backup: log16/baseline-50k/vqvae/training-config.yaml
Browse files
log16/baseline-50k/vqvae/training-config.yaml
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
task: summarization
|
| 2 |
+
experiment: ablation-vqvae-log16
|
| 3 |
+
experiment_dir: /workspace/ncm/runs/ablation-vqvae-log16/vqvae
|
| 4 |
+
training:
|
| 5 |
+
seed: 42
|
| 6 |
+
log_interval: 10
|
| 7 |
+
eval_interval: 1000
|
| 8 |
+
checkpoint_interval: 5000
|
| 9 |
+
val_iters: 100
|
| 10 |
+
n_epochs: 3
|
| 11 |
+
total_iters: null
|
| 12 |
+
tqdm_interval: 10
|
| 13 |
+
batch_size: 64
|
| 14 |
+
eval_batch_size: 16
|
| 15 |
+
demo_batch_size: 4
|
| 16 |
+
stop_after_iters: 50000
|
| 17 |
+
model:
|
| 18 |
+
n_layers: 6
|
| 19 |
+
context_length: 1024
|
| 20 |
+
embed_dim: 384
|
| 21 |
+
in_vocab_size: 50304
|
| 22 |
+
vqvae_vocab_size:
|
| 23 |
+
- 16384
|
| 24 |
+
- 16384
|
| 25 |
+
- 16384
|
| 26 |
+
- 16384
|
| 27 |
+
- 16384
|
| 28 |
+
- 16384
|
| 29 |
+
- 16384
|
| 30 |
+
- 16384
|
| 31 |
+
- 16384
|
| 32 |
+
- 16384
|
| 33 |
+
- 16384
|
| 34 |
+
- 16384
|
| 35 |
+
- 16384
|
| 36 |
+
- 16384
|
| 37 |
+
- 16384
|
| 38 |
+
- 16384
|
| 39 |
+
compression_factor: 1
|
| 40 |
+
dropout: 0.2
|
| 41 |
+
pre_quant_groupnorm: 4
|
| 42 |
+
vector_quantizer:
|
| 43 |
+
clss: multiscale_residual_vector_quantizer
|
| 44 |
+
decay: 0.99
|
| 45 |
+
epsilon: 1.0e-05
|
| 46 |
+
commitment_cost: 0.25
|
| 47 |
+
learned_l1_sampling: true
|
| 48 |
+
learned_l1_cascade: true
|
| 49 |
+
learned_l1_cascade_base: 4
|
| 50 |
+
learned_all_sampling: false
|
| 51 |
+
quant_resi:
|
| 52 |
+
enabled: true
|
| 53 |
+
ratio: 0.5
|
| 54 |
+
share_mode: 0
|
| 55 |
+
learnable_ratio: false
|
| 56 |
+
levels:
|
| 57 |
+
use_manual_levels: true
|
| 58 |
+
manual_levels:
|
| 59 |
+
- 1
|
| 60 |
+
- 16
|
| 61 |
+
- 64
|
| 62 |
+
- 160
|
| 63 |
+
- 288
|
| 64 |
+
- 410
|
| 65 |
+
- 512
|
| 66 |
+
- 600
|
| 67 |
+
- 676
|
| 68 |
+
- 742
|
| 69 |
+
- 800
|
| 70 |
+
- 852
|
| 71 |
+
- 898
|
| 72 |
+
- 940
|
| 73 |
+
- 984
|
| 74 |
+
- 1024
|
| 75 |
+
aux:
|
| 76 |
+
fine_drop:
|
| 77 |
+
prob: 0.5
|
| 78 |
+
min_keep: 1
|
| 79 |
+
wandb:
|
| 80 |
+
project: ncp-release
|
| 81 |
+
enable: false
|
| 82 |
+
group: ablation-vqvae-log16
|
| 83 |
+
resume: null
|
| 84 |
+
gradients_and_params:
|
| 85 |
+
enable: false
|
| 86 |
+
log: all
|
| 87 |
+
log_freq: 1000
|
| 88 |
+
optimization:
|
| 89 |
+
lr: 0.0003
|
| 90 |
+
lr_decay_iters: null
|
| 91 |
+
min_lr: null
|
| 92 |
+
beta_1: 0.9
|
| 93 |
+
beta_2: 0.99
|
| 94 |
+
weight_decay: 0.01
|
| 95 |
+
max_grad_norm: 1.0
|
| 96 |
+
grad_accumulation_steps: 1
|
| 97 |
+
torch_compile:
|
| 98 |
+
enable: true
|
| 99 |
+
mode: default
|
| 100 |
+
dynamic: false
|
| 101 |
+
fullgraph: false
|
| 102 |
+
backend: inductor
|
| 103 |
+
mixed_precision:
|
| 104 |
+
enable: true
|
| 105 |
+
opt_level: O1
|
| 106 |
+
loss_scale: dynamic
|
| 107 |
+
dataset:
|
| 108 |
+
train_source: data/full_owt/train_gpt2.bin
|
| 109 |
+
validate_source: data/full_owt/valid_gpt2.bin
|
| 110 |
+
pre_tokenizer:
|
| 111 |
+
name: hf
|
| 112 |
+
tokenizer: hf
|
| 113 |
+
model_id: gpt2
|
| 114 |
+
special_tokens:
|
| 115 |
+
eos_token: <|endoftext|>
|