Model
Bert Large TensorFlow2 checkpoint pretrained using AMP and LAMB optimizer
Use the NGC CLI to download:
Copied!
1 Version
21.02.0_amp_optim-lambSelected09/22/2022 5:34 PM UTC1.25 GBAccuracy: 00 EpochsBatch Size: 0GPU: A100 Copied!
21.02.0_amp_optim-lambSelected
09/22/2022 5:34 PM UTC1.25 GBAccuracy: 00 EpochsBatch Size: 0GPU: A100
Copied!
architecture
| Key | Value |
|---|---|
| type | large |
performance
| Key | Value |
|---|---|
| training_loss | 1.656 |
training
| Key | Value |
|---|---|
| ITERATIONS_PHASE1 | 5774 |
| GLOBAL_BATCH_SIZE_PHASE1 | 79872 |
| LR_PHASE1 | 0.0065 |
| GLOBAL_BATCH_SIZE_PHASE2 | 30720 |
| training_precision | AMP |
| LR_PHASE2 | 0.004 |
| BS_PHASE2 | 40 |
| BS_PHASE1 | 312 |
| WARMUP_PHASE2 | 200 |
| dataset | Wikipedia;BookCorpus |
| ITERATIONS_PHASE2 | 1666 |
| WARMUP_PHASE1 | 2000 |