Class: NanoGPT::Web::WebTrainer

Inherits:
Object
  • Object
show all
Defined in:
lib/nano_gpt/web/web_trainer.rb

Overview

Training loop with web dashboard hooks Composes with the same model/data_loader/config as Trainer but adds stop flag checking, metric recording, and SSE broadcasting

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(model:, data_loader:, config:, training_state:, metrics_store:, sse_notifier:) ⇒ WebTrainer

Returns a new instance of WebTrainer.



13
14
15
16
17
18
19
20
21
22
23
24
25
26
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 13

def initialize(model:, data_loader:, config:, training_state:, metrics_store:, sse_notifier:)
  @model = model
  @data_loader = data_loader
  @config = config.is_a?(Hash) ? config.transform_keys(&:to_sym) : config.to_h
  @training_state = training_state
  @metrics_store = metrics_store
  @sse_notifier = sse_notifier

  @iter_num = 0
  @best_val_loss = Float::INFINITY

  setup_optimizer
  setup_lr_scheduler
end

Instance Attribute Details

#best_val_loss ⇒ Object (readonly)

Returns the value of attribute best_val_loss.



11
12
13
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 11

def best_val_loss
  @best_val_loss
end

#config ⇒ Object (readonly)

Returns the value of attribute config.



11
12
13
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 11

def config
  @config
end

#iter_num ⇒ Object (readonly)

Returns the value of attribute iter_num.



11
12
13
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 11

def iter_num
  @iter_num
end

#model ⇒ Object (readonly)

Returns the value of attribute model.



11
12
13
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 11

def model
  @model
end

Instance Method Details

#load_checkpoint(path) ⇒ Object 



116
117
118
119
120
121
122
123
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 116

def load_checkpoint(path)
  checkpoint = Torch.load(path)
  @model.load_state_dict(checkpoint["model"])
  @iter_num = checkpoint["iter_num"]
  @best_val_loss = checkpoint["best_val_loss"]
  setup_optimizer
  checkpoint
end

#train(run_id) ⇒ Object 



28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
 
# File 'lib/nano_gpt/web/web_trainer.rb', line 28

def train(run_id)
  @run_id = run_id
  @training_state.update(status: "running", run_id: run_id, max_iters: @config[:max_iters])

  puts "Starting training... max_iters=#{@config[:max_iters]} eval_interval=#{@config[:eval_interval]} eval_iters=#{@config[:eval_iters]} device=#{@config[:device]}"

  @model.train
  x, y = @data_loader.get_batch(:train)
  t0 = Time.now

  while @iter_num <= @config[:max_iters]
    # Check stop flag
    if @training_state.stop_requested?
      save_checkpoint("final")
      @training_state.update(status: "stopped")
      @metrics_store.update_run(run_id, status: "stopped", stopped_at: Time.now.iso8601,
        checkpoint_path: checkpoint_path("final"))
      @sse_notifier.broadcast(type: "status", data: @training_state.to_h)
      return
    end

    lr = @config[:decay_lr] ? @lr_scheduler.step(@optimizer, @iter_num) : @config[:learning_rate]

    if @iter_num % @config[:eval_interval] == 0
      puts "iter #{@iter_num}: running eval..."
      losses = estimate_loss
      val_loss = losses[:val]
      train_loss = losses[:train]
      puts "iter #{@iter_num}: eval done - train_loss=#{format('%.4f', train_loss)} val_loss=#{format('%.4f', val_loss)}"

      @training_state.update(best_val_loss: [@best_val_loss, val_loss].min)
      @metrics_store.record_metrics(run_id, @iter_num, { val_loss: val_loss, eval_train_loss: train_loss })
      @sse_notifier.broadcast(type: "eval", data: {
        iteration: @iter_num, val_loss: val_loss, train_loss: train_loss
      })

      if val_loss < @best_val_loss || @config[:always_save_checkpoint]
        @best_val_loss = [val_loss, @best_val_loss].min
        save_checkpoint("best") if @iter_num > 0
        @metrics_store.update_run(run_id,
          best_val_loss: @best_val_loss,
          checkpoint_path: checkpoint_path("best")
        )
      end
    end

    break if @iter_num == 0 && @config[:eval_only]

    @optimizer.zero_grad

    accumulated_loss = 0.0
    @config[:gradient_accumulation_steps].times do
      _logits, loss = @model.call(x, targets: y)
      loss = loss / @config[:gradient_accumulation_steps]
      accumulated_loss += loss.item
      loss.backward
      x, y = @data_loader.get_batch(:train)
    end

    clip_grad_norm(@model.parameters, @config[:grad_clip]) if @config[:grad_clip] > 0.0
    @optimizer.step

    t1 = Time.now
    dt = t1 - t0
    t0 = t1

    if @iter_num % @config[:log_interval] == 0
      @training_state.update(
        current_iter: @iter_num,
        current_loss: accumulated_loss
      )
      @metrics_store.record_metrics(run_id, @iter_num, { train_loss: accumulated_loss, lr: lr, iter_time_ms: dt * 1000 })
      @metrics_store.update_run(run_id, current_iter: @iter_num)
      @sse_notifier.broadcast(type: "train", data: {
        iteration: @iter_num, loss: accumulated_loss, lr: lr, time_ms: (dt * 1000).round(2)
      })
    end

    @iter_num += 1
  end

  save_checkpoint("final")
  @training_state.update(status: "completed")
  @metrics_store.update_run(run_id, status: "completed", stopped_at: Time.now.iso8601,
    checkpoint_path: checkpoint_path("best"))
  @sse_notifier.broadcast(type: "status", data: @training_state.to_h)
end