SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

upgrade sentencepiece version (#13564)

Browse Source
This commit is contained in:
elishowk
2021-09-15 15:25:03 +02:00
committed by GitHub
parent e86c02ea90
commit c783e14887
5 changed files with 104 additions and 93 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
examples/research_projects/lxmert/demo.ipynb
View File

@@ -49,7 +49,7 @@
"\n",
"\n",
"# for visualizing output\n",
"def showarray(a, fmt='jpeg'):\n",
"def showarray(a, fmt=\"jpeg\"):\n",
" a = np.uint8(np.clip(a, 0, 255))\n",
" f = io.BytesIO()\n",
" PIL.Image.fromarray(a).save(f, fmt)\n",
@@ -128,7 +128,7 @@
" scales_yx=scales_yx,\n",
" padding=\"max_detections\",\n",
" max_detections=frcnn_cfg.max_detections,\n",
" return_tensors=\"pt\"\n",
" return_tensors=\"pt\",\n",
")\n",
"# add boxes and labels to the image\n",
"\n",
@@ -174,7 +174,7 @@
" \"Where is this scene?\",\n",
" \"what is the man riding?\",\n",
" \"What is the man wearing?\",\n",
" \"What is the color of the horse?\"\n",
" \"What is the color of the horse?\",\n",
"]\n",
"test_questions_for_url2 = [\n",
" \"Where is the cat?\",\n",
@@ -200,7 +200,7 @@
" return_token_type_ids=True,\n",
" return_attention_mask=True,\n",
" add_special_tokens=True,\n",
" return_tensors=\"pt\"\n",
" return_tensors=\"pt\",\n",
" )\n",
"\n",
" # run lxmert(s)\n",

