model corrections
Browse files- config.json +2 -2
- make-tiny-albert.py +11 -6
- pytorch_model.bin +2 -2
- tokenizer_config.json +1 -1
config.json
CHANGED
@@ -11,12 +11,12 @@
|
|
11 |
"gap_size": 0,
|
12 |
"hidden_act": "gelu",
|
13 |
"hidden_dropout_prob": 0.1,
|
14 |
-
"hidden_size":
|
15 |
"initializer_range": 0.02,
|
16 |
"inner_group_num": 1,
|
17 |
"intermediate_size": 128,
|
18 |
"layer_norm_eps": 1e-12,
|
19 |
-
"max_position_embeddings":
|
20 |
"model_type": "albert",
|
21 |
"net_structure_type": 0,
|
22 |
"num_attention_heads": 2,
|
11 |
"gap_size": 0,
|
12 |
"hidden_act": "gelu",
|
13 |
"hidden_dropout_prob": 0.1,
|
14 |
+
"hidden_size": 32,
|
15 |
"initializer_range": 0.02,
|
16 |
"inner_group_num": 1,
|
17 |
"intermediate_size": 128,
|
18 |
"layer_norm_eps": 1e-12,
|
19 |
+
"max_position_embeddings": 256,
|
20 |
"model_type": "albert",
|
21 |
"net_structure_type": 0,
|
22 |
"num_attention_heads": 2,
|
make-tiny-albert.py
CHANGED
@@ -66,10 +66,11 @@ import os
|
|
66 |
# workaround for fast tokenizer protobuf issue, and it's much faster too!
|
67 |
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
|
68 |
|
69 |
-
from transformers import AlbertTokenizerFast, AlbertConfig, AlbertForMaskedLM
|
70 |
|
71 |
mname_orig = "albert-base-v1"
|
72 |
mname_tiny = "tiny-albert"
|
|
|
73 |
|
74 |
### Tokenizer
|
75 |
|
@@ -82,7 +83,7 @@ if 1: # set to 0 to skip this after running once to speed things up during tune
|
|
82 |
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
|
83 |
sys.path.append("../sentencepiece/python/src/sentencepiece")
|
84 |
import sentencepiece_model_pb2 as model
|
85 |
-
tokenizer_orig =
|
86 |
tokenizer_orig.save_pretrained(tmp_dir)
|
87 |
with open(vocab_orig_path, 'rb') as f: data = f.read()
|
88 |
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
|
@@ -94,7 +95,10 @@ if 1: # set to 0 to skip this after running once to speed things up during tune
|
|
94 |
with open(vocab_short_path, 'wb') as f: f.write(m.SerializeToString())
|
95 |
m = None
|
96 |
|
97 |
-
|
|
|
|
|
|
|
98 |
|
99 |
|
100 |
### Config
|
@@ -103,11 +107,11 @@ config_tiny = AlbertConfig.from_pretrained(mname_orig)
|
|
103 |
print(config_tiny)
|
104 |
# remember to update this to the actual config as each model is different and then shrink the numbers
|
105 |
config_tiny.update(dict(
|
106 |
-
vocab_size=vocab_keep_items
|
107 |
embedding_size=64,
|
108 |
-
hidden_size=
|
109 |
intermediate_size=128,
|
110 |
-
max_position_embeddings=
|
111 |
num_attention_heads=2,
|
112 |
num_hidden_groups=1,
|
113 |
num_hidden_layers=2,
|
@@ -122,6 +126,7 @@ model_tiny.resize_token_embeddings(len(tokenizer_fast_tiny))
|
|
122 |
|
123 |
# Test
|
124 |
inputs = tokenizer_fast_tiny("The capital of France is [MASK].", return_tensors="pt")
|
|
|
125 |
outputs = model_tiny(**inputs)
|
126 |
print("Test with normal tokenizer:", len(outputs.logits[0]))
|
127 |
|
66 |
# workaround for fast tokenizer protobuf issue, and it's much faster too!
|
67 |
os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
|
68 |
|
69 |
+
from transformers import AlbertTokenizer, AlbertTokenizerFast, AlbertConfig, AlbertForMaskedLM
|
70 |
|
71 |
mname_orig = "albert-base-v1"
|
72 |
mname_tiny = "tiny-albert"
|
73 |
+
model_max_length = 256
|
74 |
|
75 |
### Tokenizer
|
76 |
|
83 |
# HACK: need the sentencepiece source to get sentencepiece_model_pb2, as it doesn't get installed
|
84 |
sys.path.append("../sentencepiece/python/src/sentencepiece")
|
85 |
import sentencepiece_model_pb2 as model
|
86 |
+
tokenizer_orig = AlbertTokenizer.from_pretrained(mname_orig)
|
87 |
tokenizer_orig.save_pretrained(tmp_dir)
|
88 |
with open(vocab_orig_path, 'rb') as f: data = f.read()
|
89 |
# adapted from https://blog.ceshine.net/post/trim-down-sentencepiece-vocabulary/
|
95 |
with open(vocab_short_path, 'wb') as f: f.write(m.SerializeToString())
|
96 |
m = None
|
97 |
|
98 |
+
# albert breaks without having tokenizer.model_max_length match config.max_position_embeddings
|
99 |
+
# these values are hardcoded in the source for official models, so we have to explicitly set those here
|
100 |
+
tokenizer_fast_tiny = AlbertTokenizerFast(vocab_file=vocab_short_path,
|
101 |
+
model_max_length=model_max_length)
|
102 |
|
103 |
|
104 |
### Config
|
107 |
print(config_tiny)
|
108 |
# remember to update this to the actual config as each model is different and then shrink the numbers
|
109 |
config_tiny.update(dict(
|
110 |
+
vocab_size=vocab_keep_items,
|
111 |
embedding_size=64,
|
112 |
+
hidden_size=32,
|
113 |
intermediate_size=128,
|
114 |
+
max_position_embeddings=model_max_length,
|
115 |
num_attention_heads=2,
|
116 |
num_hidden_groups=1,
|
117 |
num_hidden_layers=2,
|
126 |
|
127 |
# Test
|
128 |
inputs = tokenizer_fast_tiny("The capital of France is [MASK].", return_tensors="pt")
|
129 |
+
#print(inputs)
|
130 |
outputs = model_tiny(**inputs)
|
131 |
print("Test with normal tokenizer:", len(outputs.logits[0]))
|
132 |
|
pytorch_model.bin
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e8c50a0be432afd0e75428e41f98be8937fc3f234d99844b1a46603b91e14c62
|
3 |
+
size 730318
|
tokenizer_config.json
CHANGED
@@ -1 +1 @@
|
|
1 |
-
{"do_lower_case": true, "remove_space": true, "keep_accents": false, "bos_token": "[CLS]", "eos_token": "[SEP]", "unk_token": "<unk>", "sep_token": "[SEP]", "pad_token": "<pad>", "cls_token": "[CLS]", "mask_token": {"content": "[MASK]", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "sp_model_kwargs": {}, "tokenizer_class": "AlbertTokenizer"}
|
1 |
+
{"do_lower_case": true, "remove_space": true, "keep_accents": false, "bos_token": "[CLS]", "eos_token": "[SEP]", "unk_token": "<unk>", "sep_token": "[SEP]", "pad_token": "<pad>", "cls_token": "[CLS]", "mask_token": {"content": "[MASK]", "single_word": false, "lstrip": true, "rstrip": false, "normalized": true, "__type": "AddedToken"}, "model_max_length": 256, "sp_model_kwargs": {}, "tokenizer_class": "AlbertTokenizer"}
|