Adding base config and organizing configs
Browse files- configs/base/config.json +25 -0
- config.json → configs/large/config.json +0 -0
- run.sh +4 -7
- run_stream.sh +1 -1
configs/base/config.json
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"RobertaForMaskedLM"
|
4 |
+
],
|
5 |
+
"attention_probs_dropout_prob": 0.1,
|
6 |
+
"bos_token_id": 0,
|
7 |
+
"eos_token_id": 2,
|
8 |
+
"gradient_checkpointing": false,
|
9 |
+
"hidden_act": "gelu",
|
10 |
+
"hidden_dropout_prob": 0.1,
|
11 |
+
"hidden_size": 768,
|
12 |
+
"initializer_range": 0.02,
|
13 |
+
"intermediate_size": 3072,
|
14 |
+
"layer_norm_eps": 1e-05,
|
15 |
+
"max_position_embeddings": 514,
|
16 |
+
"model_type": "roberta",
|
17 |
+
"num_attention_heads": 12,
|
18 |
+
"num_hidden_layers": 12,
|
19 |
+
"pad_token_id": 1,
|
20 |
+
"position_embedding_type": "absolute",
|
21 |
+
"transformers_version": "4.9.0.dev0",
|
22 |
+
"type_vocab_size": 1,
|
23 |
+
"use_cache": true,
|
24 |
+
"vocab_size": 50265
|
25 |
+
}
|
config.json → configs/large/config.json
RENAMED
File without changes
|
run.sh
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
# From https://arxiv.org/pdf/1907.11692.pdf
|
2 |
-
|
3 |
./run_mlm_flax.py \
|
4 |
-
--output_dir="./" \
|
5 |
--model_type="roberta" \
|
6 |
-
--config_name="./" \
|
7 |
--tokenizer_name="./" \
|
8 |
--dataset_name="mc4" \
|
9 |
--dataset_config_name="es" \
|
@@ -25,7 +25,4 @@ HUB_TOKEN=`cat $HOME/.huggingface/token`
|
|
25 |
--num_train_steps="500000" \
|
26 |
--eval_steps="10000" \
|
27 |
--logging_steps="500" \
|
28 |
-
--dtype="bfloat16"
|
29 |
-
--push_to_hub_model_id="flax-community/bertin-roberta-large-spanish" \
|
30 |
-
--push_to_hub_token="$HUB_TOKEN"
|
31 |
-
--push_to_hub 2>&1 | tee run.log
|
|
|
1 |
# From https://arxiv.org/pdf/1907.11692.pdf
|
2 |
+
python -c "import jax; print('TPUs', jax.device_count())"
|
3 |
./run_mlm_flax.py \
|
4 |
+
--output_dir="./outputs" \
|
5 |
--model_type="roberta" \
|
6 |
+
--config_name="./configs/large" \
|
7 |
--tokenizer_name="./" \
|
8 |
--dataset_name="mc4" \
|
9 |
--dataset_config_name="es" \
|
|
|
25 |
--num_train_steps="500000" \
|
26 |
--eval_steps="10000" \
|
27 |
--logging_steps="500" \
|
28 |
+
--dtype="bfloat16" 2>&1 | tee run.log
|
|
|
|
|
|
run_stream.sh
CHANGED
@@ -3,7 +3,7 @@ python -c "import jax; print('TPUs', jax.device_count())"
|
|
3 |
python ./run_mlm_flax_stream.py \
|
4 |
--output_dir="./outputs" \
|
5 |
--model_type="roberta" \
|
6 |
-
--config_name="./
|
7 |
--tokenizer_name="./" \
|
8 |
--dataset_name="./mc4" \
|
9 |
--dataset_config_name="es" \
|
|
|
3 |
python ./run_mlm_flax_stream.py \
|
4 |
--output_dir="./outputs" \
|
5 |
--model_type="roberta" \
|
6 |
+
--config_name="./configs/base" \
|
7 |
--tokenizer_name="./" \
|
8 |
--dataset_name="./mc4" \
|
9 |
--dataset_config_name="es" \
|