Hi, I’m using bart-large-mnli without pipeline, and so using manual pytorch, as explained in the example at the bottom of the HF page.
device = 'cpu'
premise = "Last week I upgraded my iOS version and ever since then my phone has been overheating whenever I use your app."
candidate_labels = ['mobile', 'website', 'billing', 'account access']
entail_contradiction_logits_matrix = []
entail_logits_matrix = []
output = []
multi_class = True
for label in candidate_labels:
hypothesis = f'This example is {label}.'
# run through model pre-trained on MNLI
x = tokenizer.encode_plus(premise, hypothesis, return_tensors='pt',
truncation_strategy='only_first')['input_ids']
logits = nli_model(x)[0]
# we throw away "neutral" (dim 1) and take the probability of
# "entailment" (2) as the probability of the label being true
if multi_class:
entail_contradiction_logits_matrix.append(logits[:,[0,2]])
probs = logits[:,[0,2]].softmax(dim=1)
prob_label_is_true = probs[:,1]
output.append(prob_label_is_true.item())
else:
entail_logits_matrix.append(logits[:,[2]])
if multi_class== False:
entails_tensor = torch.cat(entail_logits_matrix, dim=1)
probs = entails_tensor.softmax(dim=1)
output = probs.detach().to('cpu').numpy()
print("premise:", premise)
print("candidate_labels:", candidate_labels)
print(output)
So, everything works fine
##OUTPUT
premise: Last week I upgraded my iOS version and ever since then my phone has been overheating whenever I use your app.
candidate_labels: ['mobile', 'website', 'billing', 'account access']
[0.9908875226974487, 0.0016617656219750643, 0.37062010169029236, 0.445667564868927]
Now I want to create a dataloader and classify multiple examples at a time. (I’ve replaced unimportant code with ‘…’)
def generate_data_loader(self, examples):
'''
Generate a Dataloader given the input examples
examples: a list of tuples (text, type_text, lang_text, doc_id, parag_id)
'''
encoded_text_array = []
label_type_text_array = []
label_lang_text_array = []
max_seq_length = 256 # the maximum length to be considered in input
batch_size = 64 # the batch size
# Tokenization
for (text, type_text, lang_text, doc_id, parag_id) in examples:
premise = text
hypothesis_list = self.generate_hypothesis(...)
for hypothesis in hypothesis_list:
encoded_text = self.tokenizer.encode_plus(premise, hypothesis, add_special_tokens=True, truncation_strategy='only_first', return_tensors='pt', max_length=max_seq_length, padding='max_length')['input_ids']
encoded_text_array.append(encoded_text)
[...]
encoded_text_array = torch.stack(encoded_text_array)
label_lang_text_array = torch.tensor(..., dtype=torch.long)
label_type_text_array = torch.tensor(..., dtype=torch.long)
# Building the TensorDataset
dataset = TensorDataset(encoded_text_array, label_lang_text_array,
label_type_text_array, ...)
sampler = SequentialSampler
# Building the DataLoader
return DataLoader(
dataset, # The training samples.
sampler = sampler(dataset), # the adopted sampler
batch_size = batch_size) # Trains with this batch size.
and finally
for batch in dataloader:
b_encoded_text = batch[0].to(self.device)
[...]
b_encoded_text = torch.unbind(b_encoded_text)
with torch.no_grad():
self.model.eval()
logits = self.model(b_encoded_text)[0]
But I get an error.
<ipython-input-36-44a7c2660e53> in model_predictor(self, dataloader)
208 # Each batch is classifed
209 #logits_subjobj, logits_priv, logits_pos, logits_neg, _ = self.model(b_input_ids, b_input_mask)
--> 210 logits = self.model(b_encoded_text)[0]
211 if multi_class:
212 print("logits", logits)
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []
/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py in forward(self, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, head_mask, decoder_head_mask, cross_attn_head_mask, encoder_outputs, inputs_embeds, decoder_inputs_embeds, labels, use_cache, output_attentions, output_hidden_states, return_dict)
1482 output_attentions=output_attentions,
1483 output_hidden_states=output_hidden_states,
-> 1484 return_dict=return_dict,
1485 )
1486 hidden_states = outputs[0] # last hidden state
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1100 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1101 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102 return forward_call(*input, **kwargs)
1103 # Do not call functions when jit is used
1104 full_backward_hooks, non_full_backward_hooks = [], []
/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py in forward(self, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask, head_mask, decoder_head_mask, cross_attn_head_mask, encoder_outputs, past_key_values, inputs_embeds, decoder_inputs_embeds, use_cache, output_attentions, output_hidden_states, return_dict)
1185
1186 decoder_input_ids = shift_tokens_right(
-> 1187 input_ids, self.config.pad_token_id, self.config.decoder_start_token_id
1188 )
1189
/usr/local/lib/python3.7/dist-packages/transformers/models/bart/modeling_bart.py in shift_tokens_right(input_ids, pad_token_id, decoder_start_token_id)
64 Shift input ids one token to the right.
65 """
---> 66 shifted_input_ids = input_ids.new_zeros(input_ids.shape)
67 shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
68 shifted_input_ids[:, 0] = decoder_start_token_id
AttributeError: 'tuple' object has no attribute 'new_zeros'
Apart from the specific error, I’m sure I haven’t understood how to pass the data with the dataloader to the model. How is this done? Thank you in advance
