haritzpuerto
commited on
Commit
•
4b2efc2
1
Parent(s):
0297073
Update inference.py
Browse files- inference.py +3 -299
inference.py
CHANGED
@@ -1,10 +1,8 @@
|
|
1 |
-
from
|
2 |
-
import
|
3 |
import numpy as np
|
4 |
|
5 |
-
|
6 |
-
from transformers.modeling_outputs import TokenClassifierOutput, BaseModelOutputWithPoolingAndCrossAttentions
|
7 |
-
from transformers.models.bert.modeling_bert import BertPooler, BertEncoder
|
8 |
|
9 |
class PredictionRequest():
|
10 |
input_question: str
|
@@ -107,297 +105,3 @@ class MetaQA():
|
|
107 |
agent_name = self.metaqa_model.config.agents[idx]
|
108 |
agent_score = input_predictions[idx][1]
|
109 |
return (pred, agent_name, metaqa_score, agent_score)
|
110 |
-
|
111 |
-
|
112 |
-
class MetaQA_Model(BertPreTrainedModel):
|
113 |
-
def __init__(self, config):
|
114 |
-
super().__init__(config)
|
115 |
-
self.bert = MetaQABertModel(config)
|
116 |
-
self.num_agents = config.num_agents
|
117 |
-
|
118 |
-
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
119 |
-
self.list_MoSeN = nn.ModuleList([nn.Linear(config.hidden_size, 1) for i in range(self.num_agents)])
|
120 |
-
self.input_size_ans_sel = 1 + config.hidden_size
|
121 |
-
interm_size = int(config.hidden_size/2)
|
122 |
-
self.ans_sel = nn.Sequential(nn.Linear(self.input_size_ans_sel, interm_size),
|
123 |
-
nn.ReLU(),
|
124 |
-
nn.Dropout(config.hidden_dropout_prob),
|
125 |
-
nn.Linear(interm_size, 2))
|
126 |
-
|
127 |
-
self.init_weights()
|
128 |
-
|
129 |
-
def forward(
|
130 |
-
self,
|
131 |
-
input_ids=None,
|
132 |
-
attention_mask=None,
|
133 |
-
token_type_ids=None,
|
134 |
-
position_ids=None,
|
135 |
-
head_mask=None,
|
136 |
-
inputs_embeds=None,
|
137 |
-
labels=None,
|
138 |
-
output_attentions=None,
|
139 |
-
output_hidden_states=None,
|
140 |
-
return_dict=None,
|
141 |
-
ans_sc=None,
|
142 |
-
agent_sc=None,
|
143 |
-
):
|
144 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
145 |
-
|
146 |
-
outputs = self.bert(
|
147 |
-
input_ids,
|
148 |
-
attention_mask=attention_mask,
|
149 |
-
token_type_ids=token_type_ids,
|
150 |
-
position_ids=position_ids,
|
151 |
-
head_mask=head_mask,
|
152 |
-
inputs_embeds=inputs_embeds,
|
153 |
-
output_attentions=output_attentions,
|
154 |
-
output_hidden_states=output_hidden_states,
|
155 |
-
return_dict=return_dict,
|
156 |
-
ans_sc=ans_sc,
|
157 |
-
agent_sc=agent_sc,
|
158 |
-
)
|
159 |
-
# domain classification
|
160 |
-
pooled_output = outputs[1]
|
161 |
-
|
162 |
-
pooled_output = self.dropout(pooled_output)
|
163 |
-
list_domains_logits = []
|
164 |
-
for MoSeN in self.list_MoSeN:
|
165 |
-
domain_logits = MoSeN(pooled_output)
|
166 |
-
list_domains_logits.append(domain_logits)
|
167 |
-
domain_logits = torch.stack(list_domains_logits)
|
168 |
-
# shape = (num_agents, batch_size, 1)
|
169 |
-
# we have to transpose the shape to (batch_size, num_agents, 1)
|
170 |
-
domain_logits = domain_logits.transpose(0,1)
|
171 |
-
|
172 |
-
# ans classifier
|
173 |
-
sequence_output = outputs[0] # (batch_size, seq_len, hidden_size)
|
174 |
-
# select the [RANK] token embeddings
|
175 |
-
idx_rank = (input_ids == 1).nonzero() # (batch_size x num_agents, 2)
|
176 |
-
idx_rank = idx_rank[:,1].view(-1, self.num_agents)
|
177 |
-
list_emb = []
|
178 |
-
for i in range(idx_rank.shape[0]):
|
179 |
-
rank_emb = sequence_output[i][idx_rank[i], :]
|
180 |
-
# rank shape = (1, hidden_size)
|
181 |
-
list_emb.append(rank_emb)
