MengniWang
commited on
Commit
•
03e8016
1
Parent(s):
aa281eb
add batch inference code
Browse files- evaluation.ipynb +100 -0
evaluation.ipynb
CHANGED
@@ -103,6 +103,106 @@
|
|
103 |
"print('acc: ', acc)"
|
104 |
]
|
105 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
106 |
{
|
107 |
"attachments": {},
|
108 |
"cell_type": "markdown",
|
|
|
103 |
"print('acc: ', acc)"
|
104 |
]
|
105 |
},
|
106 |
+
{
|
107 |
+
"cell_type": "code",
|
108 |
+
"execution_count": null,
|
109 |
+
"metadata": {
|
110 |
+
"vscode": {
|
111 |
+
"languageId": "plaintext"
|
112 |
+
}
|
113 |
+
},
|
114 |
+
"outputs": [],
|
115 |
+
"source": [
|
116 |
+
"# batch inference\n",
|
117 |
+
"\n",
|
118 |
+
"from transformers import AutoTokenizer\n",
|
119 |
+
"import torch\n",
|
120 |
+
"import numpy as np\n",
|
121 |
+
"from datasets import load_dataset\n",
|
122 |
+
"import onnxruntime as ort\n",
|
123 |
+
"from torch.nn.functional import pad\n",
|
124 |
+
"from torch.utils.data import DataLoader\n",
|
125 |
+
"\n",
|
126 |
+
"batch_size = 2\n",
|
127 |
+
"pad_max = 196\n",
|
128 |
+
"\n",
|
129 |
+
"# load model\n",
|
130 |
+
"model_id = \"EleutherAI/gpt-j-6B\"\n",
|
131 |
+
"tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
|
132 |
+
"\n",
|
133 |
+
"def tokenize_function(examples):\n",
|
134 |
+
" example = tokenizer(examples['text'])\n",
|
135 |
+
" return example\n",
|
136 |
+
"\n",
|
137 |
+
"# create dataloader\n",
|
138 |
+
"class Dataloader:\n",
|
139 |
+
" def __init__(self, pad_max=196, batch_size=1, sub_folder='validation'):\n",
|
140 |
+
" self.pad_max = pad_max\n",
|
141 |
+
" self.batch_size=batch_size\n",
|
142 |
+
" dataset = load_dataset('lambada', split=sub_folder)\n",
|
143 |
+
" dataset = dataset.map(tokenize_function, batched=True)\n",
|
144 |
+
" dataset.set_format(type=\"torch\", columns=[\"input_ids\", \"attention_mask\"])\n",
|
145 |
+
" self.dataloader = DataLoader(\n",
|
146 |
+
" dataset,\n",
|
147 |
+
" batch_size=self.batch_size,\n",
|
148 |
+
" shuffle=False,\n",
|
149 |
+
" collate_fn=self.collate_batch,\n",
|
150 |
+
" )\n",
|
151 |
+
"\n",
|
152 |
+
" def collate_batch(self, batch):\n",
|
153 |
+
" input_ids_padded = []\n",
|
154 |
+
" attention_mask_padded = []\n",
|
155 |
+
" last_ind = []\n",
|
156 |
+
" for text in batch:\n",
|
157 |
+
" input_ids = text[\"input_ids\"] if text[\"input_ids\"].shape[0] <= self.pad_max else text[\"input_ids\"][0:int(self.pad_max-1)]\n",
|
158 |
+
" pad_len = self.pad_max - input_ids.shape[0]\n",
|
159 |
+
" last_ind.append(input_ids.shape[0] - 1)\n",
|
160 |
+
" input_ids = pad(input_ids, (0, pad_len), value=1)\n",
|
161 |
+
" input_ids_padded.append(input_ids)\n",
|
162 |
+
" attention_mask = torch.ones(input_ids.shape[0] + 1)\n",
|
163 |
+
" attention_mask_padded.append(attention_mask)\n",
|
164 |
+
" return (torch.vstack(input_ids_padded), torch.vstack(attention_mask_padded)), torch.tensor(last_ind)\n",
|
165 |
+
"\n",
|
166 |
+
" def __iter__(self):\n",
|
167 |
+
" try:\n",
|
168 |
+
" for (input_ids, attention_mask), last_ind in self.dataloader:\n",
|
169 |
+
" data = [input_ids.detach().cpu().numpy().astype('int64')]\n",
|
170 |
+
" data.append(attention_mask.detach().cpu().numpy().astype('int64'))\n",
|
171 |
+
" yield data, last_ind.detach().cpu().numpy()\n",
|
172 |
+
" except StopIteration:\n",
|
173 |
+
" return\n",
|
174 |
+
"\n",
|
175 |
+
"# create session\n",
|
176 |
+
"options = ort.SessionOptions()\n",
|
177 |
+
"options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL\n",
|
178 |
+
"session = ort.InferenceSession('/path/to/model.onnx', options, providers=ort.get_available_providers())\n",
|
179 |
+
"total, hit = 0, 0\n",
|
180 |
+
"\n",
|
181 |
+
"dataloader = Dataloader(pad_max=pad_max, batch_size=batch_size)\n",
|
182 |
+
"\n",
|
183 |
+
"# inference\n",
|
184 |
+
"for idx, (batch, last_ind) in enumerate(dataloader):\n",
|
185 |
+
" label = torch.from_numpy(batch[0][torch.arange(len(last_ind)), last_ind])\n",
|
186 |
+
" pad_len = pad_max - last_ind - 1\n",
|
187 |
+
" ort_inputs = {\n",
|
188 |
+
" 'input_ids': batch[0],\n",
|
189 |
+
" 'attention_mask': batch[1]\n",
|
190 |
+
" }\n",
|
191 |
+
" for i in range(28):\n",
|
192 |
+
" ort_inputs[\"past_key_values.{}.key\".format(i)] = np.zeros((batch_size,16,1,256), dtype='float32')\n",
|
193 |
+
" ort_inputs[\"past_key_values.{}.value\".format(i)] = np.zeros((batch_size,16,1,256), dtype='float32')\n",
|
194 |
+
" \n",
|
195 |
+
" predictions = session.run(None, ort_inputs)\n",
|
196 |
+
" outputs = torch.from_numpy(predictions[0])\n",
|
197 |
+
" last_token_logits = outputs[torch.arange(len(last_ind)), -2 - pad_len, :]\n",
|
198 |
+
" pred = last_token_logits.argmax(dim=-1)\n",
|
199 |
+
" total += len(label)\n",
|
200 |
+
" hit += (pred == label).sum().item()\n",
|
201 |
+
"\n",
|
202 |
+
"acc = hit / total\n",
|
203 |
+
"print('acc: ', acc)"
|
204 |
+
]
|
205 |
+
},
|
206 |
{
|
207 |
"attachments": {},
|
208 |
"cell_type": "markdown",
|