #17772 by Maria Telenczuk and Alexandre Gramfort. Same for BayesianRidge, ARDRegression in: #17746 by Maria Telenczuk. Same for Lasso, LassoCV, ElasticNet, ElasticNetCV, MultiTaskLasso, MultiTaskLassoCV, MultiTaskElasticNet, MultiTaskElasticNetCV, in: #17785 by Maria Telenczuk and Alexandre Gramfort.

So now I am confused why this hasn't been caught before. We test polyfun.py --compute-h2-bins

polyfun/test_polyfun.py

Lines 144 to 146 in 00afe71

	#polyfun stage 4
	cmd = '%s %s --compute-h2-bins --output-prefix %s --sumstats %s --w-ld-chr %s --nnls-exact'% \
	(python3_exe, script_exe, output_prefix, sumstats_file, w_ld_prefix)

Ah, it's because of the flag --nnls-exact used in the test. That bypasses the call to Lasso():

polyfun/ldsc_polyfun/jackknife.py

Lines 261 to 270 in 00afe71

	#estimate taus
	if nn: # non-negative least squares
	if nnls_exact:
	self.est = np.atleast_2d(nnls(x, np.array(y).T[0])[0])
	else:
	xtx = x.T.dot(x)
	lasso = Lasso(alpha=1e-100, fit_intercept=False, normalize=False, precompute=xtx, positive=True, max_iter=10000, random_state=0)
	self.est = lasso.fit(x,y[:,0]).coef_.reshape((1, x.shape[1]))
	else:
	self.est = np.atleast_2d(np.linalg.lstsq(x, np.array(y).T[0])[0])

polyfun/ldsc_polyfun/jackknife.py

Lines 290 to 298 in 00afe71

	if nnls_exact:
	jk_est = np.atleast_2d(nnls(x_noblock, y_noblock[:,0])[0])
	else:
	x_block = x[s[i] : s[i+1]]
	xtx_noblock = xtx - x_block.T.dot(x_block)
	lasso_noblock = Lasso(alpha=1e-100, fit_intercept=False, normalize=False, precompute=xtx_noblock, positive=True, max_iter=10000, random_state=0)
	jk_est = lasso_noblock.fit(x_noblock, y_noblock[:,0]).coef_.reshape((1, x.shape[1]))
	###z = nnls(x_noblock, y_noblock[:,0])[0]
	###assert np.allclose(z, jk_est[0])

[Y-Isaac]

@jdblischak ohh, I get it, thanks for your help! Now I'm going to close this issue.

[omerwe]

@Y-Isaac thanks for flagging this! I've accepted pull request #183, so the problem should be fixed for everyone now (thanks @jdblischak!)