File size: 11,396 Bytes
fb238e8 |
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 |
# encoding=utf-8
import torch, math
import torch.nn.functional as F
def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
# This function has been mostly taken from huggingface conversational ai code at
# https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313
if top_k > 0:
# Remove all tokens with a probability less than the last token of the top-k
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = filter_value
if top_p > 0.0:
# convert to 1D
sorted_logits, sorted_indices = torch.sort(logits, dim=-1, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above the threshold
sorted_indices_to_remove = cumulative_probs > top_p
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
for i in range(sorted_indices.size()[0]):
indices_to_remove = sorted_indices[i][sorted_indices_to_remove[i]]
logits[i][indices_to_remove] = filter_value
return logits
def enforce_repetition_penalty(lprobs, prev_output_tokens, repetition_penalty=1.5):
"""repetition penalty (from CTRL paper https://arxiv.org/abs/1909.05858). """
for previous_token in set(prev_output_tokens):
# if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
if lprobs[previous_token] < 0:
lprobs[previous_token] *= repetition_penalty
else:
lprobs[previous_token] /= repetition_penalty
def switch(next_value, init, is_update): # 换成真实token
is_update = is_update.type_as(next_value)
return (1-is_update)*init + is_update*next_value
def get_atten_mask(batch_size, seq_length, memory_length=0):
memory_attention_mask = torch.ones(
(batch_size, 1, seq_length, seq_length + memory_length), dtype=torch.int16)
memory_attention_mask = torch.tril(
torch.triu(memory_attention_mask, 1 - seq_length + memory_length), memory_length)
return memory_attention_mask # [bs, 1, seq_len, seq_len+M]
def get_masks_and_position_ids(data, mem_length=None):
# Extract batch size and sequence length.
batch_size, seq_length = data.size()
# Attention mask (lower triangular).
attention_mask = torch.ones((1, seq_length, seq_length + mem_length), device=data.device)
attention_mask = torch.tril(torch.triu(attention_mask, 1 - seq_length + mem_length), mem_length)
attention_mask = attention_mask.unsqueeze(1)
# Position ids.
position_ids = torch.arange(seq_length, dtype=torch.long,
device=data.device)
position_ids = position_ids.unsqueeze(0).expand_as(data)
return attention_mask, position_ids
def sample_sequence_batch(model, context_tokens_tensor, context_length_tensor, max_out_seq=None, mems=None,
end_token_id=None, repetition_penalty=1.0, temperature=1.0, top_k=0, top_p=0.0):
"""_summary_
Args:
model (_type_): _description_
context_tokens_tensor (Tensor): [bs, seq_len]
context_length_tensor (Tensor): [bs, ]
max_out_seq (_type_, optional): _description_. Defaults to None.
mems (_type_, optional): _description_. Defaults to None.
end_token_id (_type_, optional): _description_. Defaults to None.
repetition_penalty (float, optional): _description_. Defaults to 1.0.
temperature (float, optional): _description_. Defaults to 1.0.
top_k (int, optional): _description_. Defaults to 0.
top_p (float, optional): _description_. Defaults to 0.0.
Returns:
_type_: _description_
"""
model_dtype = next(model.parameters()).dtype
org_context_length = torch.min(context_length_tensor).item()
batch_size = context_tokens_tensor.shape[0]
tokens = context_tokens_tensor[:, :org_context_length]
attention_mask = get_atten_mask(batch_size, org_context_length).cuda(context_tokens_tensor.device).to(model_dtype)
position_ids = torch.arange(org_context_length, dtype=torch.long,
device=tokens.device)
position_ids = position_ids.unsqueeze(0).expand_as(tokens)
counter, mem_length = 0, 0
if mems is None:
mems = []
if end_token_id is None:
end_token_id = 50000
if max_out_seq is None:
max_out_seq = 512
output_tokens_lists = []
# record order
origin_order = torch.tensor(range(batch_size), device=tokens.device)
output_order = []
# record log_probs
log_probs_tensor = torch.tensor([0.0] * batch_size, device=tokens.device)
log_probs_list = []
with torch.no_grad():
# while counter < (max_out_seq - org_context_length):
while counter < max_out_seq:
index = org_context_length + counter
if counter == 0:
output = model.forward(input_ids=tokens, position_ids=position_ids,
attention_mask=attention_mask, hidden_states=mems)
logits, mems = output.logits, output.hidden_states
else:
output = model.forward(input_ids=tokens[:, index - 1: index], position_ids=tokens.new_ones((1, 1)) * (index - 1),
