add script
Browse filesSigned-off-by: Meng, Hengyu <hengyu.meng@intel.com>
- evaluation.ipynb +159 -0
evaluation.ipynb
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"attachments": {},
|
5 |
+
"cell_type": "markdown",
|
6 |
+
"metadata": {},
|
7 |
+
"source": [
|
8 |
+
"## Introduction\n",
|
9 |
+
"\n",
|
10 |
+
"This tutorial demonstrates how to perform evaluation on a gpt-j-6B-int8 model."
|
11 |
+
]
|
12 |
+
},
|
13 |
+
{
|
14 |
+
"attachments": {},
|
15 |
+
"cell_type": "markdown",
|
16 |
+
"metadata": {},
|
17 |
+
"source": [
|
18 |
+
"## Prerequisite"
|
19 |
+
]
|
20 |
+
},
|
21 |
+
{
|
22 |
+
"cell_type": "code",
|
23 |
+
"execution_count": null,
|
24 |
+
"metadata": {
|
25 |
+
"vscode": {
|
26 |
+
"languageId": "plaintext"
|
27 |
+
}
|
28 |
+
},
|
29 |
+
"outputs": [],
|
30 |
+
"source": [
|
31 |
+
"!pip install onnx onnxruntime torch transformers datasets accelerate"
|
32 |
+
]
|
33 |
+
},
|
34 |
+
{
|
35 |
+
"attachments": {},
|
36 |
+
"cell_type": "markdown",
|
37 |
+
"metadata": {},
|
38 |
+
"source": [
|
39 |
+
"## Run\n",
|
40 |
+
"\n",
|
41 |
+
"### 1. Get lambada acc"
|
42 |
+
]
|
43 |
+
},
|
44 |
+
{
|
45 |
+
"cell_type": "code",
|
46 |
+
"execution_count": null,
|
47 |
+
"metadata": {
|
48 |
+
"vscode": {
|
49 |
+
"languageId": "plaintext"
|
50 |
+
}
|
51 |
+
},
|
52 |
+
"outputs": [],
|
53 |
+
"source": [
|
54 |
+
"from transformers import AutoTokenizer\n",
|
55 |
+
"import torch\n",
|
56 |
+
"from datasets import load_dataset\n",
|
57 |
+
"import onnxruntime as ort\n",
|
58 |
+
"from torch.nn.functional import pad\n",
|
59 |
+
"\n",
|
60 |
+
"# load model\n",
|
61 |
+
"model_id = \"EleutherAI/gpt-j-6B\"\n",
|
62 |
+
"tokenizer = AutoTokenizer.from_pretrained(model_id)\n",
|
63 |
+
"\n",
|
64 |
+
"def tokenize_function(examples):\n",
|
65 |
+
" example = tokenizer(examples['text'])\n",
|
66 |
+
" return example\n",
|
67 |
+
"\n",
|
68 |
+
"# create dataset\n",
|
69 |
+
"dataset = load_dataset('lambada', split='validation')\n",
|
70 |
+
"dataset = dataset.shuffle(seed=42)\n",
|
71 |
+
"dataset = dataset.map(tokenize_function, batched=True)\n",
|
72 |
+
"dataset.set_format(type='torch', columns=['input_ids'])\n",
|
73 |
+
"\n",
|
74 |
+
"# create session\n",
|
75 |
+
"options = ort.SessionOptions()\n",
|
76 |
+
"options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL\n",
|
77 |
+
"session = ort.InferenceSession('/path/to/model.onnx', options, providers=ort.get_available_providers())\n",
|
78 |
+
"total, hit = 0, 0\n",
|
79 |
+
"index = 1\n",
|
80 |
+
"\n",
|
81 |
+
"# inference\n",
|
82 |
+
"for idx, batch in enumerate(dataset):\n",
|
83 |
+
" input_ids = batch['input_ids'].unsqueeze(0)\n",
|
84 |
+
" label = input_ids[:, -1]\n",
|
85 |
+
" pad_len = 0 ##set to 0\n",
|
86 |
+
