test_that("Calculations are correct - two class", {
  # Powers paper
  lst <- data_powers()
  df_2_1 <- lst$df_2_1

  expect_equal(
    f_meas_vec(truth = df_2_1$truth, estimate = df_2_1$prediction),
    0.5882353,
    tolerance = 0.0001
  )
})

test_that("Calculations are correct - three class", {
  multi_ex <- data_three_by_three()
  micro <- data_three_by_three_micro()

  expect_equal(
    f_meas(multi_ex, estimator = "macro")[[".estimate"]],
    macro_metric(f_meas_binary)
  )
  expect_equal(
    f_meas(multi_ex, estimator = "macro_weighted")[[".estimate"]],
    macro_weighted_metric(f_meas_binary)
  )
  expect_equal(
    f_meas(multi_ex, estimator = "micro")[[".estimate"]],
    0.6
  )
})

test_that("All interfaces gives the same results", {
  lst <- data_altman()
  pathology <- lst$pathology
  path_tbl <- lst$path_tbl
  path_mat <- unclass(path_tbl)

  exp <- f_meas_vec(pathology$pathology, pathology$scan)

  expect_identical(
    f_meas(path_tbl)[[".estimate"]],
    exp
  )
  expect_identical(
    f_meas(path_mat)[[".estimate"]],
    exp
  )
  expect_identical(
    f_meas(pathology, truth = pathology, estimate = scan)[[".estimate"]],
    exp
  )
})

test_that("Calculations handles NAs", {
  # Powers paper
  lst <- data_powers()
  df_2_1 <- lst$df_2_1

  expect_equal(
    f_meas(df_2_1, truth = truth, estimate = pred_na)[[".estimate"]],
    0.5652174,
    tolerance = 0.0001
  )
})

test_that("Case weights calculations are correct", {
  df <- data.frame(
    truth = factor(c("x", "x", "y"), levels = c("x", "y")),
    estimate = factor(c("x", "y", "x"), levels = c("x", "y")),
    case_weights = c(1L, 1L, 2L)
  )

  expect_identical(
    f_meas_vec(
      truth = df$truth,
      estimate = df$estimate,
      case_weights = df$case_weights
    ),
    2 / 5
  )

  py_res <- read_pydata("py-f_meas")
  py_res_.5 <- read_pydata("py-f_meas_beta_.5")
  r_metric <- f_meas

  two_class_example$weights <- read_weights_two_class_example()

  expect_equal(
    r_metric(two_class_example, truth, predicted, case_weights = weights)[[
      ".estimate"
    ]],
    py_res$case_weight$binary
  )
  expect_equal(
    r_metric(
      two_class_example,
      truth,
      predicted,
      case_weights = weights,
      beta = .5
    )[[".estimate"]],
    py_res_.5$case_weight$binary
  )

  py_res <- read_pydata("py-f_meas")
  py_res_.5 <- read_pydata("py-f_meas_beta_.5")
  r_metric <- f_meas

  hpc_cv$weights <- read_weights_hpc_cv()

  expect_equal(
    r_metric(hpc_cv, obs, pred, case_weights = weights)[[".estimate"]],
    py_res$case_weight$macro
  )
  expect_equal(
    r_metric(hpc_cv, obs, pred, estimator = "micro", case_weights = weights)[[
      ".estimate"
    ]],
    py_res$case_weight$micro
  )
  expect_equal(
    r_metric(
      hpc_cv,
      obs,
      pred,
      estimator = "macro_weighted",
      case_weights = weights
    )[[".estimate"]],
    py_res$case_weight$weighted
  )

  expect_equal(
    r_metric(hpc_cv, obs, pred, beta = .5, case_weights = weights)[[
      ".estimate"
    ]],
    py_res_.5$case_weight$macro
  )
  expect_equal(
    r_metric(
      hpc_cv,
      obs,
      pred,
      estimator = "micro",
      beta = .5,
      case_weights = weights
    )[[".estimate"]],
    py_res_.5$case_weight$micro
  )
  expect_equal(
    r_metric(
      hpc_cv,
      obs,
      pred,
      estimator = "macro_weighted",
      beta = .5,
      case_weights = weights
    )[[".estimate"]],
    py_res_.5$case_weight$weighted
  )
})

test_that("work with class_pred input", {
  skip_if_not_installed("probably")

  cp_truth <- probably::as_class_pred(two_class_example$truth, which = 1)
  cp_estimate <- probably::as_class_pred(two_class_example$predicted, which = 2)

  fct_truth <- two_class_example$truth
  fct_truth[1] <- NA

  fct_estimate <- two_class_example$predicted
  fct_estimate[2] <- NA

  expect_identical(
    f_meas_vec(fct_truth, cp_estimate),
    f_meas_vec(fct_truth, fct_estimate)
  )

  expect_identical(
    f_meas_vec(fct_truth, cp_estimate, na_rm = FALSE),
    NA_real_
  )

  expect_snapshot(
    error = TRUE,
    f_meas_vec(cp_truth, cp_estimate)
  )
})

test_that("works with hardhat case weights", {
  lst <- data_altman()
  df <- lst$pathology
  imp_wgt <- hardhat::importance_weights(seq_len(nrow(df)))
  freq_wgt <- hardhat::frequency_weights(seq_len(nrow(df)))

  expect_no_error(
    f_meas_vec(df$pathology, df$scan, case_weights = imp_wgt)
  )

  expect_no_error(
    f_meas_vec(df$pathology, df$scan, case_weights = freq_wgt)
  )
})

