Saving weights and logs of step 10000

events.out.tfevents.1642349949.t1v-n-e1a08808-w-0.1786795.0.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dec4f33794c71cb36e2174526c6bf55d893d81add611f5fa83f296421cacbb8
+size 1470136

flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:636f2a88b1b680099c0beec3f038cdd04ebdbb1ffb5f71884b15e5d8b3be650c
+oid sha256:b23fc406171b4262000759c2ab5d18fabe36374906b73b564b726f4a8907d3ec
 size 498796983

run_mlm_flax.py

@@ -129,6 +129,10 @@ class DataTrainingArguments:
     static_learning_rate: bool = field(
         default=False, metadata={"help": "Use a non decaying learning rate"}
     )
+    end_learning_rate: float = field(
+        default=0, metadata={"help": "End learning rate. Will be ignored it startic learning rate is set"}
+    )
+    
     auth_token: bool = field(
         default=False, metadata={"help": "Use authorisation token"}
     )
@@ -510,6 +514,8 @@ if __name__ == "__main__":
 
     if data_args.static_learning_rate:
         end_lr_value = training_args.learning_rate
+    elif data_args.end_learning_rate:
+        end_lr_value = data_args.end_learning_rate
     else:
         end_lr_value = 0
 
@@ -622,7 +628,6 @@ if __name__ == "__main__":
 
         # Generate an epoch by shuffling sampling indices from the train dataset
         num_train_samples = len(tokenized_datasets["train"])
-        print(f'Total number of training samples: {num_train_samples}')
         train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
         train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size)
 

run_step2.sh

@@ -14,7 +14,7 @@
     --learning_rate="4e-4" \
     --warmup_steps="0" \
     --overwrite_output_dir \
-    --num_train_epochs="2" \
+    --num_train_epochs="1" \
     --adam_beta1="0.9" \
     --adam_beta2="0.98" \
     --adam_epsilon="1e-6" \

run_step3.sh

@@ -0,0 +1,29 @@
+./run_mlm_flax.py \
+    --output_dir="./" \
+    --model_type="roberta" \
+    --model_name_or_path="./" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --train_file /mnt/disks/flaxdisk/corpus/train_3_4.json \
+    --validation_file /mnt/disks/flaxdisk/corpus/validation.json \
+    --cache_dir="/mnt/disks/flaxdisk/cache/" \
+    --max_seq_length="128" \
+    --weight_decay="0.01" \
+    --per_device_train_batch_size="192" \
+    --per_device_eval_batch_size="192" \
+    --learning_rate="4e-4" \
+    --end_learning_rate="2e-4" \
+    --warmup_steps="0" \
+    --overwrite_output_dir \
+    --num_train_epochs="1" \
+    --adam_beta1="0.9" \
+    --adam_beta2="0.98" \
+    --adam_epsilon="1e-6" \
+    --logging_steps="10000" \
+    --save_steps="10000" \
+    --eval_steps="10000" \
+    --preprocessing_num_workers="64" \
+    --auth_token="True" \
+    --dtype="bfloat16" \
+    --push_to_hub
+