birdflow metadata: hyperparameters vs. length of loss_values

[slager]

Seems to be a mismatch here between training_steps and length of loss_values. I specified 600 for the modelfit, so I think the 1000 may be a bug?

> str(bf[['metadata']][c('hyperparameters', 'loss_values')])
List of 2
 $ hyperparameters:List of 7
  ..$ dist_pow      : num 0.4
  ..$ dist_weight   : num 1e-04
  ..$ ent_weight    : num 0.001
  ..$ learning_rate : num 0.1
  ..$ obs_weight    : num 1
  ..$ rng_seed      : int 17
  ..$ training_steps: int 1000
 $ loss_values    :'data.frame':	600 obs. of  4 variables:
  ..$ dist : num [1:600] 128 126 125 124 123 ...
  ..$ ent  : num [1:600] 331 331 332 332 333 ...
  ..$ obs  : num [1:600] 0.345 0.339 0.333 0.326 0.32 ...
  ..$ total: num [1:600] 1.063 1.037 1.011 0.985 0.961 ...

[slager]

It's possible the whole hyperparameters section is hard-coded in the hdf5

> i <- 1
> ll_df$model[i]
[1] "buwtea_2021_93km_obs1.0_ent0.004_dist0.016_pow0.3.hdf5"
> bf <- import_birdflow(file.path(my_dir, ll_df$model[i]))
> str(bf[['metadata']][c('hyperparameters', 'loss_values')])
List of 2
 $ hyperparameters:List of 7
  ..$ dist_pow      : num 0.4
  ..$ dist_weight   : num 1e-04
  ..$ ent_weight    : num 0.001
  ..$ learning_rate : num 0.1
  ..$ obs_weight    : num 1
  ..$ rng_seed      : int 17
  ..$ training_steps: int 1000
 $ loss_values    :'data.frame':	600 obs. of  4 variables:
  ..$ dist : num [1:600] 128 126 125 124 123 ...
  ..$ ent  : num [1:600] 331 331 332 332 333 ...
  ..$ obs  : num [1:600] 0.345 0.339 0.333 0.326 0.32 ...
  ..$ total: num [1:600] 1.063 1.037 1.011 0.985 0.961 ...
> i <- 2
> 
> ll_df$model[i]
[1] "buwtea_2021_93km_obs1.0_ent0.0035_dist0.014_pow0.3.hdf5"
> bf <- import_birdflow(file.path(my_dir, ll_df$model[i]))
> str(bf[['metadata']][c('hyperparameters', 'loss_values')])
List of 2
 $ hyperparameters:List of 7
  ..$ dist_pow      : num 0.4
  ..$ dist_weight   : num 1e-04
  ..$ ent_weight    : num 0.001
  ..$ learning_rate : num 0.1
  ..$ obs_weight    : num 1
  ..$ rng_seed      : int 17
  ..$ training_steps: int 1000
 $ loss_values    :'data.frame':	600 obs. of  4 variables:
  ..$ dist : num [1:600] 128 127 125 124 123 ...
  ..$ ent  : num [1:600] 331 331 332 332 333 ...
  ..$ obs  : num [1:600] 0.345 0.338 0.332 0.326 0.32 ...
  ..$ total: num [1:600] 0.973 0.95 0.927 0.905 0.884 ...

[slager]

https://github.com/Miguel-Fuentes/birdflow/blob/82220ee8220b5c3631867469be174ea908102e00/update_hdf.py#L112

[slager]

My python PR fixes it on the Python end, I think. Import_birdflow may want to handle normalized differently to create logical instead of factor, see here:

  ..$ hyperparameters         :List of 8
  .. ..$ dist_pow      : num 0.1
  .. ..$ dist_weight   : num 0.008
  .. ..$ ent_weight    : num 0.0015
  .. ..$ learning_rate : num 0.1
  .. ..$ normalized    : Factor w/ 2 levels "FALSE","TRUE": 2
  .. ..$ obs_weight    : num 1
  .. ..$ rng_seed      : int 17
  .. ..$ training_steps: int 600

[ethanplunkett]

Thanks. I'll convert the "normalized" hyper-parameter to logical while reading from the hdf5. I don't think any other changes are necessary on the R side.