" input_ids = pad(input_ids, (0, pad_len), value=1)\n",
|
87 |
+
" ort_inputs = {\n",
|
88 |
+
" 'input_ids': input_ids.detach().cpu().numpy(),\n",
|
89 |
+
" 'attention_mask': torch.ones(input_ids.shape).detach().cpu().numpy().astype('int64')\n",
|
90 |
+
" }\n",
|
91 |
+
" predictions = session.run(None, ort_inputs)\n",
|
92 |
+
" outputs = torch.from_numpy(predictions[0]) \n",
|
93 |
+
" last_token_logits = outputs[:, -2 - pad_len, :]\n",
|
94 |
+
" pred = last_token_logits.argmax(dim=-1)\n",
|
95 |
+
" total += label.size(0)\n",
|
96 |
+
" hit += (pred == label).sum().item()\n",
|
97 |
+
"acc = hit / total\n",
|
98 |
+
"print('acc: ', acc)"
|
99 |
+
]
|
100 |
+
},
|
101 |
+
{
|
102 |
+
"attachments": {},
|
103 |
+
"cell_type": "markdown",
|
104 |
+
"metadata": {},
|
105 |
+
"source": [
|
106 |
+
"### 2. Text Generation"
|
107 |
+
]
|
108 |
+
},
|
109 |
+
{
|
110 |
+
"cell_type": "code",
|
111 |
+
"execution_count": null,
|
112 |
+
"metadata": {
|
113 |
+
"vscode": {
|
114 |
+
"languageId": "plaintext"
|
115 |
+
}
|
116 |
+
},
|
117 |
+
"outputs": [],
|
118 |
+
"source": [
|
119 |
+
"import os\n",
|
120 |
+
"import time\n",
|
121 |
+
"import sys\n",
|
122 |
+
"\n",
|
123 |
+
"# create session\n",
|
124 |
+
"sess_options = ort.SessionOptions()\n",
|
125 |
+
"sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL\n",
|
126 |
+
"session = ort.InferenceSession('/path/to/model.onnx', sess_options)\n",
|
127 |
+
"\n",
|
128 |
+
"# input prompt\n",
|
129 |
+
"# 32 tokens input\n",
|
130 |
+
"prompt = \"Once upon a time, there existed a little girl, who liked to have adventures.\" + \\\n",
|
131 |
+
" \" She wanted to go to places and meet new people, and have fun.\"\n",
|
132 |
+
"\n",
|
133 |
+
"print(\"prompt: \", prompt)\n",
|
134 |
+
"\n",
|
135 |
+
"# start\n",
|
136 |
+
"input_ids = tokenizer(prompt, return_tensors=\"pt\").input_ids\n",
|
137 |
+
"for i in range(32):\n",
|
138 |
+
" inp = {'input_ids': input_ids.detach().cpu().numpy(),\n",
|
139 |
+
" 'attention_mask': torch.ones(input_ids.shape).detach().cpu().numpy().astype('int64')}\n",
|
140 |
+
" output = session.run(None, inp)\n",
|
141 |
+
" logits = output[0]\n",
|
142 |
+
" logits = torch.from_numpy(logits)\n",
|
143 |
+
" next_token_logits = logits[:, -1, :]\n",
|
144 |
+
" probs = torch.nn.functional.softmax(next_token_logits, dim=-1)\n",
|
145 |
+
" next_tokens = torch.argmax(probs, dim=-1)\n",
|
146 |
+
" input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)\n",
|
147 |
+
"print(tokenizer.decode(input_ids[0]))"
|
148 |
+
]
|
149 |
+
}
|
150 |
+
],
|
151 |
+
"metadata": {
|
152 |
+
"language_info": {
|
153 |
+
"name": "python"
|
154 |
+
},
|
155 |
+
"orig_nbformat": 4
|
156 |
+
},
|
157 |
+
"nbformat": 4,
|
158 |
+
"nbformat_minor": 2
|
159 |
+
}
|