add division lines for print outs

Author: kchu25

src/_training/train.jl

@@ -105,7 +105,8 @@ train_model(model, opt, train_dl, val_dl, ydim; compiled_loss=my_compiled_loss)
 """
 function train_model(model, opt_state, train_dl, val_dl, output_dim;
                      max_epochs=50, patience=10, min_delta=1e-4, print_every=100,
-                     test_set=false, compute_loss=nothing, compiled_loss=masked_mse)
+                     test_set=false, compute_loss=nothing, compiled_loss=masked_mse,
+                     show_header=true)
 
     # Build compute_loss from compiled_loss if not provided directly
     if isnothing(compute_loss) && !isnothing(compiled_loss)
@@ -127,10 +128,12 @@ function train_model(model, opt_state, train_dl, val_dl, output_dim;
     val_losses = DEFAULT_FLOAT_TYPE[]
     val_r2_scores = DEFAULT_FLOAT_TYPE[]
 
-    vprintln(VERBOSITY_NORMAL, "Starting training for up to $max_epochs epochs...")
-    vprintln(VERBOSITY_NORMAL, "Early stopping: patience=$patience, min_delta=$min_delta")
-    vprintln(VERBOSITY_NORMAL, "Batch size: $(train_dl.batchsize), Total batches per epoch: $(length(train_dl))")
-    vprintln(VERBOSITY_NORMAL, "-" ^ 50)
+    if show_header
+        vprintln(VERBOSITY_NORMAL, "Starting training for up to $max_epochs epochs...")
+        vprintln(VERBOSITY_NORMAL, "Early stopping: patience=$patience, min_delta=$min_delta")
+        vprintln(VERBOSITY_NORMAL, "Batch size: $(train_dl.batchsize), Total batches per epoch: $(length(train_dl))")
+        vprintln(VERBOSITY_NORMAL, "-" ^ 50)
+    end
 
     for epoch in 1:max_epochs
         # Train one epoch

src/final_and_code/_helpers.jl

@@ -41,7 +41,7 @@ end
 
 """Train final model and load best weights. Returns: (trained_model, stats)"""
 function _train_final_model!(setup, dl_train, dl_test; max_epochs=50, patience=10, print_every=100)
-    vprintln(VERBOSITY_NORMAL, "🎯 Training final model (train+val combined)...")
+    print_phase_banner(VERBOSITY_QUIET, "🎯 FINAL MODEL TRAINING (train+val combined)")
     best_state, stats = train_model(setup.model, setup.opt_state, dl_train, dl_test, setup.Ydim;
                                     max_epochs, patience, print_every, test_set=true, compiled_loss=setup.compiled_loss)
 

src/final_and_code/train_code_processor.jl

@@ -33,10 +33,9 @@ function train_code_processor(model, dataloader, proc_wrap;
     processor, opt_state = _init_processor(proc_wrap, model.hp, use_hard_mask, seed)
 
     # Print header
-    vprintln(VERBOSITY_NORMAL, "=" ^ 60)
-    vprintln(VERBOSITY_NORMAL, "🔧 Training code processor (arch: $(proc_wrap.arch_type), use_hard_mask: $use_hard_mask)")
+    print_phase_banner(VERBOSITY_QUIET, "🔧 CODE PROCESSOR TRAINING")
+    vprintln(VERBOSITY_NORMAL, "Arch: $(proc_wrap.arch_type), use_hard_mask: $use_hard_mask")
     vprintln(VERBOSITY_NORMAL, "Seed: $seed, Epochs: $max_epochs")
-    vprintln(VERBOSITY_NORMAL, "=" ^ 60)
 
     # Training loop
     step = Ref(0)
@@ -50,10 +49,10 @@ function train_code_processor(model, dataloader, proc_wrap;
     end
 
     # Print summary
-    vprintln(VERBOSITY_QUIET, "=" ^ 60)
+    vprintln(VERBOSITY_QUIET, "-" ^ 60)
     vprintln(VERBOSITY_QUIET, "Code processor training complete!")
     vprintln(VERBOSITY_QUIET, "Final loss: $(round(loss_history[end], digits=6))")
-    vprintln(VERBOSITY_QUIET, "=" ^ 60)
+    vprintln(VERBOSITY_QUIET, "-" ^ 60)
 
     return processor, loss_history
 end

src/tune/_helpers.jl

@@ -45,7 +45,8 @@ function _run_trial(trial, raw_data, create_model, randomize_batchsize,
                                                rng=MersenneTwister(rand(Random.GLOBAL_RNG, 1:typemax(Int))))
 
     model_state, stats = train_model(setup.model, setup.opt_state, dl_train, dl_val, setup.Ydim;
-                                     max_epochs, patience, print_every, compiled_loss=setup.compiled_loss)
+                                     max_epochs, patience, print_every, compiled_loss=setup.compiled_loss,
+                                     show_header=false)
 
     r2, loss = stats[:best_r2], stats[:best_val_loss]
     num_params = sum(length, Flux.trainables(setup.model))

src/tune/tuning.jl

@@ -15,6 +15,8 @@ function tune_hyperparameters(raw_data, create_model::Function;
                               print_every=100, save_folder=nothing, use_cuda=true,
                               loss_spec=(loss=Flux.mse, agg=StatsBase.mean), model_kwargs...)
 
+    print_phase_banner(VERBOSITY_QUIET, "🔍 HYPERPARAMETER TUNING — $n_trials trials")
+
     results = DataFrame(seed=Int[], best_r2=DEFAULT_FLOAT_TYPE[], val_loss=DEFAULT_FLOAT_TYPE[], num_params=Int[])
     best_r2, best_model, best_seed, best_batch = -Inf, nothing, nothing, nothing
     save_file = isnothing(save_folder) ? nothing : _setup_save_file(save_folder)

src/verbosity.jl

@@ -79,3 +79,22 @@ function vprint(level::Integer, args...)
     _should_log(level) && print(args...)
     nothing
 end
+
+"""
+    print_phase_banner(level, title; width=60)
+
+Print a centered, full-width banner bounded by long-dash rules announcing a
+high-level phase (e.g. tuning, final training, code-processor training).
+Gated by `level` exactly like [`vprintln`](@ref); defaults to a banner width
+of `width` characters and is preceded by a blank line so it stands out.
+"""
+function print_phase_banner(level::Integer, title::AbstractString; width::Integer=60)
+    _should_log(level) || return nothing
+    rule = "-" ^ width
+    pad = max(0, (width - length(title)) ÷ 2)
+    println()
+    println(rule)
+    println(" " ^ pad, title)
+    println(rule)
+    nothing
+end