77
examples/research_projects/movement-pruning/Saving_PruneBERT.ipynb
View File

@@ -44,7 +44,7 @@
"\n",
"from transformers import *\n",
"\n",
"os.chdir('../../')"
"os.chdir(\"../../\")"
]
},
{
@@ -70,7 +70,7 @@
"# Load fine-pruned model and quantize the model\n",
"\n",
"model = BertForQuestionAnswering.from_pretrained(\"huggingface/prunebert-base-uncased-6-finepruned-w-distil-squad\")\n",
"model.to('cpu')\n",
"model.to(\"cpu\")\n",
"\n",
"quantized_model = torch.quantization.quantize_dynamic(\n",
" model=model,\n",
@@ -92,10 +92,14 @@
"source": [
"# Saving the original (encoder + classifier) in the standard torch.save format\n",
"\n",
"dense_st = {name: param for name, param in model.state_dict().items() \n",
" if \"embedding\" not in name and \"pooler\" not in name}\n",
"torch.save(dense_st, 'dbg/dense_squad.pt',)\n",
"dense_mb_size = os.path.getsize(\"dbg/dense_squad.pt\")\n"
"dense_st = {\n",
" name: param for name, param in model.state_dict().items() if \"embedding\" not in name and \"pooler\" not in name\n",
"}\n",
"torch.save(\n",
" dense_st,\n",
" \"dbg/dense_squad.pt\",\n",
")\n",
"dense_mb_size = os.path.getsize(\"dbg/dense_squad.pt\")"
]
},
{
@@ -214,7 +218,7 @@
" elementary_qtz_st[f\"{name}.int_repr.indices\"] = np.uint16(int_repr_cs.indices) # np.array uint16\n",
" elementary_qtz_st[f\"{name}.int_repr.shape\"] = int_repr_cs.shape # tuple(int, int)\n",
" else:\n",
" elementary_qtz_st[name] = param\n"
" elementary_qtz_st[name] = param"
]
},
{
@@ -225,7 +229,7 @@
"source": [
"# Create mapping from torch.dtype to string description (we could also used an int8 instead of string)\n",
"str_2_dtype = {\"qint8\": torch.qint8}\n",
"dtype_2_str = {torch.qint8: \"qint8\"}\n"
"dtype_2_str = {torch.qint8: \"qint8\"}"
]
},
{
@@ -246,11 +250,17 @@
"source": [
"# Saving the pruned (encoder + classifier) in the standard torch.save format\n",
"\n",
"dense_optimized_st = {name: param for name, param in elementary_qtz_st.items() \n",
" if \"embedding\" not in name and \"pooler\" not in name}\n",
"torch.save(dense_optimized_st, 'dbg/dense_squad_optimized.pt',)\n",
"print(\"Encoder Size (MB) - Sparse & Quantized - `torch.save`:\",\n",
" round(os.path.getsize(\"dbg/dense_squad_optimized.pt\")/1e6, 2))\n"
"dense_optimized_st = {\n",
" name: param for name, param in elementary_qtz_st.items() if \"embedding\" not in name and \"pooler\" not in name\n",
"}\n",
"torch.save(\n",
" dense_optimized_st,\n",
" \"dbg/dense_squad_optimized.pt\",\n",
")\n",
"print(\n",
" \"Encoder Size (MB) - Sparse & Quantized - `torch.save`:\",\n",
" round(os.path.getsize(\"dbg/dense_squad_optimized.pt\") / 1e6, 2),\n",
")"
]
},
{
@@ -287,7 +297,7 @@
"# Save the decomposed state_dict with an HDF5 file\n",
"# Saving only the encoder + QA Head\n",
"\n",
"with h5py.File('dbg/squad_sparse.h5','w') as hf:\n",
"with h5py.File(\"dbg/squad_sparse.h5\", \"w\") as hf:\n",
" for name, param in elementary_qtz_st.items():\n",
" if \"embedding\" in name:\n",
" print(f\"Skip {name}\")\n",
@@ -323,13 +333,13 @@
" hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
"\n",
"\n",
"with open('dbg/metadata.json', 'w') as f:\n",
"with open(\"dbg/metadata.json\", \"w\") as f:\n",
" f.write(json.dumps(qtz_st._metadata))\n",
"\n",
"size = os.path.getsize(\"dbg/squad_sparse.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
"print(\"\")\n",
"print(\"Encoder Size (MB) - Dense: \", round(dense_mb_size / 1e6, 2))\n",
"print(\"Encoder Size (MB) - Sparse & Quantized:\", round(size/1e6, 2))\n"
"print(\"Encoder Size (MB) - Sparse & Quantized:\", round(size / 1e6, 2))"
]
},
{
@@ -350,7 +360,7 @@
"# Save the decomposed state_dict to HDF5 storage\n",
"# Save everything in the architecutre (embedding + encoder + QA Head)\n",
"\n",
"with h5py.File('dbg/squad_sparse_with_embs.h5','w') as hf:\n",
"with h5py.File(\"dbg/squad_sparse_with_embs.h5\", \"w\") as hf:\n",
" for name, param in elementary_qtz_st.items():\n",
" # if \"embedding\" in name:\n",
" # print(f\"Skip {name}\")\n",
@@ -386,12 +396,11 @@
" hf.create_dataset(name, data=param, compression=\"gzip\", compression_opts=9)\n",
"\n",
"\n",
"\n",
"with open('dbg/metadata.json', 'w') as f:\n",
"with open(\"dbg/metadata.json\", \"w\") as f:\n",
" f.write(json.dumps(qtz_st._metadata))\n",
"\n",
"size = os.path.getsize(\"dbg/squad_sparse_with_embs.h5\") + os.path.getsize(\"dbg/metadata.json\")\n",
"print('\\nSize (MB):', round(size/1e6, 2))\n"
"print(\"\\nSize (MB):\", round(size / 1e6, 2))"
]
},
{
@@ -411,10 +420,10 @@
"\n",
"reconstructed_elementary_qtz_st = {}\n",
"\n",
"hf = h5py.File('dbg/squad_sparse_with_embs.h5','r')\n",
"hf = h5py.File(\"dbg/squad_sparse_with_embs.h5\", \"r\")\n",
"\n",
"for attr_name, attr_param in hf.attrs.items():\n",
" if 'shape' in attr_name:\n",
" if \"shape\" in attr_name:\n",
" attr_param = tuple(attr_param)\n",
" elif \".scale\" in attr_name:\n",
" if \"_packed_params\" in attr_name:\n",
@@ -489,22 +498,24 @@
" indices = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.indices\"]\n",
" shape = reconstructed_elementary_qtz_st[f\"{prefix_}.int_repr.shape\"]\n",
"\n",
" int_repr = sparse.csr_matrix(arg1=(data, indices, indptr),\n",
" shape=shape)\n",
" int_repr = sparse.csr_matrix(arg1=(data, indices, indptr), shape=shape)\n",
" int_repr = torch.tensor(int_repr.todense())\n",
"\n",
" scale = reconstructed_elementary_qtz_st[f\"{prefix_}.scale\"]\n",
" zero_point = reconstructed_elementary_qtz_st[f\"{prefix_}.zero_point\"]\n",
" weight = torch._make_per_tensor_quantized_tensor(int_repr,\n",
" scale,\n",
" zero_point)\n",
" weight = torch._make_per_tensor_quantized_tensor(int_repr, scale, zero_point)\n",
"\n",
" reconstructed_qtz_st[f\"{prefix_}\"] = weight\n",
" elif \"int_repr.data\" in name or \"int_repr.shape\" in name or \"int_repr.indices\" in name or \\\n",
" \"weight.scale\" in name or \"weight.zero_point\" in name:\n",
" elif (\n",
" \"int_repr.data\" in name\n",
" or \"int_repr.shape\" in name\n",
" or \"int_repr.indices\" in name\n",
" or \"weight.scale\" in name\n",
" or \"weight.zero_point\" in name\n",
" ):\n",
" continue\n",
" else:\n",
" reconstructed_qtz_st[name] = param\n"
" reconstructed_qtz_st[name] = param"
]
},
{
@@ -556,8 +567,8 @@
"source": [
"# Load the re-constructed state dict into a model\n",
"\n",
"dummy_model = BertForQuestionAnswering.from_pretrained('bert-base-uncased')\n",
"dummy_model.to('cpu')\n",
"dummy_model = BertForQuestionAnswering.from_pretrained(\"bert-base-uncased\")\n",
"dummy_model.to(\"cpu\")\n",
"\n",
"reconstructed_qtz_model = torch.quantization.quantize_dynamic(\n",
" model=dummy_model,\n",
@@ -566,7 +577,7 @@
")\n",
"\n",
"reconstructed_qtz_st = OrderedDict(reconstructed_qtz_st)\n",
"with open('dbg/metadata.json', 'r') as read_file:\n",
"with open(\"dbg/metadata.json\", \"r\") as read_file:\n",
" metadata = json.loads(read_file.read())\n",
"reconstructed_qtz_st._metadata = metadata\n",
"\n",