test_that("na_rm argument check", {
  expect_snapshot(
    error = TRUE,
    f_meas_vec(1, 1, na_rm = "yes")
  )
})

test_that("sklearn equivalent", {
  py_res <- read_pydata("py-f_meas")
  py_res_.5 <- read_pydata("py-f_meas_beta_.5")
  r_metric <- f_meas

  expect_equal(
    r_metric(two_class_example, truth, predicted)[[".estimate"]],
    py_res$binary
  )
  expect_equal(
    r_metric(two_class_example, truth, predicted, beta = .5)[[".estimate"]],
    py_res_.5$binary
  )

  py_res <- read_pydata("py-f_meas")
  py_res_.5 <- read_pydata("py-f_meas_beta_.5")
  r_metric <- f_meas

  expect_equal(
    r_metric(hpc_cv, obs, pred)[[".estimate"]],
    py_res$macro
  )
  expect_equal(
    r_metric(hpc_cv, obs, pred, estimator = "micro")[[".estimate"]],
    py_res$micro
  )
  expect_equal(
    r_metric(hpc_cv, obs, pred, estimator = "macro_weighted")[[".estimate"]],
    py_res$weighted
  )

  expect_equal(
    r_metric(hpc_cv, obs, pred, beta = .5)[[".estimate"]],
    py_res_.5$macro
  )
  expect_equal(
    r_metric(hpc_cv, obs, pred, estimator = "micro", beta = .5)[[".estimate"]],
    py_res_.5$micro
  )
  expect_equal(
    r_metric(hpc_cv, obs, pred, "macro_weighted", beta = .5)[[".estimate"]],
    py_res_.5$weighted
  )
})

test_that("`event_level = 'second'` works", {
  lst <- data_powers()
  df <- lst$df_2_1

  df_rev <- df
  df_rev$truth <- stats::relevel(df_rev$truth, "Irrelevant")
  df_rev$prediction <- stats::relevel(df_rev$prediction, "Irrelevant")

  expect_equal(
    f_meas_vec(df$truth, df$prediction),
    f_meas_vec(df_rev$truth, df_rev$prediction, event_level = "second")
  )
})

test_that("`NA` values propagate from binary `precision()` (#77)", {
  truth <- factor(c(rep("a", 2), rep("b", 2)))
  estimate <- factor(rep("b", length(truth)), levels(truth))

  expect_snapshot({
    out <- precision_vec(truth, estimate)
    expect <- f_meas_vec(truth, estimate)
  })

  expect_identical(out, expect)
})

test_that("`NA` values propagate from binary `recall()` (#77)", {
  estimate <- factor(c(rep("a", 2), rep("b", 2)))
  truth <- factor(rep("b", length(estimate)), levels(estimate))

  expect_snapshot({
    out <- recall_vec(truth, estimate)
    expect <- f_meas_vec(truth, estimate)
  })

  expect_identical(out, expect)
})

test_that("Binary returns `NA` with a warning when results are undefined (#98)", {
  # Recall - (tp + fn = 0)
  levels <- c("a", "b")
  truth <- factor(c("b", "b"), levels = levels)
  estimate <- factor(c("a", "b"), levels = levels)

  expect_snapshot(
    out <- f_meas_vec(truth, estimate)
  )
  expect_identical(out, NA_real_)

  # Precision - (tp + fp = 0)
  levels <- c("a", "b")
  truth <- factor("a", levels = levels)
  estimate <- factor("b", levels = levels)

  expect_snapshot(
    out <- f_meas_vec(truth, estimate)
  )
  expect_identical(out, NA_real_)
})

test_that("Multiclass returns averaged value a warning when results is undefined (#98)", {
  # recall - (tp + fn = 0)
  levels <- c("a", "b", "c")

  truth <- factor(c("a", "b", "b"), levels = levels)
  estimate <- factor(c("a", "b", "c"), levels = levels)

  expect_snapshot(
    out <- f_meas_vec(truth, estimate)
  )

  expect_identical(out, 5 / 6)
})

test_that("`NA` is still returned if there are some undefined values but `na.rm = FALSE`", {
  levels <- c("a", "b", "c")
  truth <- factor(c("a", "b", "b"), levels = levels)
  estimate <- factor(c("a", NA, "c"), levels = levels)
  expect_equal(f_meas_vec(truth, estimate, na_rm = FALSE), NA_real_)
  expect_warning(f_meas_vec(truth, estimate, na_rm = FALSE), NA)
})

test_that("bad argument check", {
  expect_snapshot(
    error = TRUE,
    f_meas_vec(1, 1, beta = "yes")
  )
})

test_that("range values are correct", {
  direction <- metric_direction(f_meas)
  range <- metric_range(f_meas)
  perfect <- ifelse(direction == "minimize", range[1], range[2])
  worst <- ifelse(direction == "minimize", range[2], range[1])

  df <- tibble::tibble(
    truth = factor(c("A", "A", "B", "B", "B")),
    off = factor(c("B", "B", "A", "A", "A"))
  )

  expect_equal(
    f_meas_vec(df$truth, df$truth),
    perfect
  )

  if (direction == "minimize") {
    expect_gt(f_meas_vec(df$truth, df$off), perfect)
    expect_lte(f_meas_vec(df$truth, df$off), worst)
  }
  if (direction == "maximize") {
    expect_lt(f_meas_vec(df$truth, df$off), perfect)
    expect_gte(f_meas_vec(df$truth, df$off), worst)
  }
})