Class: Secryst::Trainer

Inherits:
Object
  • Object
show all
Defined in:
lib/secryst/trainer.rb

Instance Method Summary collapse

Constructor Details

#initialize(model:, batch_size:, lr:, data_input:, data_target:, hyperparameters:, max_epochs: nil, log_interval: 1, checkpoint_every:, checkpoint_dir:, scheduler_step_size:, gamma:) ⇒ Trainer

Returns a new instance of Trainer.



4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
 
# File 'lib/secryst/trainer.rb', line 4

def initialize(
  model:,
  batch_size:,
  lr:,
  data_input:,
  data_target:,
  hyperparameters:,
  max_epochs: nil,
  log_interval: 1,
  checkpoint_every:,
  checkpoint_dir:,
  scheduler_step_size:,
  gamma:
)
  @data_input = File.readlines(data_input, chomp: true)
  @data_target = File.readlines(data_target, chomp: true)

  @device = "cpu"
  @lr = lr
  @scheduler_step_size = scheduler_step_size
  @gamma = gamma
  @batch_size = batch_size
  @model_name = model
  @max_epochs = max_epochs
  @log_interval = log_interval
  @checkpoint_every = checkpoint_every
  @checkpoint_dir = checkpoint_dir
  FileUtils.mkdir_p(@checkpoint_dir)
  generate_vocabs_and_data
  save_vocabs

  case model
  when 'transformer'
    @model = Secryst::Transformer.new(hyperparameters.merge({
      input_vocab_size: @input_vocab.length,
      target_vocab_size: @target_vocab.length,
    }))
  else
    raise ArgumentError, 'Only transformer model is currently supported'
  end
end

Instance Method Details

#train ⇒ Object 



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
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
 
# File 'lib/secryst/trainer.rb', line 46

def train
  best_model = nil
  best_val_loss = 1.0/0.0 # infinity

  return unless @model_name == 'transformer'

  criterion = Torch::NN::CrossEntropyLoss.new(ignore_index: index_of('<pad>')).to(@device)
  optimizer = Torch::Optim::SGD.new(@model.parameters, lr: @lr)
  scheduler = Torch::Optim::LRScheduler::StepLR.new(optimizer, step_size: @scheduler_step_size, gamma: @gamma)

  total_loss = 0.0
  start_time = Time.now
  ntokens = @target_vocab.length
  epoch = 0

  loop do
    epoch_start_time = Time.now
    @model.train
    @train_data.each.with_index do |batch, i|
      inputs, targets, decoder_inputs, src_mask, tgt_mask, memory_mask = batch
      inputs = Torch.tensor(inputs).t
      decoder_inputs = Torch.tensor(decoder_inputs).t
      targets = Torch.tensor(targets).t
      src_key_padding_mask = inputs.t.eq(1)
      tgt_key_padding_mask = decoder_inputs.t.eq(1)

      optimizer.zero_grad
      opts = {
        # src_mask: src_mask,
        tgt_mask: tgt_mask,
        # memory_mask: memory_mask,
        src_key_padding_mask: src_key_padding_mask,
        tgt_key_padding_mask: tgt_key_padding_mask,
        memory_key_padding_mask: src_key_padding_mask,
      }
      output = @model.call(inputs, decoder_inputs, opts)
      loss = criterion.call(output.transpose(0,1).reshape(-1, ntokens), targets.t.view(-1))
      loss.backward
      ClipGradNorm.clip_grad_norm(@model.parameters, max_norm: 0.5)
      optimizer.step

      # puts "i[#{i}] loss: #{loss}"
      total_loss += loss.item()
      if ( (i + 1) % @log_interval == 0 )
        cur_loss = total_loss / @log_interval
        elapsed = Time.now - start_time
        puts "| epoch #{epoch} | #{i + 1}/#{@train_data.length} batches | "\
              "lr #{scheduler.get_lr()[0].round(4)} | ms/batch #{(1000*elapsed.to_f / @log_interval).round} | "\
              "loss #{cur_loss.round(5)} | ppl #{Math.exp(cur_loss).round(5)}"
        total_loss = 0
        start_time = Time.now
      end
    end

    if epoch > 0 && epoch % @checkpoint_every == 0
      puts ">> Saving checkpoint '#{@checkpoint_dir}/checkpoint-#{epoch}.pth'"
      Torch.save(@model.state_dict, "#{@checkpoint_dir}/checkpoint-#{epoch}.pth")
    end

    # Evaluate
    @model.eval()
    total_loss = 0.0
    Torch.no_grad do
      @eval_data.each.with_index do |batch, i|
        inputs, targets, decoder_inputs, src_mask, tgt_mask, memory_mask = batch
        inputs = Torch.tensor(inputs).t
        decoder_inputs = Torch.tensor(decoder_inputs).t
        targets = Torch.tensor(targets).t
        src_key_padding_mask = inputs.t.eq(1)
        tgt_key_padding_mask = decoder_inputs.t.eq(1)

        opts = {
          # src_mask: src_mask,
          tgt_mask: tgt_mask,
          # memory_mask: memory_mask,
          src_key_padding_mask: src_key_padding_mask,
          tgt_key_padding_mask: tgt_key_padding_mask,
          memory_key_padding_mask: src_key_padding_mask,
        }
        output = @model.call(inputs, decoder_inputs, **opts)
        output_flat = output.transpose(0,1).reshape(-1, ntokens)

        total_loss += criterion.call(output_flat, targets.t.view(-1)).item
      end
      total_loss = total_loss / @eval_data.length
      puts('-' * 89)
      puts "| end of epoch #{epoch} | time: #{(Time.now - epoch_start_time).round(3)}s | "\
              " valid loss #{total_loss.round(5)} | valid ppl #{Math.exp(total_loss).round(5)} "
      puts('-' * 89)
      if total_loss < best_val_loss
        best_model = @model
        best_val_loss = total_loss
      end
    end
    scheduler.step

    epoch += 1
    break if @max_epochs && @max_epochs < epoch
  end
end