File size: 32,300 Bytes
e09b20c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
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
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
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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
import itertools
import random
import os
import time
import torch
import math
import inspect
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import json
from tqdm import tqdm
import loguru
import shutil


# model architecture


class Head(nn.Module):
    """One head of self-attention."""

    def __init__(self, head_size, n_embd, dropout, block_size):
        super().__init__()
        self.head_size = head_size
        self.key = nn.Linear(n_embd, head_size, bias=False)
        self.query = nn.Linear(n_embd, head_size, bias=False)
        self.value = nn.Linear(n_embd, head_size, bias=False)
        self.pos_embeds = nn.Embedding(block_size, head_size)
        self.attn_dropout = nn.Dropout(dropout)
        self.register_buffer(
            "pos_embeds_indices",
            torch.arange(0, block_size, 1, dtype=torch.int64)
        )
        self.register_buffer(
            "causal_mask",
            (torch.tril(torch.ones((block_size, block_size), dtype=torch.bool)) == False)
        )
    # ======


    def forward(self, x):
        B, T, C = x.shape
        k = self.key(x)  # (B, T, C)
        q = self.query(x)  # (B, T, C)
        v = self.value(x)  # (B, T, C)
        r = self.pos_embeds(self.pos_embeds_indices[-T:])  # (T, C)
        s_rel = q @ r.t()  # (B, T, T)
        s_rel = F.pad(s_rel, (1, 0))  # (B, T, 1+T)
        s_rel = s_rel.reshape(B, s_rel.shape[-1], s_rel.shape[-2])  # (B, 1+T, T)
        s_rel = s_rel[:, 1:, :]  # (B, T, T)
        attn = ((q @ k.transpose(-2, -1)) + s_rel) * (self.head_size**-0.5)  # (B, T, T)
        attn = attn.masked_fill(self.causal_mask[:T, :T], float("-inf"))  # (B, T, T)
        attn = F.softmax(attn, dim=-1)  # (B, T, T)
        attn = self.attn_dropout(attn)  # (B, T, T)
        out = attn @ v  # (B, T, C)

        return out
    # ======
# ======


class MultiHeadAttention(nn.Module):
    """Multiple heads of self-attention in parallel."""

    def __init__(self, num_heads, head_size, n_embd, dropout, block_size):
        super().__init__()
        self.heads = nn.ModuleList([Head(head_size, n_embd, dropout, block_size) for _ in range(num_heads)])
        self.proj = nn.Linear(n_embd, n_embd)
        self.dropout = nn.Dropout(dropout)
    # ======


    def forward(self, x):
        out = torch.cat([h(x) for h in self.heads], dim=-1)
        out = self.dropout(self.proj(out))
        return out
    # ======
# ======


class FeedForward(nn.Module):
    """A simple linear layer followed by a non-linearity."""

    def __init__(self, n_embd, ff_width, dropout):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(n_embd, ff_width * n_embd, bias=False),
            nn.SiLU(),
            nn.Linear(ff_width * n_embd, n_embd, bias=False),
            nn.Dropout(dropout),
        )
    # ======


    def forward(self, x):
        return self.net(x)
    # ======
# ======


class Block(nn.Module):
    """Transformer block: communication followed by feedforward."""

    def __init__(self, n_embd, n_head, ff_width, dropout, block_size):
        super().__init__()
        head_size = n_embd // n_head
        self.sa = MultiHeadAttention(n_head, head_size, n_embd, dropout, block_size)
        self.ffwd = FeedForward(n_embd, ff_width, dropout)
        self.ln1 = nn.RMSNorm(normalized_shape=n_embd, eps=1e-5)
        self.ln2 = nn.RMSNorm(normalized_shape=n_embd, eps=1e-5)
    # ======


    def forward(self, x):
        x = x + self.sa(self.ln1(x))
        x = x + self.ffwd(self.ln2(x))
        return x
    # ======
# ======


class GPT(nn.Module):
    """Transformer architecture for flowshop scheduling."""

