Upload train_and_inference.ipynb
Browse files- code/train_and_inference.ipynb +107 -0
code/train_and_inference.ipynb
ADDED
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
+
"metadata": {
|
7 |
+
"id": "_nSGBgG98qRC"
|
8 |
+
},
|
9 |
+
"outputs": [],
|
10 |
+
"source": [
|
11 |
+
"# Trainer\n",
|
12 |
+
"import torch\n",
|
13 |
+
"from tqdm import tqdm\n",
|
14 |
+
"\n",
|
15 |
+
"iterator = tqdm(dataloader, desc=\"Training\", postfix={\"train_loss\":0.0})\n",
|
16 |
+
"\n",
|
17 |
+
"for item in iterator:\n",
|
18 |
+
" item = tokenizer.bos_token + \" \" + item[0] + \" \" + tokenizer.eos_token\n",
|
19 |
+
" encoded_inp = tokenizer(item, return_tensors='pt').input_ids.to(\"cuda\")\n",
|
20 |
+
" logits = mamba_model(encoded_inp)\n",
|
21 |
+
"\n",
|
22 |
+
" labels = encoded_inp.to(logits.device)\n",
|
23 |
+
" shift_logits = logits[:, :-1, :].contiguous()\n",
|
24 |
+
" labels = labels[:, 1:].contiguous()\n",
|
25 |
+
" loss_fct = torch.nn.CrossEntropyLoss()\n",
|
26 |
+
" loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1))\n",
|
27 |
+
"\n",
|
28 |
+
" optimizer.zero_grad(set_to_none=True)\n",
|
29 |
+
" loss.backward()\n",
|
30 |
+
" optimizer.step()\n",
|
31 |
+
"\n",
|
32 |
+
" # moving data's from gpu to cpu\n",
|
33 |
+
" loss = loss.detach().cpu().numpy()\n",
|
34 |
+
" logits = logits.detach().cpu().numpy()\n",
|
35 |
+
" labels = labels.detach().cpu().numpy()\n",
|
36 |
+
" encoded_inp = encoded_inp.detach().cpu().numpy()\n",
|
37 |
+
" shift_logits = shift_logits.detach().cpu().numpy()\n",
|
38 |
+
"\n",
|
39 |
+
" iterator.set_postfix({\"train_loss\": loss.item()}, refresh=False)"
|
40 |
+
]
|
41 |
+
},
|
42 |
+
{
|
43 |
+
"cell_type": "code",
|
44 |
+
"execution_count": 14,
|
45 |
+
"metadata": {
|
46 |
+
"id": "feaR0XKtOGug"
|
47 |
+
},
|
48 |
+
"outputs": [],
|
49 |
+
"source": [
|
50 |
+
"# Inference\n",
|
51 |
+
"import torch\n",
|
52 |
+
"import torch.nn.functional as F\n",
|
53 |
+
"\n",
|
54 |
+
"\n",
|
55 |
+
"def generate(model,\n",
|
56 |
+
" tokenizer,\n",
|
57 |
+
" prompt: str,\n",
|
58 |
+
" n_tokens_to_gen: int = 200,\n",
|
59 |
+
" sample: bool = True,\n",
|
60 |
+
" top_k: int = 40):\n",
|
61 |
+
" model.eval()\n",
|
62 |
+
"\n",
|
63 |
+
" input_ids = tokenizer(prompt, return_tensors='pt').input_ids.to(\"cuda\")\n",
|
64 |
+
"\n",
|
65 |
+
" for token_n in range(n_tokens_to_gen):\n",
|
66 |
+
" with torch.no_grad():\n",
|
67 |
+
" indices_to_input = input_ids\n",
|
68 |
+
" next_token_logits = mamba_model(indices_to_input)[:, -1]\n",
|
69 |
+
"\n",
|
70 |
+
" probs = F.softmax(next_token_logits, dim=-1)\n",
|
71 |
+
" (batch, vocab_size) = probs.shape\n",
|
72 |
+
"\n",
|
73 |
+
" if top_k is not None:\n",
|
74 |
+
" (values, indices) = torch.topk(probs, k=top_k)\n",
|
75 |
+
" probs[probs < values[:, -1, None]] = 0\n",
|
76 |
+
" probs = probs / probs.sum(axis=1, keepdims=True)\n",
|
77 |
+
"\n",
|
78 |
+
" if sample:\n",
|
79 |
+
" next_indices = torch.multinomial(probs, num_samples=1)\n",
|
80 |
+
" else:\n",
|
81 |
+
" next_indices = torch.argmax(probs, dim=-1)[:, None]\n",
|
82 |
+
"\n",
|
83 |
+
" input_ids = torch.cat([input_ids, next_indices], dim=1)\n",
|
84 |
+
"\n",
|
85 |
+
" output_completions = [tokenizer.decode(output.tolist()) for output in input_ids][0]\n",
|
86 |
+
"\n",
|
87 |
+
" return output_completions"
|
88 |
+
]
|
89 |
+
}
|
90 |
+
],
|
91 |
+
"metadata": {
|
92 |
+
"accelerator": "GPU",
|
93 |
+
"colab": {
|
94 |
+
"gpuType": "T4",
|
95 |
+
"provenance": []
|
96 |
+
},
|
97 |
+
"kernelspec": {
|
98 |
+
"display_name": "Python 3",
|
99 |
+
"name": "python3"
|
100 |
+
},
|
101 |
+
"language_info": {
|
102 |
+
"name": "python"
|
103 |
+
}
|
104 |
+
},
|
105 |
+
"nbformat": 4,
|
106 |
+
"nbformat_minor": 0
|
107 |
+
}
|