geohot/tinygrad
1
0
Fork 0
forked from tinygrad/tinygrad
Code Pull Requests Actions Activity

update resnet hparams based on BS=1632 RCP (#4210)

Browse Source 
https://github.com/mlcommons/logging/blob/master/mlperf_logging/rcp_checker/training_4.0.0/rcps_resnet.json
This commit is contained in:
chenyu
2024-04-18 12:01:46 -04:00
committed by GitHub
parent fa57c3e7ce
commit f7416916df
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/mlperf/model_train.py
+4 -4
View File
@@ -47,12 +47,12 @@ def train_resnet():
parameters = get_parameters(model)
# ** hyperparameters **
epochs = config["epochs"] = getenv("EPOCHS", 41)
epochs = config["epochs"] = getenv("EPOCHS", 37)
BS = config["BS"] = getenv("BS", 104 * len(GPUS)) # fp32 GPUS<=6 7900xtx can fit BS=112
EVAL_BS = config["EVAL_BS"] = getenv("EVAL_BS", BS)
base_lr = config["base_lr"] = getenv("LR", 8.5 * (BS/2048))
lr_warmup_epochs = config["lr_warmup_epochs"] = getenv("WARMUP_EPOCHS", 5)
decay = config["decay"] = getenv("DECAY", 2e-4)
base_lr = config["base_lr"] = getenv("LR", 7.4 * (BS/1632))
lr_warmup_epochs = config["lr_warmup_epochs"] = getenv("WARMUP_EPOCHS", 2)
decay = config["decay"] = getenv("DECAY", 5e-5)
loss_scaler = config["LOSS_SCALER"] = getenv("LOSS_SCALER", 128.0 if dtypes.default_float == dtypes.float16 else 1.0)