ferrite-dl 0.2.0

Deep learning library written in pure Rust
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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320

use std::rc::Rc;

use crate::{match_storage, match_storage_assign, DeviceStorage, PermuteGrad, Storage, Tensor};


pub trait TransformOps {
  fn apply<F>(&self, op: F) -> Self
  where
    F: Fn(f32) -> f32;
  
  fn apply_assign<F>(&mut self, op: F)
  where
    F: Fn(f32) -> f32;

  fn elementwise_op<F>(&self, other: &Self, op: F) -> Self
  where
  F: Fn(f32, f32) -> f32;

  fn scalar_op<F>(&self, scalar: f32, op: F) -> Self
  where
  F: Fn(f32, f32) -> f32;

  fn elementwise_op_assign<F>(&mut self, other: &Self, op: F)
  where
  F: Fn(f32, f32) -> f32;

  fn scalar_op_assign<F>(&mut self, scalar: f32, op: F)
  where
  F: Fn(f32, f32) -> f32;

  fn sum_dim(&self, dims: &[bool]) -> Self;
  fn reshape(&mut self, new_shape: Vec<usize>);
  fn permute(&mut self, dims: &[usize]);
  fn transpose(&self) -> Self;
  fn flatten(&mut self);
  fn squeeze(&mut self);
  fn unsqueeze(&mut self, dim: usize);

  fn broadcast(&self, new_shape: &[usize]) -> Self;
  fn compute_broadcast_shape(&self, target_shape: &[usize]) -> Vec<usize>;
  fn compute_broadcast_strides(&self, broadcast_shape: &[usize]) -> Vec<usize>;
  fn pad_shape(&self, target_rank: usize) -> Vec<usize>;
  fn broadcast_tensors(a: &Self, b: &Self) -> (Self, Self) where Self: Sized;

}

macro_rules! match_storage {
  // Binary operation with two storage arguments
  (binary $self:expr, $method:ident, $other:expr $(, $args:expr)*) => {
    match ($self, $other) {
      (Storage::Cpu(cpu_self), Storage::Cpu(cpu_other)) => {
        Storage::Cpu(cpu_self.$method(cpu_other $(, $args)*))
      }
      _ => unimplemented!("Cross-device operations not supported"),
    }
  };

  // Unary operation with single storage argument
  (unary $self:expr, $method:ident $(, $args:expr)*) => {
    match $self {
      Storage::Cpu(cpu) => Storage::Cpu(cpu.$method($($args),*)),
      _ => unimplemented!("Device not supported"),
    }
  };
}


impl TransformOps for Storage {
  fn apply<F>(&self, op: F) -> Self
  where
    F: Fn(f32) -> f32 {
    match_storage!(unary self, apply, op)
  }

  fn apply_assign<F>(&mut self, op: F)
  where
    F: Fn(f32) -> f32 {
    match_storage_assign!(unary self, apply_assign, op)
  }

  fn elementwise_op<F>(&self, other: &Self, op: F) -> Self
  where
  F: Fn(f32, f32) -> f32 {
    match_storage!(binary self, elementwise_op, other, op)
  }

  fn scalar_op<F>(&self, scalar: f32, op: F) -> Self
  where
  F: Fn(f32, f32) -> f32 {
    match_storage!(unary self, scalar_op, scalar, op)
  }

  fn elementwise_op_assign<F>(&mut self, other: &Self, op: F)
  where
  F: Fn(f32, f32) -> f32 {
    match_storage_assign!(binary self, elementwise_op_assign, other, op)
  }

  fn scalar_op_assign<F>(&mut self, scalar: f32, op: F)
  where
  F: Fn(f32, f32) -> f32 {
    match_storage_assign!(unary self, scalar_op_assign, scalar, op)
  }

  fn sum_dim(&self, dims: &[bool]) -> Self {
    match_storage!(unary self, sum_dim, dims)
  }

  fn reshape(&mut self, new_shape: Vec<usize>) {
    match_storage_assign!(unary self, reshape, new_shape)
  }

  fn permute(&mut self, dims: &[usize]) {
    match_storage_assign!(unary self, permute, dims)
  }

  fn transpose(&self) -> Self {
    match_storage!(unary self, transpose)
  }

  fn flatten(&mut self) {
    match_storage_assign!(unary self, flatten)
  }

  fn squeeze(&mut self) {
    match_storage_assign!(unary self, squeeze)
  }

  fn unsqueeze(&mut self, dim: usize) {
    match_storage_assign!(unary self, unsqueeze, dim)
  }

  fn broadcast(&self, new_shape: &[usize]) -> Self {
    match_storage!(unary self, broadcast, new_shape)
  }

  fn compute_broadcast_shape(&self, target_shape: &[usize]) -> Vec<usize> {
    let self_rank = self.shape().len();
    let target_rank = target_shape.len();
    let max_rank = std::cmp::max(self_rank, target_rank);
    
    // Pad shapes with 1s to match maximum rank
    let self_padded = self.pad_shape(max_rank);
    let mut result_shape = Vec::with_capacity(max_rank);

