Spaces:
Runtime error
Runtime error
File size: 9,414 Bytes
030a0f8 9320186 030a0f8 9852b1b e852933 030a0f8 e852933 030a0f8 e94abdd 030a0f8 d0289f9 9852b1b c2ef356 23b915b c2ef356 6ba5cb1 a905c49 030a0f8 e94abdd 030a0f8 9852b1b 23b915b a905c49 9852b1b c2ef356 4e3c392 1e2066e 9852b1b 030a0f8 124c8d3 030a0f8 e2f0b3b 030a0f8 e2f0b3b 030a0f8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 |
from typing import Optional, Union
import torch
import transformers
import streamlit as st
from plotly import graph_objects as go
from utils import get_lm
class Generator:
def __init__(self, lm_model_name, device, entropy=None):
self.device = device
self.tokenizer = transformers.AutoTokenizer.from_pretrained(
lm_model_name
)
self.lm = get_lm(lm_model_name).to(device)
self.lm.eval()
self.lm.config.pad_token_id = self.lm.config.eos_token_id
self.tokenizer.add_special_tokens(
{"pad_token": self.tokenizer.decode(self.lm.config.eos_token_id)}
)
self.caif_sampler = None
self.ordinary_sampler = None
self.entropy_based_stats = {
"skips": 0,
"avg_entropy": 0,
"count": 0,
}
self.entropy = entropy
def set_caif_sampler(self, sampler):
self.caif_sampler = sampler
def set_ordinary_sampler(self, sampler):
self.ordinary_sampler = sampler
def sample_sequences(
self,
num_samples: int,
input_prompt: Optional[str],
max_length: int,
caif_period: int,
caif_tokens_num: Union[int, None] = None,
entropy: float = None,
progress_bar=None,
**sampler_kwargs
):
self.entropy = entropy
input_ids, past, ended_sequences = self.get_input_ids(
input_prompt,
num_samples,
)
text = st.empty()
plot = st.empty()
gen_history = []
layout = go.Layout({
"xaxis": {
"title": "# Tokens"
},
"yaxis": {
"title": "Desired Attribute"
},
"plot_bgcolor": '#FFFFFF',
"template": "plotly_white",
"hovermode": "x",
})
inp_len = len(input_ids[0])
if self.caif_sampler is not None:
current_decoded = self.tokenizer.decode(input_ids[0])
probs = torch.exp(
self.caif_sampler.get_classifier_log_probs(
current_decoded, target_cls_id=sampler_kwargs["target_cls_id"]
)
).item()
gen_history += [probs]
for i in range(max_length):
is_caif_step = (
i % caif_period == 0 and self.caif_sampler is not None
)
input_ids, past, ended_sequences = self.generation_step(
input_ids,
past,
ended_sequences,
is_caif_step,
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
progress_bar.progress((i+1)/max_length)
if ended_sequences.all():
break
current_decoded = self.tokenizer.decode(input_ids[0])
if self.caif_sampler is not None:
probs = torch.exp(
self.caif_sampler.get_classifier_log_probs(
current_decoded, target_cls_id=sampler_kwargs["target_cls_id"]
)
).item()
gen_history += [probs]
scatter_data = go.Scatter({
"x": list(range(len(gen_history))),
"y": gen_history,
"hovertext": ["[PROMPT]"] + [self.tokenizer.decode(t) for t in input_ids[0][inp_len:]]
})
fig = go.Figure([scatter_data], layout=layout)
plot.plotly_chart(fig, use_container_width=True)
if i == 0:
with st.expander("What is it?"):
st.write("You can see how the probability of the desired attribute varies for every generation step.")