    def __init__(self, intermediate_schedules, vocab_size, n_embd, n_head, n_layer, dropout, ff_width, block_size):
        super().__init__()
        self.jobs_embedding_table = nn.Embedding(vocab_size, n_embd)
        self.blocks = nn.Sequential(*[Block(n_embd=n_embd, n_head=n_head, ff_width=ff_width, dropout=dropout, block_size=block_size) for _ in range(n_layer)])
        self.ln_f = nn.LayerNorm(n_embd)
        self.makespan_head = nn.Linear(n_embd, 1)
        self.intermediate_schedules = intermediate_schedules
    # ======


    def forward(self, idx, targets):
        B, T = idx.shape

        # idx and targets are both (B,T,) tensor of integers
        x = self.jobs_embedding_table(idx)  # (B,T,C,)
        x = self.blocks(x)  # (B,T,C,)
        x = self.ln_f(x)  # (B,T,C,)
        if not self.intermediate_schedules:
            x = x[:, -1, :].squeeze()  # (B,C,)
            makespans = self.makespan_head(x).squeeze()  # (B,)
            targets = targets[:,-1].squeeze() # (B,)
        else:
            x = x.reshape(B*T, -1)  # (B*T,C)
            makespans = self.makespan_head(x).squeeze()  # (B*T,)
            targets = targets.reshape(B*T)  # (B*T,)
        # ======
        loss = F.smooth_l1_loss(makespans, targets, beta=0.1)
        return makespans, loss
    # ======


    def generate(self, job_embeds):
        """

        Function to generate Transformer predicted makespan from input schedule of job embeddings.

        Input:

            - job_embeds: a (T,C,) tensor of job embeddings

        Output:

            - makespan: a scalar tensor of the predicted makespan

        """

        job_embeds = job_embeds.unsqueeze(0) # (1,T,C)
        x = self.blocks(job_embeds)  # (1,T,C,)
        x = self.ln_f(x)  # (1,T,C,)
        makespans = self.makespan_head(x).squeeze()  # (T,)
        makespan = makespans[-1]

        return makespan
    # ======


def train(

    testing: bool,

    seed: int,

    data_dir: str,

    n_embd: int,

    n_head: int,

    n_layer: int,

    intermediate_schedules: bool,

    dropout: float,

    ff_width: int,

    train_batch_size: int,

    val_batch_size: int,

    nb_epochs: int,

    early_stopping_patience: int,

    checkpoint_interval_ratio: float,

    decay_lr: bool,

    lr_partitions_ratios: list[float],

    init_lr: float,

    max_lr: float,

    min_lr: float,

    lr_warmup_iters_ratio: float,

    lr_decay_iters_ratio: float,

    beta1: float,

    beta2: float,

    weight_decay: float,

    grad_clip: float,

    compile: bool,

    compile_mode: str,

    save_only_last_checkpoint: bool,

    output_dir: str,

):

    os.makedirs(output_dir, exist_ok=True)

    # check if experiment termination flag file exists
    if not testing:
        if os.path.exists(os.path.join(output_dir, ".terminated_phase1")):
            print("Phase 1 already terminated. Exiting...")
            return
        # ======
        if not os.path.exists(os.path.join(output_dir, "viz_train.ipynb")):
            shutil.copy("viz_train.ipynb", output_dir)
        # ======
    else:

        # delete all the files that are created by this script like log files, batch_losses.npy and so on ...
        files_to_delete = [
            "train.log",
            "train_parameters.json",
            "batch_losses.npy",
            "last_batch_loss_idx.npy",
            "val_losses.npy",
            "last_val_loss_idx.npy",
            "train_pbar_epoch.log",
            "train_pbar_val.log",
            ".terminated_phase1",
            "viz_train.ipynb",
        ]
        for f in files_to_delete:
            f_path = os.path.join(output_dir, f)
            if os.path.exists(f_path): os.remove(f_path)
        # ======
        checkpoints_dir = os.path.join(output_dir, "checkpoints")
        if os.path.exists(checkpoints_dir): shutil.rmtree(checkpoints_dir)
        shutil.copy("viz_train.ipynb", output_dir)
    # ======

