Update train_script.py
Browse filesAdded loading pretrained model
- train_script.py +7 -2
train_script.py
CHANGED
@@ -4,7 +4,7 @@ import os
|
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
|
7 |
-
from transformers import GPT2TokenizerFast, GPT2LMHeadModel
|
8 |
from transformers import DataCollatorWithPadding, GPT2Config, DataCollatorForLanguageModeling
|
9 |
from transformers import Trainer, TrainingArguments, RobertaTokenizerFast
|
10 |
|
@@ -58,7 +58,7 @@ dataset = dataset.with_format("torch")
|
|
58 |
tokenizer = RobertaTokenizerFast.from_pretrained(ENCODER_MODEL_NAME, max_len=TOKENIZER_MAX_LEN)
|
59 |
collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
|
60 |
|
61 |
-
|
62 |
config = GPT2Config(
|
63 |
vocab_size=len(tokenizer),
|
64 |
n_positions=TOKENIZER_MAX_LEN,
|
@@ -71,6 +71,11 @@ config = GPT2Config(
|
|
71 |
|
72 |
model = ConditionalGPT2LMHeadModel(config)
|
73 |
|
|
|
|
|
|
|
|
|
|
|
74 |
# change trainer args as needed
|
75 |
args = TrainingArguments(
|
76 |
output_dir=TRAINER_SAVE_DIR,
|
|
|
4 |
import torch
|
5 |
import torch.nn as nn
|
6 |
|
7 |
+
from transformers import GPT2TokenizerFast, GPT2LMHeadModel, AutoModelForCausalLM
|
8 |
from transformers import DataCollatorWithPadding, GPT2Config, DataCollatorForLanguageModeling
|
9 |
from transformers import Trainer, TrainingArguments, RobertaTokenizerFast
|
10 |
|
|
|
58 |
tokenizer = RobertaTokenizerFast.from_pretrained(ENCODER_MODEL_NAME, max_len=TOKENIZER_MAX_LEN)
|
59 |
collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
|
60 |
|
61 |
+
# train from scratch
|
62 |
config = GPT2Config(
|
63 |
vocab_size=len(tokenizer),
|
64 |
n_positions=TOKENIZER_MAX_LEN,
|
|
|
71 |
|
72 |
model = ConditionalGPT2LMHeadModel(config)
|
73 |
|
74 |
+
# alternatively, load a pre-trained model
|
75 |
+
# commit_hash = '0ba58478f467056fe33003d7d91644ecede695a7'
|
76 |
+
# model = AutoModelForCausalLM.from_pretrained("entropy/roberta_zinc_decoder",
|
77 |
+
# trust_remote_code=True, revision=commit_hash)
|
78 |
+
|
79 |
# change trainer args as needed
|
80 |
args = TrainingArguments(
|
81 |
output_dir=TRAINER_SAVE_DIR,
|