wip gpt2

generators/gpt2.py

@@ -0,0 +1,32 @@
+import pytorch_lightning.utilities
+# hack until https://github.com/minimaxir/aitextgen/issues/200 is fixed
+pytorch_lightning.utilities._TPU_AVAILABLE = False
+
+from aitextgen.TokenDataset import TokenDataset
+from aitextgen.tokenizers import train_tokenizer
+from aitextgen.utils import GPT2ConfigCPU
+from aitextgen import aitextgen
+
+# The name of the downloaded Shakespeare text for training
+file_name = "littlethief.txt"
+
+# Train a custom BPE Tokenizer on the downloaded text
+# This will save one file: `aitextgen.tokenizer.json`, which contains the
+# information needed to rebuild the tokenizer.
+train_tokenizer(file_name)
+tokenizer_file = "aitextgen.tokenizer.json"
+
+# GPT2ConfigCPU is a mini variant of GPT-2 optimized for CPU-training
+# e.g. the # of input tokens here is 64 vs. 1024 for base GPT-2.
+config = GPT2ConfigCPU()
+
+# Instantiate aitextgen using the created tokenizer and config
+ai = aitextgen(tokenizer_file=tokenizer_file, config=config)
+
+# You can build datasets for training by creating TokenDatasets,
+# which automatically processes the dataset with the appropriate size.
+data = TokenDataset(file_name, tokenizer_file=tokenizer_file, block_size=64)
+
+# Train the model! It will save pytorch_model.bin periodically and after completion to the `trained_model` folder.
+# On a 2020 8-core iMac, this took ~25 minutes to run.
+ai.train(data, batch_size=8, num_steps=50000, generate_every=5000, save_every=5000)

requirements/gpt2.txt

@@ -1,6 +1,2 @@
-fire ~= 0.1.3
+pytorch-lightning==1.7.7
-# this package claims perpetual backwards compatibility with the "re" module so we allow any version
+aitextgen
-regex
-requests ~= 2.0
-tqdm ~= 4.0
-toposort ~= 1.0