{
"cells": [
{
"cell_type": "code",
"execution_count": 51,
"id": "1eccc83e-bc68-4082-a3cc-b055779b6ee8",
"metadata": {},
"outputs": [],
"source": [
"# References:\n",
"# https://www.tanishq.ai/blog/posts/2021-11-16-gradio-huggingface.html"
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "5b74867e-7ec1-4cda-9d96-0f5cd9cd4810",
"metadata": {},
"outputs": [],
"source": [
"import numpy as np\n",
"import pandas as pd\n",
"import gradio as gr\n",
"import torch\n",
"from torch import nn\n",
"import pickle\n",
"from torch import tensor\n",
"import torch.nn.functional as F\n",
"import pandas as pd"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "7d6e9e70-83fe-4209-8f06-6542cf6ba11b",
"metadata": {},
"outputs": [],
"source": [
"with open(\"meta.pkl\", \"rb\") as f:\n",
" meta = pickle.load(f)\n",
"t2i = meta['t2i']\n",
"i2t = meta['i2t']\n",
"encode = lambda x: [t2i[c] for c in x]\n",
"decode = lambda x: \"\".join([i2t[i] for i in x])"
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "c4a0b480-6775-4d82-9395-9b5a455012ad",
"metadata": {},
"outputs": [],
"source": [
"batch_size = 128 # B, batch size\n",
"block_size = 48 # T, context len for poem is shorter, to set to 48\n",
"vocab_size = len(t2i.keys())\n",
"nn_emb_size = 64 # nn_emb\n",
"n_head = 16\n",
"n_layers = 8\n",
"\n",
"#device = \"cuda\"\n",
"devicd = \"cpu\""
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "0e4e72ce-5f61-4831-b7e8-703ed171936b",
"metadata": {},
"outputs": [],
"source": [
"def encode_pad(s):\n",
" if len(s) >= block_size:\n",
" sample = s[:block_size]\n",
" else:\n",
" sample = s\n",
" sample = encode(s)\n",
" sample = [0]*(block_size-len(sample)) + sample \n",
" inp = tensor(sample[:block_size])[None,...]\n",
" return inp"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "a9bc886f-4ec8-458a-b847-c9996df57fa9",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"Model(\n",
" (tk_emb): Embedding(7475, 64)\n",
" (pos_emb): Embedding(48, 64)\n",
" (ln): LayerNorm((64,), eps=1e-05, elementwise_affine=True)\n",
" (attention_blocks): ModuleList(\n",
" (0-7): 8 x AttentionBlock(\n",
" (emb_proj): Linear(in_features=64, out_features=192, bias=True)\n",
" (ln_1): LayerNorm((64,), eps=1e-05, elementwise_affine=True)\n",
" (mult_head): MultiheadAttention(\n",
" (out_proj): NonDynamicallyQuantizableLinear(in_features=64, out_features=64, bias=True)\n",
" )\n",
" (ln_2): LayerNorm((64,), eps=1e-05, elementwise_affine=True)\n",
" (ff): Sequential(\n",
" (0): Linear(in_features=64, out_features=256, bias=True)\n",
" (1): GELU(approximate='none')\n",
" (2): Dropout(p=0.2, inplace=False)\n",
" (3): Linear(in_features=256, out_features=64, bias=True)\n",
" (4): GELU(approximate='none')\n",
" (5): Dropout(p=0.2, inplace=False)\n",
" )\n",
" )\n",
" )\n",
" (ln_h): Linear(in_features=64, out_features=7475, bias=True)\n",
")"
]
},
"execution_count": 9,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"class AttentionBlock(nn.Module):\n",
" def __init__(self, nn_emb = nn_emb_size, block_size = block_size, n_head = n_head):\n",
" super().__init__()\n",
" self.nn_emb = nn_emb_size\n",
" self.block_size = block_size\n",
" self.n_head = n_head\n",
"\n",
" self.emb_proj = nn.Linear(nn_emb, nn_emb * 3)\n",
" self.ln_1 = nn.LayerNorm(nn_emb) \n",
" self.mult_head = nn.MultiheadAttention(nn_emb, n_head, dropout=0.2, batch_first=True)\n",
" self.ln_2 = nn.LayerNorm(nn_emb) \n",
" self.ff = nn.Sequential(nn.Linear(nn_emb, nn_emb * 4),nn.GELU(), nn.Dropout(0.2), nn.Linear(nn_emb * 4, nn_emb), nn.GELU(), nn.Dropout(0.2))\n",
"\n",
" def forward(self,x): # (B, T, nn_emb)\n",
" x1 = x\n",
" x = self.emb_proj(x) # (B, T, nn_emb*3)\n",
" q,k,v = x.split(self.nn_emb, dim=2)\n",
" x,_ = self.mult_head(q, k, v, key_padding_mask=None, need_weights=False, attn_mask=torch.nn.Transformer.generate_square_subsequent_mask(self.nn_emb), average_attn_weights=True, is_causal=True) # (B,T,nn_emb)\n",
" x = x+x1\n",
" x = self.ff(self.ln_2(x)) + x\n",
" return x\n",
" \n",
" \n",
"class Model(nn.Module):\n",
" def __init__(self, nn_emb = nn_emb_size, block_size = block_size,vocab_size = vocab_size, n_head = n_head, n_layers = n_layers): \n",
" super().__init__()\n",
" self.vocab_size = vocab_size\n",