    # check if GPU is available
    assert torch.cuda.is_available(), "This code requires a GPU to run. Please run it on a machine with a CUDA-compatible GPU."
    device = "cuda"

    # setup logging
    loguru.logger.add(os.path.join(output_dir, "train.log"))

    # set random seeds
    torch.manual_seed(seed)
    random.seed(seed)
    np.random.seed(seed)

    # setup model architecture parameters
    with open(os.path.join(data_dir, "metadata.json"), "r") as f:
        metadata = json.load(f)
    block_size = metadata["nb_jobs"]  # context window size
    vocab_size = metadata["nb_jobs"]  # vocabulary size
    n_embd = n_embd  # embedding dimension
    n_head = n_head  # number of attention heads
    assert n_embd % n_head == 0
    n_layer = n_layer  # number of transformer blocks
    intermediate_schedules = intermediate_schedules
    ff_width = ff_width

    # setup training parameters and utils
    train_batch_size = train_batch_size  # batch size for training
    val_batch_size = val_batch_size # batch size for validation
    nb_epochs = nb_epochs # number of pseudo-epochs to train for
    early_stopping_patience = early_stopping_patience  # number of epochs without improvement to trigger early stopping
    dropout = dropout


    class FlowshopDataset(torch.utils.data.Dataset):
        """Dataset for flowshop scheduling problem."""

        def __init__(self, dataset_path, split, load_in_memory=True):
            metadata_path = os.path.join(dataset_path, "metadata.json")
            with open(metadata_path, "r") as f:
                metadata = json.load(f)
            nb_samples = metadata["nb_samples"]
            nb_jobs = metadata["nb_jobs"]
            schedules_path = os.path.join(dataset_path, f"schedules_{split}.npy")
            schedules = np.lib.format.open_memmap(
                schedules_path, dtype=np.int32, mode="r", shape=(nb_samples, nb_jobs)
            )
            loguru.logger.info(f"Loaded schedules from {schedules_path} with shape {schedules.shape}")
            makespans_path = os.path.join(dataset_path, f"makespans_{split}.npy")
            makespans = np.lib.format.open_memmap(
                makespans_path, dtype=np.float32, mode="r", shape=(nb_samples,)
            )
            loguru.logger.info(f"Loaded makespans from {makespans_path} with shape {makespans.shape}")
            if load_in_memory:
                schedules = np.array(schedules)
                makespans = np.array(makespans)
            self.schedules = torch.from_numpy(schedules)
            self.makespans = torch.from_numpy(makespans)
        # ======


        def __len__(self):
            return len(self.schedules)
        # ======


        def __getitem__(self, idx):
            return self.schedules[idx], self.makespans[idx]
        # ======


    train_dataset = FlowshopDataset(data_dir, split="train", load_in_memory=True)
    train_data_loader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=train_batch_size,
        shuffle=False,
        drop_last=False,
    )
    ## log the shape of the items returned by the train_data_loader
    for schedules, makespans in train_data_loader:
        loguru.logger.info(f"schedules.shape: {schedules.shape}")
        loguru.logger.info(f"makespans.shape: {makespans.shape}")
        break
    nb_iters = nb_epochs * len(train_data_loader)
    checkpoint_interval = int(checkpoint_interval_ratio * len(train_data_loader))
    decay_lr = decay_lr
    lr_partitions_ratios = lr_partitions_ratios + [None]
    lr_partitions_iters = [int(r * nb_iters) for r in lr_partitions_ratios[:-1]]
    lr_partitions_iters = lr_partitions_iters + [nb_iters - sum(lr_partitions_iters)]
    assert sum(lr_partitions_iters) == nb_iters
    init_lr = init_lr #1e-4
    max_lr = max_lr #1e-3
    min_lr = min_lr #5*1e-5
    lr_warmup_iters_ratio = lr_warmup_iters_ratio #0.1
    lr_decay_iters_ratio = lr_decay_iters_ratio #0.95
    beta1 = beta1 # Adam beta1
    beta2 = beta2 # Adam beta2
    weight_decay = weight_decay # 1e-1  # weight decay
    grad_clip = grad_clip # 1.0  # gradient clipping value
    compile = compile
    compile_mode = compile_mode
    save_only_last_checkpoint = save_only_last_checkpoint


    def human_readable(num):
        """Define function to make large numbers of parameters human-readable."""

