MCPcopy Create free account
hub / github.com/Alpha-VLLM/LLaMA2-Accessory / get_train_sampler

Function get_train_sampler

Large-DiT-ImageNet/train.py:53–73  ·  view source on GitHub ↗
(dataset, rank, world_size, global_batch_size, max_steps,
                      resume_step, seed)

Source from the content-addressed store, hash-verified

51
52
53def get_train_sampler(dataset, rank, world_size, global_batch_size, max_steps,
54 resume_step, seed):
55 sample_indices = torch.empty([max_steps * global_batch_size // world_size],
56 dtype=torch.long)
57 epoch_id, fill_ptr, offs = 0, 0, 0
58 while fill_ptr < sample_indices.size(0):
59 g = torch.Generator()
60 g.manual_seed(seed + epoch_id)
61 epoch_sample_indices = torch.randperm(len(dataset), generator=g)
62 epoch_id += 1
63 epoch_sample_indices = epoch_sample_indices[
64 (rank + offs) % world_size::world_size
65 ]
66 offs = (offs + world_size - len(dataset) % world_size) % world_size
67 epoch_sample_indices = epoch_sample_indices[
68 :sample_indices.size(0) - fill_ptr
69 ]
70 sample_indices[fill_ptr: fill_ptr + epoch_sample_indices.size(0)] = \
71 epoch_sample_indices
72 fill_ptr += epoch_sample_indices.size(0)
73 return sample_indices[resume_step * global_batch_size // world_size:].tolist()
74
75
76@torch.no_grad()

Callers 1

mainFunction · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected