Train amplify
main(num_nodes=1, devices=1, min_seq_length=512, max_seq_length=512, result_dir=Path('./results'), num_steps=1000000, warmup_steps=1000, decay_steps=900000, limit_val_batches=1.0, val_check_interval=10000, log_every_n_steps=100, num_dataset_workers=27, biobert_spec_option=BiobertSpecOption.esm2_bert_layer_with_transformer_engine_spec, lr=0.001, micro_batch_size=64, accumulate_grad_batches=1, experiment_name='amplify', resume_if_exists=False, precision='bf16-mixed', wandb_entity=None, wandb_project=None, wandb_offline=False, wandb_tags=None, wandb_group=None, wandb_id=None, wandb_anonymous=False, wandb_log_model=False, pipeline_model_parallel_size=1, tensor_model_parallel_size=1, create_tensorboard_logger=False, nemo1_init_path=None, restore_from_checkpoint_path=None, save_last_checkpoint=True, metric_to_monitor_for_checkpoints='val_loss', save_top_k=2, nsys_profiling=False, nsys_start_step=0, nsys_end_step=None, nsys_ranks=[0], random_mask_strategy=RandomMaskStrategy.ALL_TOKENS, num_layers=24, hidden_size=640, num_attention_heads=10, ffn_hidden_size=2560, no_overlap_grad_reduce=False, overlap_param_gather=False, no_average_in_collective=False, grad_reduce_in_fp32=False)
Train an AMPLIFY model on UR100P data.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
num_nodes
|
int
|
Number of nodes to run on |
1
|
devices
|
int
|
number of devices |
1
|
min_seq_length
|
Optional[int]
|
Whether to pad sequences to a minimum length. If None, no extra padding is added |
512
|
max_seq_length
|
int
|
The maximum sequence length for the AMPLIFY transformer |
512
|
result_dir
|
Path
|
directory to store results, logs and checkpoints |
Path('./results')
|
num_steps
|
int
|
number of steps to train the model for |
1000000
|
warmup_steps
|
int
|
number of steps for the learning rate warmup phase |
1000
|
decay_steps
|
int
|
number of steps for the learning rate decay phase |
900000
|
limit_val_batches
|
int
|
limit the number of validation global batches to this many |
1.0
|
val_check_interval
|
int
|
number of steps to periodically check the validation loss and save |
10000
|
log_every_n_steps
|
Optional[int]
|
frequency for logging (steps) |
100
|
num_dataset_workers
|
int
|
num dataset workers |
27
|
biobert_spec_option
|
BiobertSpecOption
|
the biobert spec option (architecture) to use for this run |
esm2_bert_layer_with_transformer_engine_spec
|
lr
|
float
|
learning rate |
0.001
|
micro_batch_size
|
int
|
micro batch size, from this and parallelism settings we infer the global batch size |
64
|
accumulate_grad_batches
|
int
|
number of batches to accumulate before performing a gradient update |
1
|
experiment_name
|
str
|
experiment name, this is the name used for the wandb run, and the sub-directory of the result_dir that stores the logs and checkpoints. |
'amplify'
|
resume_if_exists
|
bool
|
attempt to resume if the checkpoint exists [FIXME @skothenhill this doesn't work yet] |
False
|
precision
|
PrecisionTypes
|
precision to use for training (bf16-mixed, 16-mixed, 32) |
'bf16-mixed'
|
wandb_entity
|
str
|
The team posting this run (default: your username or your default team) |
None
|
wandb_project
|
str
|
The name of the project to which this run will belong. |
None
|
wandb_tags
|
List[str]
|
Tags associated with this run. |
None
|
wandb_group
|
str
|
A unique string shared by all runs in a given group |
None
|
wandb_offline
|
bool
|
Run offline (data can be streamed later to wandb servers). |
False
|
wandb_id
|
str
|
Sets the version, mainly used to resume a previous run. |
None
|
wandb_anonymous
|
bool
|
Enables or explicitly disables anonymous logging. |
False
|
wandb_log_model
|
bool
|
Save checkpoints in wandb dir to upload on W&B servers. |
False
|
pipeline_model_parallel_size
|
int
|
degree of pipeline model parallelism |
1
|
tensor_model_parallel_size
|
int
|
degree of tensor model parallelism |
1
|
create_tensorboard_logger
|
bool
|
create the tensorboard logger |
False
|
nemo1_init_path
|
Optional[Path]
|
path to a NeMo v1 checkpoint to initialize from |
None
|
restore_from_checkpoint_path
|
Optional[str]
|
If set, restores the model from the directory passed in. Expects the checkpoint to be created by using the ModelCheckpoint class and always_save_context=True. |
None
|
save_last_checkpoint
|
bool
|
whether to save the last checkpoint |
True
|
metric_to_monitor_for_checkpoints
|
str
|
metric to monitor for checkpoints |
'val_loss'
|
save_top_k
|
int
|
number of top checkpoints to save |
2
|
nsys_profiling
|
bool
|
whether to enable nsys profiling |
False
|
nsys_start_step
|
int
|
start step for nsys profiling |
0
|
nsys_end_step
|
Optional[int]
|
end step for nsys profiling |
None
|
nsys_ranks
|
List[int]
|
ranks for nsys profiling |
[0]
|
random_mask_strategy
|
RandomMaskStrategy
|
random mask strategy |
ALL_TOKENS
|
num_layers
|
int
|
number of layers |
24
|
hidden_size
|
int
|
hidden size |
640
|
num_attention_heads
|
int
|
number of attention heads |
10
|
ffn_hidden_size
|
int
|
feed forward hidden size |
2560
|
no_overlap_grad_reduce
|
bool
|
disable overlap gradient reduction |
False
|
overlap_param_gather
|
bool
|
overlap parameter gather |
False
|
no_average_in_collective
|
bool
|
disable average in collective |
False
|
grad_reduce_in_fp32
|
bool
|
gradient reduction in fp32 |
False
|
Source code in bionemo/amplify/train_amplify.py
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 |
|