        magnitude = 0
        while abs(num) >= 1000:
            magnitude += 1
            num /= 1000.0
        return "%.0f%s" % (num, ["", "K", "M", "G", "T", "P"][magnitude])
    # ======


    def get_lr(it):
        """Get the learning rate for the current iteration."""

        i = 0
        tmp_it = it - lr_partitions_iters[i]
        while tmp_it >= 0:
            i += 1
            tmp_it -= lr_partitions_iters[i]

        lr_partition_iters = lr_partitions_iters[i]
        warmup_iters = int(lr_partition_iters * lr_warmup_iters_ratio)
        lr_decay_iters = int(lr_partition_iters * lr_decay_iters_ratio)

        it = it - sum(lr_partitions_iters[:i])
    
        # 1) linear warmup for warmup_iters steps
        if it < warmup_iters:
            return (it / warmup_iters) * (max_lr - init_lr) + init_lr
        
        # 2) if it > lr_decay_iters, return min learning rate
        if it > lr_decay_iters:
            return min_lr
        
        # 3) in between, use cosine decay down to min learning rate
        decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
        assert 0 <= decay_ratio <= 1
        coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))  # coeff ranges 0..1
        return min_lr + coeff * (max_lr - min_lr)
    # ======


    # log parameters
    loguru.logger.info(f"data_dir: {data_dir}")
    loguru.logger.info(f"block_size: {block_size}")
    loguru.logger.info(f"vocab_size: {vocab_size}")
    loguru.logger.info(f"n_embd: {n_embd}")
    loguru.logger.info(f"n_head: {n_head}")
    loguru.logger.info(f"n_layer: {n_layer}")
    loguru.logger.info(f"ff_width: {ff_width}")
    loguru.logger.info(f"train_batch_size: {train_batch_size}")
    loguru.logger.info(f"val_batch_size: {val_batch_size}")
    loguru.logger.info(f"dropout: {dropout}")
    loguru.logger.info(f"nb_epochs: {nb_epochs}")
    loguru.logger.info(f"early_stopping_patience: {early_stopping_patience}")
    loguru.logger.info(f"nb_iters: {nb_iters}")
    loguru.logger.info(f"checkpoint_interval: {checkpoint_interval}")
    loguru.logger.info(f"decay_lr: {decay_lr}")
    loguru.logger.info(f"lr_partitions_ratios: {lr_partitions_ratios}")
    loguru.logger.info(f"lr_partitions_iters: {lr_partitions_iters}")
    loguru.logger.info(f"init_lr: {init_lr}")
    loguru.logger.info(f"max_lr: {max_lr}")
    loguru.logger.info(f"min_lr: {min_lr}")
    loguru.logger.info(f"lr_warmup_iters_ratio: {lr_warmup_iters_ratio}")
    loguru.logger.info(f"lr_decay_iters_ratio: {lr_decay_iters_ratio}")
    loguru.logger.info(f"beta1: {beta1}")
    loguru.logger.info(f"beta2: {beta2}")
    loguru.logger.info(f"weight_decay: {weight_decay}")
    loguru.logger.info(f"grad_clip: {grad_clip}")
    loguru.logger.info(f"compile: {compile}")
    loguru.logger.info(f"compile_mode: {compile_mode}")
    loguru.logger.info(f"intermediate_schedules: {intermediate_schedules}")
    loguru.logger.info(f"save_only_last_checkpoint: {save_only_last_checkpoint}")