|
182 |
-
|
183 |
-
rank_emb = torch.stack(list_emb)
|
184 |
-
|
185 |
-
rank_emb = self.dropout(rank_emb)
|
186 |
-
rank_emb = torch.cat((rank_emb, domain_logits), dim=2)
|
187 |
-
# rank emb shape = (batch_size, num_agents, hidden_size+1)
|
188 |
-
logits = self.ans_sel(rank_emb) # (batch_size, num_agents, 2)
|
189 |
-
|
190 |
-
if not return_dict:
|
191 |
-
output = (logits,) + outputs[2:]
|
192 |
-
return output
|
193 |
-
|
194 |
-
return TokenClassifierOutput(
|
195 |
-
loss=None,
|
196 |
-
logits=logits,
|
197 |
-
hidden_states=outputs.hidden_states,
|
198 |
-
attentions=outputs.attentions,
|
199 |
-
)
|
200 |
-
|
201 |
-
|
202 |
-
class MetaQABertModel(BertPreTrainedModel):
|
203 |
-
def __init__(self, config, add_pooling_layer=True):
|
204 |
-
super().__init__(config)
|
205 |
-
self.config = config
|
206 |
-
|
207 |
-
self.embeddings = MetaQABertEmbeddings(config)
|
208 |
-
self.encoder = BertEncoder(config)
|
209 |
-
self.pooler = BertPooler(config) if add_pooling_layer else None
|
210 |
-
|
211 |
-
self.init_weights()
|
212 |
-
|
213 |
-
def get_input_embeddings(self):
|
214 |
-
return self.embeddings.word_embeddings
|
215 |
-
|
216 |
-
def set_input_embeddings(self, value):
|
217 |
-
self.embeddings.word_embeddings = value
|
218 |
-
|
219 |
-
def _prune_heads(self, heads_to_prune):
|
220 |
-
for layer, heads in heads_to_prune.items():
|
221 |
-
self.encoder.layer[layer].attention.prune_heads(heads)
|
222 |
-
|
223 |
-
def forward(
|
224 |
-
self,
|
225 |
-
input_ids=None,
|
226 |
-
attention_mask=None,
|
227 |
-
token_type_ids=None,
|
228 |
-
position_ids=None,
|
229 |
-
head_mask=None,
|
230 |
-
inputs_embeds=None,
|
231 |
-
encoder_hidden_states=None,
|
232 |
-
encoder_attention_mask=None,
|
233 |
-
past_key_values=None,
|
234 |
-
use_cache=None,
|
235 |
-
output_attentions=None,
|
236 |
-
output_hidden_states=None,
|
237 |
-
return_dict=None,
|
238 |
-
ans_sc=None,
|
239 |
-
agent_sc=None,
|
240 |
-
):
|
241 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
242 |
-
output_hidden_states = (
|
243 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
244 |
-
)
|
245 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
246 |
-
|
247 |
-
if self.config.is_decoder:
|
248 |
-
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
249 |
-
else:
|
250 |
-
use_cache = False
|
251 |
-
|
252 |
-
if input_ids is not None and inputs_embeds is not None:
|
253 |
-
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
254 |
-
elif input_ids is not None:
|
255 |
-
input_shape = input_ids.size()
|
256 |
-
batch_size, seq_length = input_shape
|
257 |
-
elif inputs_embeds is not None:
|
258 |
-
input_shape = inputs_embeds.size()[:-1]
|
259 |
-
batch_size, seq_length = input_shape
|
260 |
-
else:
|
261 |
-
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
262 |
-
|
263 |
-
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
264 |
-
|
265 |
-
# past_key_values_length
|
266 |
-
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
|
267 |
-
|
268 |
-
if attention_mask is None:
|
269 |
-
attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
|
270 |
-
|
271 |
-
if token_type_ids is None:
|
272 |
-
if hasattr(self.embeddings, "token_type_ids"):
|
273 |
-
buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
|
274 |
-
buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
|
275 |
-
token_type_ids = buffered_token_type_ids_expanded
|
276 |
-
else:
