[pymvpa] Fwd: Searchlight Accuracy Decimal Precision Issue
adkinsty at umich.edu
Thu May 17 23:19:25 BST 2018
Below is the output of data.summary() for one subject:
Dataset: 211x260502 at float32, <sa: chunks,targets>, <fa: voxel_indices>, <a:
stats: mean=-0.0847961 std=0.967814 var=0.936665 min=-22.0178 max=24.7773
No details due to large number of targets or chunks. Increase maxc and maxt
Summary for targets across chunks
targets mean std min max #chunks
0 0.502 0.5 0 1 106
1 0.498 0.5 0 1 105
Sequence statistics for 211 entries from set [0, 1]
Counter-balance table for orders up to 2:
Targets/Order O1 | O2 |
0: 0 105 | 105 0 |
1: 105 0 | 0 104 |
Correlations: min=-0.99 max=0.98 mean=-0.0048 sum(abs)=1e+02
On Thu, May 17, 2018 at 7:21 AM, Yaroslav Halchenko <yoh at onerussian.com>
> Please share output of
> print data.summary()
> Right before your give it to searchlight
> On May 17, 2018 3:59:39 AM EDT, Tyler Adkins <adkinsty at umich.edu> wrote:
>> I am running a searchlight analysis and I encountered a strange feature
>> of the resulting classification accuracies. All of the non-zero accuracies
>> have only one decimal point of precision (e.g., .5, .6., .7., .8, etc.).
>> Unfortunately, I'm not able to figure out what aspect of my code is leading
>> to this truncating or rounding of the accuracies.
>> I attached my code below, but here's some information about my analysis.
>> Please let me know if you would like me to provide any other information
>> about the analysis. My searchlight repeats a 10-fold cross-validation
>> procedure for a linear support vector classifier with default parameters.
>> The number of classes is 2 and the total number of samples is roughly 240.
>> The sequence of samples is randomized and balanced so that there is an
>> equal number of instances of the two classes in each of the 10 folds. The
>> searchlight also applies the mean_sample() postproc so that the resulting
>> classification accuracies are averaged over the cross-validation folds.
>> As mentioned above, I'm unsure what about my script is leading to the
>> rounding/truncating of the accuracies. Most of the script is copied from
>> one of the PyMVPA searchlight tutorials, which further adds to my
>> confusion, since the tutorial searchlight clearly outputs accuracies with
>> greater than 1-decimal precision.
>> I would greatly appreciate any ideas you might have about what could be
>> causing this problem or how to address it.
>> Thank you for your time.
>> Tyler Adkins
>> PhD Pre-candidate | Cognition and Cognitive Neuroscience
>> University of Michigan Department of Psychology
>> 530 Church Street Ann Arbor, MI 48109
>> Email: adkinsty at umich.edu
>> Office: 3036 East Hall
>> Lab: B018 East Hall
> Sent from a phone which beats iPhone.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Pkg-ExpPsy-PyMVPA