Update README.md
Browse files
README.md
CHANGED
@@ -38,6 +38,8 @@ Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanow
|
|
38 |
|
39 |
## How to use the model
|
40 |
|
|
|
|
|
41 |
The model can be loaded with the `zero-shot-classification` pipeline like so:
|
42 |
|
43 |
```python
|
@@ -61,6 +63,32 @@ candidate_labels = ["politics", "economy", "entertainment", "environment"]
|
|
61 |
classifier(sequence_to_classify, candidate_labels, multi_label=True)
|
62 |
```
|
63 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
### Eval results
|
65 |
The model was evaluated using the XNLI test sets on 14 languages: English (en), Arabic (ar), Bulgarian (bg), German (de), Greek (el), Spanish (es), French (fr), Russian (ru), Swahili (sw), Thai (th), Turkish (tr), Urdu (ur), Vietnam (vi) and Chinese (zh). The metric used is accuracy.
|
66 |
|
|
|
38 |
|
39 |
## How to use the model
|
40 |
|
41 |
+
### With the zero-shot classification pipeline
|
42 |
+
|
43 |
The model can be loaded with the `zero-shot-classification` pipeline like so:
|
44 |
|
45 |
```python
|
|
|
63 |
classifier(sequence_to_classify, candidate_labels, multi_label=True)
|
64 |
```
|
65 |
|
66 |
+
### With manual PyTorch
|
67 |
+
|
68 |
+
The model can also be applied on NLI tasks like so:
|
69 |
+
|
70 |
+
```python
|
71 |
+
import torch
|
72 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
73 |
+
|
74 |
+
# device = "cuda:0" or "cpu"
|
75 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
76 |
+
|
77 |
+
model_name = "mjwong/mcontriever-msmarco-xnli"
|
78 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
79 |
+
model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
80 |
+
|
81 |
+
premise = "But I thought you'd sworn off coffee."
|
82 |
+
hypothesis = "I thought that you vowed to drink more coffee."
|
83 |
+
|
84 |
+
input = tokenizer(premise, hypothesis, truncation=True, return_tensors="pt")
|
85 |
+
output = model(input["input_ids"].to(device))
|
86 |
+
prediction = torch.softmax(output["logits"][0], -1).tolist()
|
87 |
+
label_names = ["entailment", "neutral", "contradiction"]
|
88 |
+
prediction = {name: round(float(pred) * 100, 2) for pred, name in zip(prediction, label_names)}
|
89 |
+
print(prediction)
|
90 |
+
```
|
91 |
+
|
92 |
### Eval results
|
93 |
The model was evaluated using the XNLI test sets on 14 languages: English (en), Arabic (ar), Bulgarian (bg), German (de), Greek (el), Spanish (es), French (fr), Russian (ru), Swahili (sw), Thai (th), Turkish (tr), Urdu (ur), Vietnam (vi) and Chinese (zh). The metric used is accuracy.
|
94 |
|