" self.block_size = block_size\n",
" self.nn_emb = nn_emb\n",
" self.n_head = n_head\n",
" self.n_layers = n_layers\n",
" \n",
" self.tk_emb = nn.Embedding(vocab_size, nn_emb)\n",
" self.pos_emb = nn.Embedding(block_size, nn_emb)\n",
" self.ln = nn.LayerNorm(nn_emb)\n",
" #self.emb_proj = nn.Linear(nn_emb, nn_emb * 3)\n",
" #self.atten = nn.MultiheadAttention(nn_emb, n_head, dropout=0.2, batch_first=True)\n",
" self.attention_blocks = nn.ModuleList( [AttentionBlock(nn_emb, block_size, n_head)] * n_layers)\n",
" #self.h = nn.Sequential(nn.Linear(nn_emb, nn_emb),nn.GELU(), nn.Dropout(0.2), nn.Linear(nn_emb, nn_emb), nn.GELU(), nn.Dropout(0.2))\n",
" self.ln_h = nn.Linear(nn_emb, self.vocab_size)\n",
"\n",
" def forward(self, inp, targ = None): # inp is (B, T), targ is (B, T)\n",
" inp.to(device)\n",
" tk = self.tk_emb(inp) # (B,T,nn_emb)\n",
" positions = torch.arange(self.block_size).to(device)\n",
" #print(positions)\n",
" pos = self.pos_emb(positions) # (T,nn_emb)\n",
" x = tk + pos # (B,T,nn_emb)\n",
" #x = self.ln(x) \n",
" #a = x\n",
" #x = self.emb_proj(x) # (B,t,nn_emb*3)\n",
" for blk in self.attention_blocks:\n",
" x = blk(x)\n",
" #q,k,v = x.split(self.nn_emb, dim=2)\n",
" #x,_ = self.atten(q, k, v, key_padding_mask=None, need_weights=False, attn_mask=torch.nn.Transformer.generate_square_subsequent_mask(self.nn_emb), average_attn_weights=True, is_causal=True) # (B,T,nn_emb)\n",
" #x = x + a\n",
" #x = self.ln(x) \n",
" #x = x+self.h(x) # (B,T,nn_emb)\n",
" x = self.ln(x) # (B,T,nn_emb) \n",
" x = self.ln_h(x) # (B,T,vocab_size)\n",
" if targ == None:\n",
" loss = None\n",
" else:\n",
" targ.to(device)\n",
" loss = F.cross_entropy(x.view(-1, x.shape[-1]), targ.view(-1))\n",
" return x, loss\n",
"\n",
"m = Model()\n",
"m.to(device)"
]
},
{
"cell_type": "code",
"execution_count": 20,
"id": "95545bf7-51fa-45a8-b34d-0231aa95e300",
"metadata": {},
"outputs": [],
"source": [
"with open(\"model_v4.pkl\",\"rb\") as f:\n",
" m=pickle.load(f)"
]
},
{
"cell_type": "code",
"execution_count": 21,
"id": "c2393e78-a1c6-4671-9170-4ea33cdb50d1",
"metadata": {},
"outputs": [],
"source": [
"top_k = 20\n",
"def generate(s, num = 60):\n",
"\n",
" for i in range(num + num):\n",
" inp = s[-block_size:]\n",
" inp = encode_pad(inp).to(device)\n",
" out, loss = m(inp)\n",
" out = out[:,-1,:]\n",
" if top_k is not None:\n",
" v, _ = torch.topk(out, min(top_k, out.size(-1)))\n",
" out[out < v[:, [-1]]] = -float('Inf') \n",
" prob = torch.softmax(out[:,:], dim=-1)\n",
" g = torch.multinomial(prob, num_samples=1)\n",
" next_c = i2t[g[0].item()]\n",
" if next_c in s and next_c != '。' and next_c != '，':\n",
" continue\n",
" s = s + next_c\n",
"\n",
" if (len(s) > num and s[-1] == \"。\"):\n",
" break\n",
" return s"
]
},
{
"cell_type": "code",
"execution_count": 24,
"id": "170b95ca-74b9-4360-84cc-6a8dfa3f8c42",
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"'终南。若问黄云一路在，更有东城上去时。不须为别故园庐，独坐江山半夜凉。此地无馀春树晚，今朝日暮向来迟。西北天津长望后，三湘月下烟中。'"
]
},
"execution_count": 24,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"generate('终南。')"
]
},
{
"cell_type": "code",
"execution_count": 26,
"id": "edca19ab-087b-4368-84d0-8eee7388c200",
"metadata": {
"scrolled": true
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Running on local URL: http://127.0.0.1:7867\n",
"\n",
"To create a public link, set `share=True` in `launch()`.\n"
]
},
{
"data": {
"text/html": [
""
],
"text/plain": [
""
]
},
"metadata": {},
"output_type": "display_data"
},
{
"data": {
"text/plain": []
},
"execution_count": 26,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"\n",
"inputs = [gr.Textbox(label=\"Input\",\n",
" info=\"Enter some Chinese text to start generate\",\n",
" lines=3,\n",
" value=\"终南。\",)]\n",
"\n",
"outputs = [ gr.Textbox(\n",
" label=\"Output\",\n",
" info=\"Generated Poem\",\n",
" lines=3,\n",
" value=\"\", )]\n",
"gr.Interface(fn=generate, inputs=inputs, outputs=outputs, title=\"Enter Chinese text to generate Chinese Poem.\").launch(share=False)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "6112eaea-16d6-4d43-8b95-3999c605643b",
"metadata": {},
"outputs": [],
"source": []
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.8.10"
}
},
"nbformat": 4,
"nbformat_minor": 5
}