Upload unilm/dit/classification/deepspeed_configs/config.json with huggingface_hub
Browse files
unilm/dit/classification/deepspeed_configs/config.json
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train_batch_size": 2048,
|
3 |
+
"train_micro_batch_size_per_gpu": 32,
|
4 |
+
"steps_per_print": 1000,
|
5 |
+
"optimizer": {
|
6 |
+
"type": "Adam",
|
7 |
+
"adam_w_mode": true,
|
8 |
+
"params": {
|
9 |
+
"lr": 0.001,
|
10 |
+
"weight_decay": 0.05,
|
11 |
+
"bias_correction": true,
|
12 |
+
"betas": [
|
13 |
+
0.9,
|
14 |
+
0.999
|
15 |
+
]
|
16 |
+
}
|
17 |
+
},
|
18 |
+
"gradient_clipping": 1,
|
19 |
+
"fp16": {
|
20 |
+
"enabled": true,
|
21 |
+
"loss_scale": 0,
|
22 |
+
"initial_scale_power": 7,
|
23 |
+
"loss_scale_window": 128
|
24 |
+
},
|
25 |
+
"zero_optimization": {
|
26 |
+
"stage": 1
|
27 |
+
}
|
28 |
+
}
|