text.text(current_decoded)
return (
[
self.tokenizer.decode(sequence, skip_special_tokens=True)
for sequence in input_ids
],
input_ids,
)
def generation_step(
self,
input_ids,
past,
ended_sequences,
is_caif_step: bool,
caif_tokens_num=None,
**sampler_kwargs
):
prepared_inputs = self.lm.prepare_inputs_for_generation(
input_ids, past, use_cache=True
)
outputs = self.lm(
**prepared_inputs,
output_attentions=False,
output_hidden_states=False,
return_dict=True
)
past = outputs.past_key_values
if self.entropy is not None:
normalized = torch.nn.functional.log_softmax(
outputs.logits, dim=-1
)
p = torch.exp(normalized)
output_probs = p
output_information = -normalized
output_entropy = (output_probs * output_information).sum(-1)[:, -1]
batch_size = output_entropy.shape[0]
caif_mask = torch.ge(output_entropy, self.entropy)
ordinary_mask = ~caif_mask
self.entropy_based_stats["skips"] += caif_mask.sum() / batch_size
self.entropy_based_stats["count"] += 1
self.entropy_based_stats["avg_entropy"] += (
output_entropy.sum() / batch_size
)
flatten_entropy = output_entropy.view(-1).cpu().tolist()
if "entropy" not in self.entropy_based_stats.keys():
self.entropy_based_stats["entropy"] = flatten_entropy
else:
self.entropy_based_stats["entropy"] += flatten_entropy
if caif_mask.sum() == 0:
next_tokens_sampler = self.ordinary_sampler
next_tokens = next_tokens_sampler(
input_ids,
outputs.logits,
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
next_tokens = (
next_tokens * (1 - ended_sequences.long())
+ self.lm.config.eos_token_id * ended_sequences.long()
).long()
elif caif_mask.sum() == batch_size:
next_tokens_sampler = self.caif_sampler
next_tokens = next_tokens_sampler(
input_ids,
outputs.logits,
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
next_tokens = (
next_tokens * (1 - ended_sequences.long())
+ self.lm.config.eos_token_id * ended_sequences.long()
).long()
else:
next_tokens_caif = self.caif_sampler(
input_ids[caif_mask],
outputs.logits[caif_mask],
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
next_tokens_ordinary = self.ordinary_sampler(
input_ids[ordinary_mask],
outputs.logits[ordinary_mask],
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
next_tokens_caif = (
next_tokens_caif * (1 - ended_sequences[caif_mask].long())
+ self.lm.config.eos_token_id
* ended_sequences[caif_mask].long()
).long()
next_tokens_ordinary = (
next_tokens_ordinary
* (1 - ended_sequences[ordinary_mask].long())
+ self.lm.config.eos_token_id
* ended_sequences[ordinary_mask].long()
).long()
next_tokens = torch.ones(batch_size).long().to(self.device)
next_tokens[caif_mask] = next_tokens_caif
next_tokens[ordinary_mask] = next_tokens_ordinary
else:
if is_caif_step:
next_tokens_sampler = self.caif_sampler
else:
next_tokens_sampler = self.ordinary_sampler
next_tokens = next_tokens_sampler(
input_ids,
outputs.logits,
caif_tokens_num=caif_tokens_num,
**sampler_kwargs
)
next_tokens = (
next_tokens * (1 - ended_sequences.long())
+ self.lm.config.eos_token_id * ended_sequences.long()
).long()
input_ids = torch.cat(
[input_ids, next_tokens[:, None].to(self.device)], dim=-1
)
ended_sequences += next_tokens == self.lm.config.eos_token_id
return input_ids, past, ended_sequences
def get_input_ids(self, input_prompt, num_samples):
#input_ids = torch.tensor([[self.lm.config.bos_token_id]])
if input_prompt is not None:
input_prompt = self.tokenizer(
input_prompt, return_tensors="pt"
).input_ids
input_ids = input_prompt
input_ids = input_ids.repeat(num_samples, 1).to(self.device)
past = None
ended_sequences = torch.zeros(
input_ids.shape[0], device=self.device
).bool()
return input_ids, past, ended_sequences
@staticmethod
def sample(unscaled_probs, values):
samples = torch.multinomial(unscaled_probs, 1)
return torch.take_along_dim(values, samples, dim=1)
|