    # save parameters into a train_parameters.json
    train_params = {
        "data_dir": data_dir,
        "block_size": block_size,
        "vocab_size": vocab_size,
        "n_embd": n_embd,
        "n_head": n_head,
        "n_layer": n_layer,
        "ff_width": ff_width,
        "train_batch_size": train_batch_size,
        "val_batch_size": val_batch_size,
        "dropout": dropout,
        "nb_epochs": nb_epochs,
        "early_stopping_patience": early_stopping_patience,
        "nb_iters": nb_iters,
        "checkpoint_interval": checkpoint_interval,
        "decay_lr": decay_lr,
        "lr_partitions_ratios": lr_partitions_ratios,
        "lr_partitions_iters": lr_partitions_iters,
        "init_lr": init_lr,
        "max_lr": max_lr,
        "min_lr": min_lr,
        "lr_warmup_iters_ratio": lr_warmup_iters_ratio,
        "lr_decay_iters_ratio": lr_decay_iters_ratio,
        "beta1": beta1,
        "beta2": beta2,
        "weight_decay": weight_decay,
        "grad_clip": grad_clip,
        "compile": compile,
        "compile_mode": compile_mode,
        "intermediate_schedules": intermediate_schedules,
        "save_only_last_checkpoint": save_only_last_checkpoint,
    }
    with open(os.path.join(output_dir, "train_parameters.json"), "w") as f: json.dump(train_params, f, indent=4)

    # load the last checkpoint if it exists, otherwise initialize the training from scratch
    try:
        last_checkpoint = torch.load(os.path.join(output_dir, "checkpoints", "last_checkpoint.pth"))
        start_epoch = last_checkpoint["epoch"]
        start_epoch_iter = last_checkpoint["epoch_iter"] + 1
        model_state_dict = last_checkpoint["model_state_dict"]
        optimizer_state_dict = last_checkpoint["optimizer_state_dict"]
        best_val_loss = last_checkpoint["best_val_loss"]
        patience_counter = last_checkpoint["patience_counter"]
        improved_this_epoch = last_checkpoint["improved_this_epoch"]
    except FileNotFoundError:
        os.makedirs(os.path.join(output_dir, "checkpoints"), exist_ok=True)
        start_epoch = 0
        start_epoch_iter = 0
        model_state_dict = None
        optimizer_state_dict = None
        best_val_loss = float("inf")
        patience_counter = 0
        improved_this_epoch = False

    # initialize the model
    model = GPT(
        intermediate_schedules=intermediate_schedules,
        vocab_size=vocab_size,
        n_embd=n_embd,
        n_head=n_head,
        n_layer=n_layer,
        dropout=dropout,
        ff_width=ff_width,
        block_size=block_size
    ).to(device)
    loguru.logger.info(f"The model has {human_readable(sum(p.numel() for p in model.parameters() if p.requires_grad))} trainable parameters")
    train_model = model
    if model_state_dict is not None:
        train_model.load_state_dict(model_state_dict)
    if compile:
        train_model = torch.compile(train_model, mode=compile_mode)

    # initialize the optimizer
    param_dict = {pn: p for pn, p in train_model.named_parameters()}
    ## filter out those that do not require grad
    param_dict = {pn: p for pn, p in param_dict.items() if p.requires_grad}
    ## create optim groups. Any parameters that is 2D will be weight decayed, otherwise no.
    ## i.e. all weight tensors in matmuls + embeddings decay, all biases and layernorms don't.
    decay_params = [p for n, p in param_dict.items() if p.dim() >= 2]
    nodecay_params = [p for n, p in param_dict.items() if p.dim() < 2]
    optim_groups = [
        {"params": decay_params, "weight_decay": weight_decay},
        {"params": nodecay_params, "weight_decay": 0.0},
    ]
    num_decay_params = sum(p.numel() for p in decay_params)
    num_nodecay_params = sum(p.numel() for p in nodecay_params)
    loguru.logger.info(f"num decayed parameter tensors: {len(decay_params)}, with {num_decay_params:,} parameters")
    loguru.logger.info(f"num non-decayed parameter tensors: {len(nodecay_params)}, with {num_nodecay_params:,} parameters")
    ## create AdamW optimizer and use the fused version if it is available
    fused_available = "fused" in inspect.signature(torch.optim.AdamW).parameters
    use_fused = fused_available and ("cuda" in device)
    loguru.logger.info(f"using fused AdamW: {use_fused}")
    extra_args = dict(fused=True) if use_fused else dict()
    optimizer = torch.optim.AdamW(optim_groups, lr=init_lr, betas=(beta1, beta2), **extra_args)
    ## load the optimizer state if it exists
    if optimizer_state_dict is not None:
        optimizer.load_state_dict(optimizer_state_dict)