|
277 |
-
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
|
278 |
-
|
279 |
-
# We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
|
280 |
-
# ourselves in which case we just need to make it broadcastable to all heads.
|
281 |
-
extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
|
282 |
-
|
283 |
-
# If a 2D or 3D attention mask is provided for the cross-attention
|
284 |
-
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
|
285 |
-
if self.config.is_decoder and encoder_hidden_states is not None:
|
286 |
-
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
|
287 |
-
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
|
288 |
-
if encoder_attention_mask is None:
|
289 |
-
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
|
290 |
-
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
|
291 |
-
else:
|
292 |
-
encoder_extended_attention_mask = None
|
293 |
-
|
294 |
-
# Prepare head mask if needed
|
295 |
-
# 1.0 in head_mask indicate we keep the head
|
296 |
-
# attention_probs has shape bsz x n_heads x N x N
|
297 |
-
# input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
|
298 |
-
# and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
|
299 |
-
head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
|
300 |
-
|
301 |
-
embedding_output = self.embeddings(
|
302 |
-
input_ids=input_ids,
|
303 |
-
position_ids=position_ids,
|
304 |
-
token_type_ids=token_type_ids,
|
305 |
-
inputs_embeds=inputs_embeds,
|
306 |
-
past_key_values_length=past_key_values_length,
|
307 |
-
ans_sc=ans_sc,
|
308 |
-
agent_sc=agent_sc,
|
309 |
-
)
|
310 |
-
encoder_outputs = self.encoder(
|
311 |
-
embedding_output,
|
312 |
-
attention_mask=extended_attention_mask,
|
313 |
-
head_mask=head_mask,
|
314 |
-
encoder_hidden_states=encoder_hidden_states,
|
315 |
-
encoder_attention_mask=encoder_extended_attention_mask,
|
316 |
-
past_key_values=past_key_values,
|
317 |
-
use_cache=use_cache,
|
318 |
-
output_attentions=output_attentions,
|
319 |
-
output_hidden_states=output_hidden_states,
|
320 |
-
return_dict=return_dict,
|
321 |
-
)
|
322 |
-
sequence_output = encoder_outputs[0]
|
323 |
-
pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
|
324 |
-
|
325 |
-
if not return_dict:
|
326 |
-
return (sequence_output, pooled_output) + encoder_outputs[1:]
|
327 |
-
|
328 |
-
return BaseModelOutputWithPoolingAndCrossAttentions(
|
329 |
-
last_hidden_state=sequence_output,
|
330 |
-
pooler_output=pooled_output,
|
331 |
-
past_key_values=encoder_outputs.past_key_values,
|
332 |
-
hidden_states=encoder_outputs.hidden_states,
|
333 |
-
attentions=encoder_outputs.attentions,
|
334 |
-
cross_attentions=encoder_outputs.cross_attentions,
|
335 |
-
)
|
336 |
-
|
337 |
-
class MetaQABertEmbeddings(nn.Module):
|
338 |
-
"""Construct the embeddings from word, position and token_type embeddings."""
|
339 |
-
|
340 |
-
def __init__(self, config):
|
341 |
-
super().__init__()
|
342 |
-
self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
|
343 |
-
self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
|
344 |
-
self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
|
345 |
-
self.ans_sc_proj = nn.Linear(1, config.hidden_size)
|
346 |
-
self.agent_sc_proj = nn.Linear(1, config.hidden_size)
|
347 |
-
# self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
|
348 |
-
# any TensorFlow checkpoint file
|
349 |
-
self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
350 |
-
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
351 |
-
# position_ids (1, len position emb) is contiguous in memory and exported when serialized
|
352 |
-
self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
|
353 |
-
self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
|
354 |
-
self.register_buffer(
|
355 |
-
"token_type_ids",
|
356 |
-
torch.zeros(self.position_ids.size(), dtype=torch.long, device=self.position_ids.device),
|
357 |
-
persistent=False,
|
358 |
-
)
|
359 |
-
|
360 |
-
|
361 |
-
def forward(
|
362 |
-
self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0,
|
363 |
-
ans_sc=None, agent_sc=None):
|
364 |
-
if input_ids is not None:
|
365 |
-
input_shape = input_ids.size()
|
366 |
-
else:
|
367 |
-
input_shape = inputs_embeds.size()[:-1]
|
368 |
-
|
369 |
-
seq_length = input_shape[1]
|
370 |
-
|
371 |
-
if position_ids is None:
|
372 |
-
position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
|
373 |
-
|
374 |
-
# Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
|
375 |
-
# when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
|
376 |
-
# issue #5664
|
377 |
-
if token_type_ids is None:
|
378 |
-
if hasattr(self, "token_type_ids"):
|
379 |
-
buffered_token_type_ids = self.token_type_ids[:, :seq_length]
|
380 |
-
buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
|
381 |
-
token_type_ids = buffered_token_type_ids_expanded
|
382 |
-
else:
|
383 |
-
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
|
384 |
-
|
385 |
-
if inputs_embeds is None:
|
386 |
-
inputs_embeds = self.word_embeddings(input_ids)
|
387 |
-
token_type_embeddings = self.token_type_embeddings(token_type_ids)
|
388 |
-
|
389 |
-
embeddings = inputs_embeds + token_type_embeddings
|
390 |
-
if self.position_embedding_type == "absolute":
|
391 |
-
position_embeddings = self.position_embeddings(position_ids)
|
392 |
-
embeddings += position_embeddings
|
393 |
-
|
394 |
-
if ans_sc is not None:
|
395 |
-
ans_sc_emb = self.ans_sc_proj(ans_sc.unsqueeze(2))
|
396 |
-
embeddings += ans_sc_emb
|
397 |
-
if agent_sc is not None:
|
398 |
-
agent_sc_emb = self.agent_sc_proj(agent_sc.unsqueeze(2))
|
399 |
-
embeddings += agent_sc_emb
|
400 |
-
|
401 |
-
embeddings = self.LayerNorm(embeddings)
|
402 |
-
embeddings = self.dropout(embeddings)
|
403 |
-
return embeddings
|
|
|
1 |
+
from transformers import AutoTokenizer
|
2 |
+
from MetaQA_Model import MetaQA_Model
|
3 |
import numpy as np
|
4 |
|
5 |
+
import torch
|
|
|
|
|
6 |
|
7 |
class PredictionRequest():
|
8 |
input_question: str
|
|
|
105 |
agent_name = self.metaqa_model.config.agents[idx]
|
106 |
agent_score = input_predictions[idx][1]
|
107 |
return (pred, agent_name, metaqa_score, agent_score)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|