Cannot tune ets_model and loss arguments in adam_reg() function.

[tavad]

I am teaching myself how this function works by tuning it. I may be mistaken, but I do not think ets and loss arguments should be present in a model that is based on ARIMA. Again I may be wrong.

I was trying to tune these parameters and it trows errors as seen in the code below. PS. thanks for your contributions!!

library(dplyr)
library(parsnip)
library(rsample)
library(timetk)
library(modeltime)
library(smooth)

Data

m750 <- m4_monthly %>% filter(id == "M750")

Split Data 80/20

set.seed(1234)
splits <- initial_time_split(m750, prop = 0.8)

set.seed(456)
time_roll <-
rolling_origin(
data = training(splits),
initial = round(nrow(training(splits)) * 0.75),
assess = 24,
cumulative = TRUE,
skip = TRUE
)

recipe

rec <- recipe(value ~ date, data = training(splits)) %>%
step_log(value)

Model Spec

model_spec <-
modeltime::adam_reg(
ets_model = tune("ets"),
# loss = tune("loss")
) %>%
set_engine('adam')

selecting terms to be tuned

ets_params <-
crossing(
error = c("A", "M"),
trend = c("N", "A", "Ad", "M", "Md"),
seasonality = c("N", "A", "M")
) %>%
mutate(comb = paste0(error, trend, seasonality)) %>%
pull(comb) %>%
sample(2) # randomly selecting 2 ets terms to be tuned

loss_params <-
c("likelihood", "MSE", "MAE", "HAM", "LASSO",
"RIDGE", "MSEh", "TMSE", "GTMSE", "MSCE") %>%
sample(2) # randomly selecting 2 loss terms to be tuned

tuning the model

adam_fit <-
workflow() %>%
add_model(model_spec) %>%
add_recipe(rec) %>%
tune_grid(
resamples = time_roll,
metrics = metric_set(rmse, rsq),
grid = crossing(
ets = loss_params,
)
)

the error

 # Error in `mutate()`:
 #   ℹ In argument: `object = purrr::map(call_info, eval_call_info)`.
 # Caused by error in `purrr::map()`:
 #   ℹ In index: 1.
 # Caused by error in `.f()`:
 #   ! Error when calling ets_model(): Error : 'ets_model' is not an exported object from 'namespace:modeltime'


 # > rlang::last_trace()
 # <error/dplyr:::mutate_error>
 #   Error in `mutate()`:
 #   ℹ In argument: `object = purrr::map(call_info, eval_call_info)`.
 # Caused by error in `purrr::map()`:
 #   ℹ In index: 1.
 # Caused by error in `.f()`:
 #   ! Error when calling ets_model(): Error : 'ets_model' is not an exported object from 'namespace:modeltime'
 # ---
 #   Backtrace:
 #   ▆
 # 1. ├─... %>% ...
 # 2. ├─tune::tune_grid(...)
 # 3. ├─tune:::tune_grid.workflow(...)
 # 4. │ └─tune:::tune_grid_workflow(...)
 # 5. │   └─tune::check_parameters(...)
 # 6. │     ├─hardhat::extract_parameter_set_dials(wflow)
 # 7. │     └─workflows:::extract_parameter_set_dials.workflow(wflow)
 # 8. │       ├─hardhat::extract_parameter_set_dials(model)
 # 9. │       └─parsnip:::extract_parameter_set_dials.model_spec(model)
 # 10. │         └─... %>% ...
 # 11. ├─dplyr::mutate(., object = purrr::map(call_info, eval_call_info))
 # 12. ├─dplyr:::mutate.data.frame(., object = purrr::map(call_info, eval_call_info))
 # 13. │ └─dplyr:::mutate_cols(.data, dplyr_quosures(...), by)
 # 14. │   ├─base::withCallingHandlers(...)
 # 15. │   └─dplyr:::mutate_col(dots[[i]], data, mask, new_columns)
 # 16. │     └─mask$eval_all_mutate(quo)
 # 17. │       └─dplyr (local) eval()
 # 18. └─purrr::map(call_info, eval_call_info)
 # 19.   └─purrr:::map_("list", .x, .f, ..., .progress = .progress)
 # 20.     ├─purrr:::with_indexed_errors(...)
 # 21.     │ └─base::withCallingHandlers(...)
 # 22.     ├─purrr:::call_with_cleanup(...)
 # 23.     └─parsnip (local) .f(.x[[i]], ...)
 # 24.       └─base::stop(paste0("Error when calling ", x$fun, "(): ", as.character(res)))
 # 

[mdancho84]

I've fixed this in the development version of modeltime.

