rust-bert 0.7.2

Ready-to-use NLP pipelines and transformer-based models (BERT, DistilBERT, GPT2,...)
Documentation 
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

use rust_bert::resources::{Resource, download_resource, RemoteResource};
use rust_bert::electra::{ElectraConfigResources, ElectraVocabResources, ElectraModelResources, ElectraConfig, ElectraForMaskedLM, ElectraDiscriminator};
use tch::{Device, nn, Tensor, no_grad};
use rust_tokenizers::{BertTokenizer, TruncationStrategy, Tokenizer, Vocab};
use rust_bert::Config;

#[test]
fn electra_masked_lm() -> failure::Fallible<()> {
    //    Resources paths
    let config_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraConfigResources::BASE_GENERATOR));
    let vocab_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraVocabResources::BASE_GENERATOR));
    let weights_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraModelResources::BASE_GENERATOR));
    let config_path = download_resource(&config_resource)?;
    let vocab_path = download_resource(&vocab_resource)?;
    let weights_path = download_resource(&weights_resource)?;

//    Set-up masked LM model
    let device = Device::Cpu;
    let mut vs = nn::VarStore::new(device);
    let tokenizer: BertTokenizer = BertTokenizer::from_file(vocab_path.to_str().unwrap(), true);
    let mut config = ElectraConfig::from_file(config_path);
    config.output_attentions = Some(true);
    config.output_hidden_states = Some(true);
    let electra_model = ElectraForMaskedLM::new(&vs.root(), &config);
    vs.load(weights_path)?;

//    Define input
    let input = ["Looks like one [MASK] is missing", "It was a very nice and [MASK] day"];
    let tokenized_input = tokenizer.encode_list(input.to_vec(), 128, &TruncationStrategy::LongestFirst, 0);
    let max_len = tokenized_input.iter().map(|input| input.token_ids.len()).max().unwrap();
    let tokenized_input = tokenized_input.
        iter().
        map(|input| input.token_ids.clone()).
        map(|mut input| {
            input.extend(vec![0; max_len - input.len()]);
            input
        }).
        map(|input|
            Tensor::of_slice(&(input))).
        collect::<Vec<_>>();
    let input_tensor = Tensor::stack(tokenized_input.as_slice(), 0).to(device);

//    Forward pass
    let (output,
        all_hidden_states,
        all_attentions) = no_grad(|| {
        electra_model
            .forward_t(Some(input_tensor),
                       None,
                       None,
                       None,
                       None,
                       false)
    });

//    Decode output
    let index_1 = output.get(0).get(4).argmax(0, false);
    let index_2 = output.get(1).get(7).argmax(0, false);
    let word_1 = tokenizer.vocab().id_to_token(&index_1.int64_value(&[]));
    let word_2 = tokenizer.vocab().id_to_token(&index_2.int64_value(&[]));

    assert_eq!(output.size(), &[2, 10, config.vocab_size]);
    assert_eq!(config.num_hidden_layers as usize, all_hidden_states.unwrap().len());
    assert_eq!(config.num_hidden_layers as usize, all_attentions.unwrap().len());
    assert_eq!("thing", word_1); // Outputs "person" : "Looks like one [person] is missing"
    assert_eq!("sunny", word_2);// Outputs "pear" : "It was a very nice and [sunny] day"
    Ok(())
}

#[test]
fn electra_discriminator() -> failure::Fallible<()> {
    //    Resources paths
    let config_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraConfigResources::BASE_DISCRIMINATOR));
    let vocab_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraVocabResources::BASE_DISCRIMINATOR));
    let weights_resource = Resource::Remote(RemoteResource::from_pretrained(ElectraModelResources::BASE_DISCRIMINATOR));
    let config_path = download_resource(&config_resource)?;
    let vocab_path = download_resource(&vocab_resource)?;
    let weights_path = download_resource(&weights_resource)?;

//    Set-up masked LM model
    let device = Device::Cpu;
    let mut vs = nn::VarStore::new(device);
    let tokenizer: BertTokenizer = BertTokenizer::from_file(vocab_path.to_str().unwrap(), true);
    let config = ElectraConfig::from_file(config_path);
    let electra_model = ElectraDiscriminator::new(&vs.root(), &config);
    vs.load(weights_path)?;

//    Define input
    let input = ["One Two Three Ten Five Six Seven Eight"];
    let tokenized_input = tokenizer.encode_list(input.to_vec(), 128, &TruncationStrategy::LongestFirst, 0);
    let max_len = tokenized_input.iter().map(|input| input.token_ids.len()).max().unwrap();
    let encoded_input = tokenized_input.
        iter().
        map(|input| input.token_ids.clone()).
        map(|mut input| {
            input.extend(vec![0; max_len - input.len()]);
            input
        }).
        map(|input|
            Tensor::of_slice(&(input))).
        collect::<Vec<_>>();
    let input_tensor = Tensor::stack(encoded_input.as_slice(), 0).to(device);

    //    Forward pass
    let (output, _, _) = no_grad(|| {
        electra_model
            .forward_t(Some(input_tensor),
                       None,
                       None,
                       None,
                       None,
                       false)
    });

//    Validate model predictions
    let expected_probabilities = vec!(0.0101, 0.0030, 0.0010, 0.0018, 0.9489, 0.0067, 0.0026, 0.0017, 0.0311, 0.0101);
    let probabilities = output.iter::<f64>().unwrap().collect::<Vec<f64>>();

    assert_eq!(output.size(), &[10]);
    for (expected, pred) in probabilities.iter().zip(expected_probabilities) {
        assert!((expected - pred).abs() < 1e-4);
    }

    Ok(())
}