Compute F1 metric. If loss is "categorical_crossentropy"
, number of targets must be 2. If
loss is "binary_crossentropy"
and number of targets > 1, will flatten y_true
and y_pred
matrices
to a single vector (rather than computing separate F1 scores for each class).
Examples
if (FALSE) { # reticulate::py_module_available("tensorflow")
y_true <- c(1,0,0,1,1,0,1,0,0)
y_pred <- c(0.9,0.05,0.05,0.9,0.05,0.05,0.9,0.05,0.05)
f1_metric <- f1_wrapper(3L, "binary_crossentropy")
f1_metric$update_state(y_true, y_pred)
f1_metric$result()
# add metric to a model
num_targets <- 1
model <- create_model_lstm_cnn(maxlen = 20,
layer_lstm = 8,
bal_acc = FALSE,
last_layer_activation = "sigmoid",
loss_fn = "binary_crossentropy",
layer_dense = c(8, num_targets))
f1_metric <- f1_wrapper(num_targets, loss = model$loss)
model %>% keras::compile(loss = model$loss,
optimizer = model$optimizer,
metrics = c(model$metrics, f1_metric))
}