doc

make-tiny-xlm-roberta.py

@@ -29,15 +29,15 @@
 # 3. clone
 # git clone https://huggingface.co/hf-internal-testing/tiny-xlm-roberta
 # cd tiny-xlm-roberta
-
+#
 # 4. start with some pre-existing script from one of the https://huggingface.co/hf-internal-testing/ tiny model repos, e.g.
-# wget https://huggingface.co/hf-internal-testing/tiny-xlm-roberta/raw/main/make-tiny-xlm-roberta.py
-# chmod a+x ./make-tiny-xlm-roberta.py
-# mv ./make-tiny-xlm-roberta.py ./make-tiny-xlm-roberta.py
+# wget https://huggingface.co/hf-internal-testing/tiny-albert/raw/main/make-tiny-albert.py
+# chmod a+x ./make-tiny-albert.py
+# mv ./make-tiny-albert.py ./make-tiny-xlm-roberta.py
 #
 # 5. automatically rename things from the old names to new ones
-# perl -pi -e 's|MT5|XLMRoberta|g' make-tiny-xlm-roberta.py
-# perl -pi -e 's|mt5|xlm-roberta|g' make-tiny-xlm-roberta.py
+# perl -pi -e 's|Albert|XLMRoberta|g' make-tiny-xlm-roberta.py
+# perl -pi -e 's|albert|xlm-roberta|g' make-tiny-xlm-roberta.py
 #
 # 6. edit and re-run this script while fixing it up
 # ./make-tiny-xlm-roberta.py
@@ -63,7 +63,10 @@
 import sys
 import os
 
-from transformers import XLMRobertaTokenizer, XLMRobertaTokenizerFast, XLMRobertaConfig, XLMRobertaForCausalLM
+# workaround for fast tokenizer protobuf issue, and it's much faster too!
+os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+
+from transformers import XLMRobertaTokenizerFast, XLMRobertaConfig, XLMRobertaForCausalLM
 
 mname_orig = "xlm-roberta-base"
 mname_tiny = "tiny-xlm-roberta"
@@ -75,13 +78,11 @@ vocab_keep_items = 5000
 tmp_dir = f"/tmp/{mname_tiny}"
 vocab_orig_path = f"{tmp_dir}/sentencepiece.bpe.model"
 vocab_short_path = f"{tmp_dir}/spiece-short.model"
-# workaround for fast tokenizer protobuf issue, and it's much faster too!
-os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
 if 1: # set to 0 to skip this after running once to speed things up during tune up
     # HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
     sys.path.append("../sentencepiece/python/src/sentencepiece")
     import sentencepiece_model_pb2 as model
-    tokenizer_orig = XLMRobertaTokenizer.from_pretrained(mname_orig)
+    tokenizer_orig = XLMRobertaTokenizerFast.from_pretrained(mname_orig)
     tokenizer_orig.save_pretrained(tmp_dir)
     with open(vocab_orig_path, 'rb') as f: data = f.read()
     # adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
@@ -94,11 +95,11 @@ if 1: # set to 0 to skip this after running once to speed things up during tune
     m = None
 
 tokenizer_fast_tiny = XLMRobertaTokenizerFast(vocab_file=vocab_short_path)
-tokenizer_tiny = XLMRobertaTokenizer(vocab_file=vocab_short_path)
 
 ### Config
 
 config_tiny = XLMRobertaConfig.from_pretrained(mname_orig)
+print(config_tiny)
 # remember to update this to the actual config as each model is different and then shrink the numbers
 config_tiny.update(dict(
     vocab_size=vocab_keep_items+12,
@@ -121,12 +122,9 @@ print("New config", config_tiny)
 
 model_tiny = XLMRobertaForCausalLM(config_tiny)
 print(f"{mname_tiny}: num of params {model_tiny.num_parameters()}")
-model_tiny.resize_token_embeddings(len(tokenizer_tiny))
-
-inputs = tokenizer_tiny("hello", return_tensors="pt")
-outputs = model_tiny(**inputs)
-print("Test with normal tokenizer:", len(outputs.logits[0]))
+model_tiny.resize_token_embeddings(len(tokenizer_fast_tiny))
 
+# Test
 inputs = tokenizer_fast_tiny("hello", return_tensors="pt")
 outputs = model_tiny(**inputs)
 print("Test with fast tokenizer:", len(outputs.logits[0]))
@@ -134,7 +132,6 @@ print("Test with fast tokenizer:", len(outputs.logits[0]))
 # Save
 model_tiny.half() # makes it smaller
 model_tiny.save_pretrained(".")
-tokenizer_tiny.save_pretrained(".")
 tokenizer_fast_tiny.save_pretrained(".")
 
 readme = "README.md"