attention_mask=tokens.new_ones(batch_size, 1, 1, mem_length + 1).to(model_dtype), hidden_states=mems)
logits, mems = output.logits, output.hidden_states
logits = logits[:, -1]
logits /= temperature
logits = top_k_logits(logits, top_k=top_k, top_p=top_p)
# if repetition_penalty != 1.0:
# for bz in range(batch_size):
# enforce_repetition_penalty(logits[bz, :], tokens[bz, :], repetition_penalty)
log_probs = F.softmax(logits, dim=-1) # [bs, vocab_size]
# if repetition_penalty != 1.0:
# for bz in range(batch_size):
# enforce_repetition_penalty(
# log_probs[bz, :], tokens[bz, :], repetition_penalty)
prev = torch.multinomial(log_probs, num_samples=1).view(-1)
if index < torch.max(context_length_tensor).item():
prev = switch(
prev, context_tokens_tensor[:, index], context_length_tensor <= index)
for i in range(batch_size):
if index > context_length_tensor[i] and prev[i] != end_token_id:
log_probs_tensor[i] += math.log(log_probs[i][prev[i]] + 1e-6) ###
if prev[i] == end_token_id:
log_probs_tensor[i] /= (context_length_tensor[i].cpu() - index)
# with torch.autocast('cpu'):
stop_idx = prev == end_token_id
if torch.all(stop_idx).item():
output_order.extend(origin_order[stop_idx].tolist())
break
finished = tokens[stop_idx]
output_tokens_lists.extend(finished.detach().cpu().tolist())
log_probs_list.extend(log_probs_tensor[stop_idx].tolist())
output_order.extend(origin_order[stop_idx].tolist())
# continue with non-ending tokens
conti_idx = (prev != end_token_id)
origin_order = origin_order[conti_idx]
tokens, prev = tokens[conti_idx], prev[conti_idx]
context_tokens_tensor = context_tokens_tensor[conti_idx]
context_length_tensor = context_length_tensor[conti_idx]
log_probs_tensor = log_probs_tensor[conti_idx]
batch_size = tokens.shape[0]
for im in range(len(mems)):
mems[im] = mems[im][conti_idx, :, :]
tokens = torch.cat((tokens, prev.view(batch_size, 1)), dim=-1)
counter += 1
output_tokens_lists.extend(tokens.detach().cpu().tolist())
log_probs_list.extend(log_probs_tensor.tolist())
output_order.extend(origin_order.tolist()) ###
output_tokens_lists = [tokens[:tokens.index(
end_token_id)] if end_token_id in tokens else tokens for tokens in output_tokens_lists]
output_tokens_lists = [tokens for _, tokens in sorted(zip(output_order, output_tokens_lists))]
output_log_porbs = [prob for _, prob in sorted(zip(output_order, log_probs_list))]
return output_tokens_lists, output_log_porbs
def sample_sequence(model, tokens, attention_mask, do_sampling=True,
repetition_penalty=1.0, max_out_seq=None, mems=None, end_token_id=None,
mem_length=0, temperature=1.0, top_k=0, top_p=0.0):
"""_summary_
Args:
model (_type_): _description_
tokens (Tensor): [1, seq_len]
attention_mask (Tensor): [1, 1, seq_len, seq_len]
do_sampling (bool, optional): _description_. Defaults to True.
repetition_penalty (float, optional): _description_. Defaults to 1.0.
max_out_seq (_type_, optional): _description_. Defaults to None.
mems (_type_, optional): _description_. Defaults to None.
end_token (_type_, optional): _description_. Defaults to None.
mem_length (int, optional): _description_. Defaults to 0.
temperature (float, optional): _description_. Defaults to 1.0.
top_k (int, optional): _description_. Defaults to 0.
top_p (float, optional): _description_. Defaults to 0.0.
Returns:
_type_: _description_
"""
counter = 0
if mems is None:
mems = []
if end_token_id is None:
end_token_id = 50000
if max_out_seq is None:
max_out_seq = 512
org_context_length = tokens.size(1)
with torch.no_grad():
# while counter < (max_out_seq - org_context_length):
while counter < max_out_seq:
if counter == 0:
logits, *mems = model(input_ids=tokens, position_ids=None,
attention_mask=attention_mask, mems=mems)
else:
index = org_context_length + counter
logits, *mems = model(input_ids=tokens[:, index - 1: index], position_ids=None,
attention_mask=tokens.new_ones(1, 1, 1, mem_length + 1), mems=mems)
logits = logits[:, -1]
logits /= temperature
if do_sampling:
logits = top_k_logits(logits, top_k=top_k, top_p=top_p)
log_probs = F.softmax(logits, dim=-1)
if repetition_penalty != 1.0:
enforce_repetition_penalty(
log_probs[0, :], tokens[0, :], repetition_penalty)
prev = torch.multinomial(log_probs, num_samples=1)[0]
is_end = (prev == end_token_id)
if is_end:
break
tokens = torch.cat((tokens, prev.view(1, 1)), dim=1)
counter += 1
output_tokens_list = tokens.detach().cpu().tolist()
if end_token_id in output_tokens_list:
output_tokens_list = output_tokens_list[:output_tokens_list.index(
end_token_id)]
return output_tokens_list[0], mems
|