hodu_nn 0.1.9

hodu nn
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
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

use crate::compat::*;
use crate::module::Module;
use hodu_core::{error::HoduResult, scalar::Scalar, tensor::Tensor, types::dtype::DType};

#[derive(Module, Clone)]
pub struct Embedding {
    num_embeddings: usize,
    embedding_dim: usize,
    weight: Tensor,
    padding_idx: Option<usize>,
    max_norm: Option<Scalar>,
    norm_type: Scalar,
}

impl Embedding {
    pub fn new(
        num_embeddings: usize,
        embedding_dim: usize,
        padding_idx: Option<usize>,
        max_norm: Option<impl Into<Scalar>>,
        norm_type: impl Into<Scalar>,
        dtype: DType,
    ) -> HoduResult<Self> {
        let norm_type = norm_type.into();
        let max_norm = max_norm.map(|v| v.into());

        // Xavier/Glorot initialization
        let k: f32 = 1.0 / (embedding_dim as f32).sqrt();
        let k_scalar = Scalar::from_f32(k, dtype);

        let zero = Scalar::zero(dtype);
        let one = Scalar::one(dtype);

        // weight: [num_embeddings, embedding_dim]
        let weight = Tensor::randn(&[num_embeddings, embedding_dim], zero, one)?;
        weight.set_requires_grad(true)?;
        // Scale by k
        let weight = weight.mul_scalar(k_scalar)?;

        // If padding_idx is specified, initialize that embedding to zeros
        let weight = if let Some(idx) = padding_idx {
            if idx >= num_embeddings {
                return Err(hodu_core::error::HoduError::InternalError(format!(
                    "padding_idx {} is out of range for num_embeddings {}",
                    idx, num_embeddings
                )));
            }
            // Set weight[idx] to zeros using index_put
            let indices = Tensor::new(vec![idx as i32])?.reshape([1])?;
            let zeros = Tensor::zeros(&[1, embedding_dim], dtype)?;
            weight.index_put(0, &indices, &zeros)?
        } else {
            weight
        };

        Ok(Self {
            num_embeddings,
            embedding_dim,
            weight,
            padding_idx,
            max_norm,
            norm_type,
        })
    }

    pub fn from_pretrained(weight: Tensor, freeze: bool) -> HoduResult<Self> {
        let weight_layout = weight.get_layout();
        let weight_shape = weight_layout.get_shape();

        if weight_shape.len() != 2 {
            return Err(hodu_core::error::HoduError::InternalError(format!(
                "Embedding weight must be 2D [num_embeddings, embedding_dim], got {}D",
                weight_shape.len()
            )));
        }

        let num_embeddings = weight_shape[0];
        let embedding_dim = weight_shape[1];

        weight.set_requires_grad(!freeze)?;

        Ok(Self {
            num_embeddings,
            embedding_dim,
            weight,
            padding_idx: None,
            max_norm: None,
            norm_type: Scalar::F32(2.0),
        })
    }

    pub fn forward(&self, input: &Tensor) -> HoduResult<Tensor> {
        // input: indices of shape [batch_size] or [batch_size, seq_len] or any shape
        // output: embeddings of shape [..., embedding_dim]

        let input_layout = input.get_layout();
        let input_shape = input_layout.get_shape();

        // Flatten input to 1D for easier processing
        let num_indices: usize = input_shape.iter().product();
        let flat_input = input.reshape([num_indices])?;

        // Use index_select on the weight tensor (same as gather on axis 0)
        // axis 0 is the num_embeddings dimension
        // This will fail if any index >= num_embeddings
        let flat_output = self.weight.index_select(0, &flat_input)?;

        // Apply max_norm constraint if specified
        let flat_output = if let Some(max_norm) = self.max_norm {
            self.apply_max_norm(&flat_output, max_norm)?
        } else {
            flat_output
        };

        // If padding_idx is specified, zero out gradients for padding embeddings
        let flat_output = if let Some(padding_idx) = self.padding_idx {
            self.handle_padding_idx(&flat_output, &flat_input, padding_idx)?
        } else {
            flat_output
        };

        // Output shape: [...original_input_shape..., embedding_dim]
        let mut output_shape = input_shape.to_vec();
        output_shape.push(self.embedding_dim);

        // Reshape to the desired output shape
        flat_output.reshape(&output_shape)
    }

    fn apply_max_norm(&self, embeddings: &Tensor, max_norm: Scalar) -> HoduResult<Tensor> {
        // Compute L2 norm along the embedding dimension (last dimension)
        // embeddings shape: [num_indices, embedding_dim]
        let embeddings_layout = embeddings.get_layout();
        let embedding_dim = embeddings_layout.get_shape().last().unwrap();

        // squared_sum = sum(x^2, dim=-1, keepdim=true)
        let squared = embeddings.mul(embeddings)?;
        let squared_sum = squared.sum(&[embedding_dim - 1], true)?;

        // norm = sqrt(squared_sum)
        let norm = squared_sum.sqrt()?;

        // scale = min(max_norm / norm, 1.0)
        let max_norm_tensor = Tensor::full(norm.get_layout().get_shape(), max_norm)?;
        let scale = max_norm_tensor.div(&norm)?;

        // Clamp scale to maximum of 1.0
        let one = Tensor::full(scale.get_layout().get_shape(), Scalar::one(embeddings.get_dtype()))?;
        let scale = scale.minimum(&one)?;

        // Apply scaling: embeddings * scale
        embeddings.mul(&scale)
    }

    fn handle_padding_idx(&self, embeddings: &Tensor, indices: &Tensor, padding_idx: usize) -> HoduResult<Tensor> {
        // Create a mask where padding indices are marked
        // mask = (indices == padding_idx)
        let padding_idx_scalar = Scalar::from_f32(padding_idx as f32, indices.get_dtype());
        let padding_idx_tensor = Tensor::full(indices.get_layout().get_shape(), padding_idx_scalar)?;

        // mask: [num_indices], true where index == padding_idx
        let mask = indices.eq(&padding_idx_tensor)?;

        // Expand mask to match embeddings shape: [num_indices, 1]
        let mask = mask.unsqueeze(-1)?;

        // Zero out embeddings where mask is true
        // result = embeddings * (1 - mask)
        let one = Tensor::ones(mask.get_layout().get_shape(), embeddings.get_dtype())?;
        let zero = Tensor::zeros(mask.get_layout().get_shape(), embeddings.get_dtype())?;

        // Convert bool mask to float: where(mask, 0.0, 1.0)
        let float_mask = mask.where3(&zero, &one)?;

        // Apply mask
        embeddings.mul(&float_mask)
    }

    pub fn parameters(&mut self) -> Vec<&mut Tensor> {
        vec![&mut self.weight]
    }

    pub fn weight(&self) -> &Tensor {
        &self.weight
    }

    pub fn num_embeddings(&self) -> usize {
        self.num_embeddings
    }

    pub fn embedding_dim(&self) -> usize {
        self.embedding_dim
    }

    pub fn padding_idx(&self) -> Option<usize> {
        self.padding_idx
    }

    pub fn max_norm(&self) -> Option<f32> {
        self.max_norm.as_ref().map(|s| s.to_f32())
    }

    pub fn norm_type(&self) -> f32 {
        self.norm_type.to_f32()
    }
}