    # set the torch precision to tf32
    torch.set_float32_matmul_precision("high")

    # initialize the np memmap array to save the batch losses
    batch_losses_path = os.path.join(output_dir, "batch_losses.npy")
    last_batch_loss_idx_path = os.path.join(output_dir, "last_batch_loss_idx.npy")
    val_losses_path = os.path.join(output_dir, "val_losses.npy")
    last_val_loss_idx_path = os.path.join(output_dir, "last_val_loss_idx.npy")

    try:
        batch_losses = np.lib.format.open_memmap(batch_losses_path, mode="r+", dtype=np.float32, shape=(nb_iters,))
        last_batch_loss_idx = np.lib.format.open_memmap(last_batch_loss_idx_path, mode="r+", dtype=np.int32, shape=())
        val_losses = np.lib.format.open_memmap(val_losses_path, mode="r+", dtype=np.float32, shape=(nb_epochs * math.ceil(len(train_data_loader)/checkpoint_interval),))
        last_val_loss_idx = np.lib.format.open_memmap(last_val_loss_idx_path, mode="r+", dtype=np.int32, shape=())
    except FileNotFoundError:
        batch_losses = np.lib.format.open_memmap(batch_losses_path, mode="w+", dtype=np.float32, shape=(nb_iters,))
        last_batch_loss_idx = np.lib.format.open_memmap(last_batch_loss_idx_path, mode="w+", dtype=np.int32, shape=())
        val_losses = np.lib.format.open_memmap(val_losses_path, mode="w+", dtype=np.float32, shape=(nb_epochs * math.ceil(len(train_data_loader)/checkpoint_interval),))
        last_val_loss_idx = np.lib.format.open_memmap(last_val_loss_idx_path, mode="w+", dtype=np.int32, shape=())
        last_batch_loss_idx[...] = 0
        last_batch_loss_idx.flush()
        last_val_loss_idx[...] = 0
        last_val_loss_idx.flush()

    # create data_loader for validation
    val_data_loader = torch.utils.data.DataLoader(
        FlowshopDataset(data_dir, split="val", load_in_memory=True),
        batch_size=val_batch_size,
        shuffle=False,
    )
    
    # launch the training loop
    early_stop = False
    for epoch in range(start_epoch, nb_epochs):
        if early_stop: break
        if start_epoch_iter == 0: improved_this_epoch = False

        # implement the logic to resume after failure
        ## create the generator, sampler, data loader
        generator = torch.Generator()
        generator.manual_seed(seed + epoch)
        train_sampler = torch.utils.data.RandomSampler(
            train_dataset,
            generator=generator
        )
        train_data_loader = torch.utils.data.DataLoader(
            train_dataset,
            batch_size=train_batch_size,
            sampler=train_sampler,
            drop_last=False if not compile else True,
        )
        train_data_loader_iterator = iter(train_data_loader)
        ## skip the iterations until start_iter for the epoch
        train_data_loader_iterator = itertools.islice(
            train_data_loader_iterator,
            start_epoch_iter,
            None,
        )

        # iterate over the training data loader
        for epoch_iter, (schedules_batch, makespans_batch) in (pbar:=tqdm(
            enumerate(train_data_loader_iterator, start=start_epoch_iter),
            total=len(train_data_loader),
            initial=start_epoch_iter,
            desc=f"Epoch {epoch+1}/{nb_epochs}",
        )):
            with open(os.path.join(output_dir, "train_pbar_epoch.log"), "w") as f: f.write(str(pbar))
        