8
examples/research_projects/visual_bert/demo.ipynb
View File

@@ -40,7 +40,7 @@
"\n",
"\n",
"# for visualizing output\n",
"def showarray(a, fmt='jpeg'):\n",
"def showarray(a, fmt=\"jpeg\"):\n",
" a = np.uint8(np.clip(a, 0, 255))\n",
" f = io.BytesIO()\n",
" PIL.Image.fromarray(a).save(f, fmt)\n",
@@ -82,7 +82,7 @@
"image_preprocess = Preprocess(frcnn_cfg)\n",
"\n",
"bert_tokenizer = BertTokenizerFast.from_pretrained(\"bert-base-uncased\")\n",
"visualbert_vqa = VisualBertForQuestionAnswering.from_pretrained(\"uclanlp/visualbert-vqa\")\n"
"visualbert_vqa = VisualBertForQuestionAnswering.from_pretrained(\"uclanlp/visualbert-vqa\")"
],
"outputs": [
{
@@ -114,7 +114,7 @@
" scales_yx=scales_yx,\n",
" padding=\"max_detections\",\n",
" max_detections=frcnn_cfg.max_detections,\n",
" return_tensors=\"pt\"\n",
" return_tensors=\"pt\",\n",
")\n",
"# add boxes and labels to the image\n",
"\n",
@@ -189,7 +189,7 @@
" return_token_type_ids=True,\n",
" return_attention_mask=True,\n",
" add_special_tokens=True,\n",
" return_tensors=\"pt\"\n",
" return_tensors=\"pt\",\n",
" )\n",
"\n",
" output_vqa = visualbert_vqa(\n",

2
setup.py
View File

@@ -134,7 +134,7 @@ _deps = [
"sacremoses",
"sagemaker>=2.31.0",
"scikit-learn",
"sentencepiece==0.1.91",
"sentencepiece>=0.1.91,!=0.1.92",
"soundfile",
"sphinx-copybutton",
"sphinx-markdown-tables",

2
src/transformers/dependency_versions_table.py
View File

@@ -52,7 +52,7 @@ deps = {
"sacremoses": "sacremoses",
"sagemaker": "sagemaker>=2.31.0",
"scikit-learn": "scikit-learn",
"sentencepiece": "sentencepiece==0.1.91",
"sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
"soundfile": "soundfile",
"sphinx-copybutton": "sphinx-copybutton",
"sphinx-markdown-tables": "sphinx-markdown-tables",