Install:

remotes::isntall_github("business-science/modeltime")

Reproducible example:

library(dplyr)
library(parsnip)
library(rsample)
library(timetk)
library(modeltime)
library(smooth)


m750 <- m4_monthly %>% filter(id == "M750")

set.seed(1234)
splits <- initial_time_split(m750, prop = 0.8)

set.seed(456)
time_roll <-
    rolling_origin(
        data = training(splits),
        initial = round(nrow(training(splits)) * 0.75),
        assess = 24,
        cumulative = TRUE,
        skip = TRUE
    )

# recipe
rec <- recipe(value ~ date, data = training(splits)) %>%
    step_log(value)

# Model Spec
model_spec <-
    modeltime::adam_reg(
        # ets_model = tune("ets"),
        loss = tune()
    ) %>%
    set_engine('adam')

# selecting terms to be tuned
ets_params <-
    crossing(
        error = c("A", "M"),
        trend = c("N", "A", "Ad", "M", "Md"),
        seasonality = c("N", "A", "M")
    ) %>%
    mutate(comb = paste0(error, trend, seasonality)) %>%
    pull(comb) %>%
    sample(2) # randomly selecting 2 ets terms to be tuned

loss_params <-
    c("likelihood", "MSE", "MAE", "HAM", "LASSO",
      "RIDGE", "MSEh", "TMSE", "GTMSE", "MSCE") %>%
    sample(2) # randomly selecting 2 loss terms to be tuned


grid_1 <- crossing(
    # ets_model = ets_params,
    loss = loss_params
)

# tuning the model
adam_fit <-
    workflow() %>%
    add_model(model_spec) %>%
    add_recipe(rec) %>%
    tune_grid(
        resamples = time_roll,
        metrics = metric_set(rmse, rsq),
        grid = grid_1
    )

Output:

frequency = 12 observations per 1 year
→ A | warning: You have not provided lambda parameter. I will set it to zero.
frequency = 12 observations per 1 yearns   A: x1
→ B | error:   The horizon "h" needs to be specified and be positive in order for the multistep loss to work.
→ C | warning: A correlation computation is required, but `estimate` is constant and has 0 standard deviation, resulting in a divide by
               0 error. `NA` will be returned.
frequency = 12 observations per 1 yearns   A: x1   B: x1   C: x1
frequency = 12 observations per 1 yearns   A: x2   B: x1   C: x1
frequency = 12 observations per 1 yearns   A: x2   B: x2   C: x1
frequency = 12 observations per 1 yearns   A: x3   B: x2   C: x2
frequency = 12 observations per 1 yearns   A: x3   B: x3   C: x2
frequency = 12 observations per 1 yearns   A: x4   B: x3   C: x3
frequency = 12 observations per 1 yearns   A: x4   B: x4   C: x4
frequency = 12 observations per 1 yearns   A: x5   B: x4   C: x4
frequency = 12 observations per 1 yearns   A: x5   B: x5   C: x5
frequency = 12 observations per 1 year
frequency = 12 observations per 1 yearns   A: x6   B: x6   C: x5
frequency = 12 observations per 1 yearns   A: x7   B: x6   C: x6
frequency = 12 observations per 1 yearns   A: x7   B: x7   C: x6
frequency = 12 observations per 1 yearns   A: x8   B: x7   C: x7
frequency = 12 observations per 1 year
frequency = 12 observations per 1 yearns   A: x9   B: x8   C: x8
frequency = 12 observations per 1 yearns   A: x9   B: x9   C: x8
frequency = 12 observations per 1 yearns   A: x10   B: x9   C: x9
frequency = 12 observations per 1 yearns   A: x10   B: x10   C: x10
frequency = 12 observations per 1 yearns   A: x11   B: x10   C: x10
frequency = 12 observations per 1 year
frequency = 12 observations per 1 yearns   A: x12   B: x11   C: x11
frequency = 12 observations per 1 yearns   A: x12   B: x12   C: x11
frequency = 12 observations per 1 yearns   A: x13   B: x12   C: x12
frequency = 12 observations per 1 yearns   A: x13   B: x13   C: x12
frequency = 12 observations per 1 yearns   A: x14   B: x13   C: x13
frequency = 12 observations per 1 yearns   A: x14   B: x14   C: x14
frequency = 12 observations per 1 yearns   A: x15   B: x14   C: x14
frequency = 12 observations per 1 year
frequency = 12 observations per 1 yearns   A: x16   B: x15   C: x15
frequency = 12 observations per 1 yearns   A: x16   B: x16   C: x15
frequency = 12 observations per 1 yearns   A: x17   B: x16   C: x16
frequency = 12 observations per 1 yearns   A: x17   B: x17   C: x16
frequency = 12 observations per 1 yearns   A: x18   B: x17   C: x17
frequency = 12 observations per 1 yearns   A: x18   B: x18   C: x18
frequency = 12 observations per 1 year
There were issues with some computations   A: x19   B: x19   C: x19