            # move the batch to the device
            schedules_batch = schedules_batch.to(device)
            makespans_batch = makespans_batch.to(device)

            # clear the gradients
            optimizer.zero_grad(set_to_none=True)

            # forward pass and compute the loss
            makespans, loss = train_model(schedules_batch, makespans_batch)

            # backward pass
            loss.backward()

            # clip the gradients if grad_clip is set to a non-zero value
            if grad_clip != 0.0: torch.nn.utils.clip_grad_norm_(train_model.parameters(), grad_clip)

            # get the learning rate for the current iteration and set it in the optimizer
            current_iter = epoch * len(train_data_loader) + epoch_iter
            lr = get_lr(current_iter) if decay_lr else init_lr
            for param_group in optimizer.param_groups: param_group["lr"] = lr

            # update the parameters
            optimizer.step()

            # save the training loss for the current iteration
            batch_losses[current_iter] = loss.item()
            last_batch_loss_idx[...] = current_iter
            batch_losses.flush()
            last_batch_loss_idx.flush()

            # validation and checkpointing
            if (epoch_iter + 1) % checkpoint_interval == 0 or (epoch_iter + 1) == len(train_data_loader):

                # initialize the total validation loss
                total_val_loss = 0

                # iterate over the validation data loader
                for schedules_batch, makespans_batch in (pbar2:=tqdm(
                    val_data_loader,
                    desc=f"Validation {epoch+(epoch_iter+1)/len(train_data_loader):.2f}",
                )):
                    with open(os.path.join(output_dir, "train_pbar_val.log"), "w") as f: f.write(str(pbar2))

                    # move the batch to the device
                    schedules_batch = schedules_batch.to(device)
                    makespans_batch = makespans_batch.to(device)

                    # compute the validation loss without gradient tracking, and update the total validation loss
                    with torch.no_grad():
                        makespans, loss = train_model(schedules_batch, makespans_batch)
                    total_val_loss += loss.item() * schedules_batch.size(0)
                # ======
                with open(os.path.join(output_dir, "train_pbar_val.log"), "w") as f: f.write(str(pbar2))

                # compute the total validation loss (averaging over the dataset)
                total_val_loss /= len(val_data_loader.dataset)
                val_loss_idx = epoch * math.ceil(len(train_data_loader)/checkpoint_interval) + epoch_iter // checkpoint_interval
                val_losses[val_loss_idx] = total_val_loss
                last_val_loss_idx[...] = val_loss_idx
                val_losses.flush()
                last_val_loss_idx.flush()

                # early stopping check
                if total_val_loss < best_val_loss:
                    best_val_loss = total_val_loss
                    improved_this_epoch = True

                # save the checkpoint
                checkpoint = {
                    "epoch": epoch,
                    "epoch_iter": epoch_iter,
                    "model_state_dict": train_model.state_dict(),
                    "optimizer_state_dict": optimizer.state_dict(),
                    "validation_loss": total_val_loss,
                    "time": time.strftime("%Y_%m_%d_%H_%M_%S"),
                    "best_val_loss": best_val_loss,
                    "patience_counter": patience_counter,
                    "improved_this_epoch": improved_this_epoch,
                }
                torch.save(
                    checkpoint, 
                    os.path.join(output_dir, "checkpoints", "last_checkpoint.pth")
                )
                if not save_only_last_checkpoint:
                    torch.save(
                        checkpoint,
                        os.path.join(output_dir, "checkpoints", f"checkpoint_epoch_{epoch+(epoch_iter+1)/len(train_data_loader):.2f}.pth")
                    )
                if best_val_loss == total_val_loss:
                    torch.save(
                        checkpoint, 
                        os.path.join(output_dir, "checkpoints", "best_checkpoint.pth")
                    )
                # ======
            # ======
        # ======
        with open(os.path.join(output_dir, "train_pbar_epoch.log"), "w") as f: f.write(str(pbar))

