Setup training for gpt2 on oscar-ta

.gitattributes

@@ -14,3 +14,4 @@
 *.pb filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

create_config.py

@@ -0,0 +1,6 @@
+from transformers import GPT2Config
+
+model_dir = "./gpt2-tamil"  # ${MODEL_DIR}
+
+config = GPT2Config.from_pretrained("gpt2", resid_pdrop=0.0, embd_pdrop=0.0, attn_pdrop=0.0)
+config.save_pretrained(model_dir)

gpt2-tamil/config.json

@@ -0,0 +1,36 @@
+{
+  "activation_function": "gelu_new",
+  "architectures": [
+    "GPT2LMHeadModel"
+  ],
+  "attn_pdrop": 0.0,
+  "bos_token_id": 50256,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 50256,
+  "gradient_checkpointing": false,
+  "initializer_range": 0.02,
+  "layer_norm_epsilon": 1e-05,
+  "model_type": "gpt2",
+  "n_ctx": 1024,
+  "n_embd": 768,
+  "n_head": 12,
+  "n_inner": null,
+  "n_layer": 12,
+  "n_positions": 1024,
+  "resid_pdrop": 0.0,
+  "scale_attn_weights": true,
+  "summary_activation": null,
+  "summary_first_dropout": 0.1,
+  "summary_proj_to_labels": true,
+  "summary_type": "cls_index",
+  "summary_use_proj": true,
+  "task_specific_params": {
+    "text-generation": {
+      "do_sample": true,
+      "max_length": 50
+    }
+  },
+  "transformers_version": "4.9.0.dev0",
+  "use_cache": true,
+  "vocab_size": 50257
+}

run_clm_flax.py

@@ -0,0 +1 @@
+/home/deepak/sources/transformers/examples/flax/language-modeling/run_clm_flax.py

train_gpt2-oscar-tamil.sh

@@ -0,0 +1,17 @@
+#!/usr/bin/env bash
+./run_clm_flax.py \
+    --output_dir="${MODEL_DIR}" \
+    --model_type="gpt2" \
+    --config_name="${MODEL_DIR}" \
+    --tokenizer_name="${MODEL_DIR}" \
+    --dataset_name="oscar" \
+    --dataset_config_name="unshuffled_deduplicated_ta" \
+    --do_train --do_eval \
+    --block_size="512" \
+    --per_device_train_batch_size="64" \
+    --per_device_eval_batch_size="64" \
+    --learning_rate="5e-3" --warmup_steps="1000" \
+    --adam_beta1="0.9" --adam_beta2="0.98" --weight_decay="0.01" \
+    --overwrite_output_dir \
+    --num_train_epochs="20" \
+    --push_to_hub

train_tokenizer.py

@@ -0,0 +1,27 @@
+from datasets import load_dataset
+from tokenizers import trainers, Tokenizer, normalizers, ByteLevelBPETokenizer
+
+model_dir = "./gpt2-tamil"  # ${MODEL_DIR}
+
+# load dataset
+dataset = load_dataset("oscar", "unshuffled_deduplicated_ta", split="train")
+
+# Instantiate tokenizer
+tokenizer = ByteLevelBPETokenizer()
+
+def batch_iterator(batch_size=1000):
+    for i in range(0, len(dataset), batch_size):
+        yield dataset[i: i + batch_size]["text"]
+
+# Customized training
+tokenizer.train_from_iterator(batch_iterator(), vocab_size=50265, min_frequency=2, special_tokens=[
+    "<s>",
+    "<pad>",
+    "</s>",
+    "<unk>",
+    "<mask>",
+])
+
+# Save files to disk
+tokenizer.save(f"{model_dir}/tokenizer.json")
+