    // Compare dimensions from right to left
    for i in 0..max_rank {
      let self_dim = self_padded[i];
      let target_dim = if i >= max_rank - target_rank {
          target_shape[i - (max_rank - target_rank)]
        } else {
          1
        };

      if self_dim == target_dim {
        result_shape.push(self_dim);
      } else if self_dim == 1 {
        result_shape.push(target_dim);
      } else if target_dim == 1 {
        result_shape.push(self_dim);
      } else {
        panic!(
          "Incompatible broadcast dimensions: {} and {}",
          self_dim, target_dim
        )
      }
    }

    result_shape
  }

  fn compute_broadcast_strides(&self, broadcast_shape: &[usize]) -> Vec<usize> {
    let self_rank = self.shape().len();
    let broadcast_rank = broadcast_shape.len();
    let rank_diff = broadcast_rank - self_rank;
    
    let mut new_strides = vec![0; broadcast_rank];
    
    // Fill in strides for dimensions that match or are broadcasted
    for i in 0..self_rank {
      let broadcast_idx = i + rank_diff;
      if broadcast_shape[broadcast_idx] == self.shape()[i] {
        new_strides[broadcast_idx] = self.stride()[i];
      } else if self.shape()[i] == 1 {
        new_strides[broadcast_idx] = 0;  // Stride of 0 for broadcasted dimensions
      } else {
        panic!("Invalid broadcast shape")
      }
    }

    new_strides
  }

  fn pad_shape(&self, target_rank: usize) -> Vec<usize> {
    let mut padded = vec![1; target_rank];
    let rank_diff = target_rank - self.shape().len();
    padded[rank_diff..].copy_from_slice(self.shape());
    padded
  }

  fn broadcast_tensors(a: &Self, b: &Self) -> (Self, Self) where Self: Sized {
    // Use a's compute_broadcast_shape to get the final shape
    let broadcast_shape = a.compute_broadcast_shape(b.shape());

    // Broadcast both tensors to the new shape
    let broadcast_a = a.broadcast(&broadcast_shape);
    let broadcast_b = b.broadcast(&broadcast_shape);
    
    (broadcast_a, broadcast_b)
  }
}


impl TransformOps for Tensor {
  fn apply<F>(&self, op: F) -> Self
  where
    F: Fn(f32) -> f32 {
    todo!()
  }

  fn apply_assign<F>(&mut self, op: F)
  where
    F: Fn(f32) -> f32 {
    todo!()
  }

  fn elementwise_op<F>(&self, other: &Self, op: F) -> Self
  where
  F: Fn(f32, f32) -> f32 {
    todo!()
  }

  fn scalar_op<F>(&self, scalar: f32, op: F) -> Self
  where
    F: Fn(f32, f32) -> f32 {
    todo!()
  }

  fn elementwise_op_assign<F>(&mut self, other: &Self, op: F)
  where
    F: Fn(f32, f32) -> f32 {
    todo!()
  }

  fn scalar_op_assign<F>(&mut self, scalar: f32, op: F)
  where
    F: Fn(f32, f32) -> f32 {
    todo!()
  }

  fn sum_dim(&self, dims: &[bool]) -> Self {
    todo!()
  }

  fn reshape(&mut self, new_shape: Vec<usize>) {
    self.tensor_mut().set_shape(new_shape);
  }

  fn transpose(&self) -> Self {
    // Transpose by swapping dimensions & strides

    let tensor = self.tensor().transpose();
    let requires_grad = *self.requires_grad();
    let mut result = Tensor::new(tensor, self.device(), requires_grad);
    
    if requires_grad {
      result.set_grad_fn(Some(Rc::new(PermuteGrad::new(self, &result))));
    }
    
    result
  }

  fn permute(&mut self, dims: &[usize]) {
    self.tensor_mut().permute(dims);
  }

  fn flatten(&mut self) {
    self.tensor_mut().flatten();
  }

  fn squeeze(&mut self) {
    self.tensor_mut().squeeze();
  } 

  fn unsqueeze(&mut self, dim: usize) {
    self.tensor_mut().unsqueeze(dim);
  }

  fn broadcast(&self, new_shape: &[usize]) -> Self {
    let tensor = self.tensor().broadcast(new_shape);
    
    // When broadcasting, we need to maintain the gradient tracking
    let requires_grad = *self.requires_grad();
    let mut result = Tensor::new(tensor, self.device(), requires_grad);
    
    // If original tensor requires gradient, the broadcasted tensor
    // should have the same gradient function
    if requires_grad {
      result.set_grad_fn(self.grad_fn());
    }
    
    result
  }

  fn compute_broadcast_shape(&self, target_shape: &[usize]) -> Vec<usize> {
    self.tensor().compute_broadcast_shape(target_shape)
  }

  fn compute_broadcast_strides(&self, broadcast_shape: &[usize]) -> Vec<usize> {
    self.tensor().compute_broadcast_strides(broadcast_shape)
  }

  fn pad_shape(&self, target_rank: usize) -> Vec<usize> {
    self.tensor().pad_shape(target_rank)
  }

  fn broadcast_tensors(a: &Self, b: &Self) -> (Self, Self) where Self: Sized {
    todo!()
  }
}