        # set the start_epoch_iter to 0 for the next epoch
        start_epoch_iter = 0

        # check if early stopping should be triggered at the end of the epoch
        if improved_this_epoch:
            patience_counter = 0
        else:
            patience_counter += 1
            if patience_counter >= early_stopping_patience:
                loguru.logger.info(f"Early stopping triggered! Validation loss hasn't improved for {early_stopping_patience} epochs.")
                early_stop = True
    # ======

    # log the best validation loss
    loguru.logger.info(f"Best validation loss: {best_val_loss:.4f}")

    # create experiment termination flag file
    with open(os.path.join(output_dir, ".terminated_phase1"), "w") as f:
        pass
    # ======
# ======


if __name__ == "__main__":

    # parse arguments
    from argparse import ArgumentParser
    parser = ArgumentParser()
    parser.add_argument("--testing", type=bool, required=True)
    parser.add_argument("--seed", type=int, required=True)
    parser.add_argument("--data_dir", type=str, required=True)
    parser.add_argument("--n_embd", type=int, required=True)
    parser.add_argument("--n_head", type=int, required=True)
    parser.add_argument("--n_layer", type=int, required=True)
    parser.add_argument("--intermediate_schedules", type=bool, required=True)
    parser.add_argument("--dropout", type=float, required=True)
    parser.add_argument("--ff_width", type=int, required=True)
    parser.add_argument("--train_batch_size", type=int, required=True)
    parser.add_argument("--val_batch_size", type=int, required=True)
    parser.add_argument("--nb_epochs", type=int, required=True)
    parser.add_argument("--early_stopping_patience", type=int, required=True)
    parser.add_argument("--checkpoint_interval_ratio", type=float, required=True)
    parser.add_argument("--decay_lr", type=bool, required=True)
    parser.add_argument("--lr_partitions_ratios", type=lambda s: [float(item) for item in s.split(',')], help='Comma-separated list of floats that do not add up to 1 (e.g., 0.1,0.5,1)', required=True)
    parser.add_argument("--init_lr", type=float, required=True)
    parser.add_argument("--max_lr", type=float, required=True)
    parser.add_argument("--min_lr", type=float, required=True)
    parser.add_argument("--lr_warmup_iters_ratio", type=float, required=True)
    parser.add_argument("--lr_decay_iters_ratio", type=float, required=True)
    parser.add_argument("--beta1", type=float, required=True)
    parser.add_argument("--beta2", type=float, required=True)
    parser.add_argument("--weight_decay", type=float, required=True)
    parser.add_argument("--grad_clip", type=float, required=True)
    parser.add_argument("--compile", type=bool, required=True)
    parser.add_argument("--compile_mode", type=str, required=True)
    parser.add_argument("--save_only_last_checkpoint", type=bool, required=True)
    parser.add_argument("--output_dir", type=str, required=True)
    args = parser.parse_args()


    train(
        testing=args.testing,
        seed=args.seed,
        data_dir=args.data_dir,
        n_embd=args.n_embd,
        n_head=args.n_head,
        n_layer=args.n_layer,
        intermediate_schedules=args.intermediate_schedules,
        dropout=args.dropout,
        ff_width=args.ff_width,
        train_batch_size=args.train_batch_size,
        val_batch_size=args.val_batch_size,
        nb_epochs=args.nb_epochs,
        early_stopping_patience=args.early_stopping_patience,
        checkpoint_interval_ratio=args.checkpoint_interval_ratio,
        decay_lr=args.decay_lr,
        lr_partitions_ratios=args.lr_partitions_ratios,
        init_lr=args.init_lr,
        max_lr=args.max_lr,
        min_lr=args.min_lr,
        lr_warmup_iters_ratio=args.lr_warmup_iters_ratio,
        lr_decay_iters_ratio=args.lr_decay_iters_ratio,
        beta1=args.beta1,
        beta2=args.beta2,
        weight_decay=args.weight_decay,
        grad_clip=args.grad_clip,
        compile=args.compile,
        compile_mode=args.compile_mode,
        save_only_last_checkpoint=args.save_only_last_checkpoint,
        output_dir=args.output_dir,
    )