1 | /* |
---|
2 | Copyright (C) 2018 Paul Brossier <piem@aubio.org> |
---|
3 | |
---|
4 | This file is part of aubio. |
---|
5 | |
---|
6 | aubio is free software: you can redistribute it and/or modify |
---|
7 | it under the terms of the GNU General Public License as published by |
---|
8 | the Free Software Foundation, either version 3 of the License, or |
---|
9 | (at your option) any later version. |
---|
10 | |
---|
11 | aubio is distributed in the hope that it will be useful, |
---|
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
---|
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
---|
14 | GNU General Public License for more details. |
---|
15 | |
---|
16 | You should have received a copy of the GNU General Public License |
---|
17 | along with aubio. If not, see <http://www.gnu.org/licenses/>. |
---|
18 | |
---|
19 | */ |
---|
20 | |
---|
21 | |
---|
22 | #include "aubio_priv.h" |
---|
23 | #include "fmat.h" |
---|
24 | #include "tensor.h" |
---|
25 | #include "conv1d.h" |
---|
26 | |
---|
27 | typedef enum |
---|
28 | { |
---|
29 | PAD_SAME = 0, |
---|
30 | PAD_VALID = 1, |
---|
31 | PAD_CAUSAL = 2, // TODO (1d only, for dilated convolution) |
---|
32 | } aubio_conv1d_padding_type; |
---|
33 | |
---|
34 | struct _aubio_conv1d_t { |
---|
35 | // define internals here |
---|
36 | uint_t n_filters; |
---|
37 | uint_t kernel_shape; // kernel sizes |
---|
38 | uint_t stride_shape; // stride sizes |
---|
39 | |
---|
40 | aubio_conv1d_padding_type padding_mode; |
---|
41 | |
---|
42 | // these will be set after calling get_output_shape |
---|
43 | aubio_tensor_t *kernel; |
---|
44 | fvec_t *bias; |
---|
45 | uint_t output_shape[2]; // shape of output |
---|
46 | uint_t padding_start; // left padding |
---|
47 | |
---|
48 | #if defined(HAVE_BLAS) |
---|
49 | aubio_tensor_t *padded_input; |
---|
50 | #endif |
---|
51 | }; |
---|
52 | |
---|
53 | static void aubio_conv1d_debug(aubio_conv1d_t *c, aubio_tensor_t *input_tensor); |
---|
54 | |
---|
55 | aubio_conv1d_t *new_aubio_conv1d(uint_t n_filters, uint_t kernel_shape[1]) |
---|
56 | { |
---|
57 | aubio_conv1d_t *c = AUBIO_NEW(aubio_conv1d_t); |
---|
58 | |
---|
59 | // validate input parameters |
---|
60 | AUBIO_GOTO_FAILURE((sint_t)n_filters >= 1); |
---|
61 | AUBIO_GOTO_FAILURE((sint_t)kernel_shape[0] >= 1); |
---|
62 | |
---|
63 | // set internal variables |
---|
64 | c->n_filters = n_filters; |
---|
65 | c->kernel_shape = kernel_shape[0]; |
---|
66 | |
---|
67 | // default to padding_mode="valid" |
---|
68 | c->padding_mode = PAD_VALID; |
---|
69 | // set default stride_shape to (1) |
---|
70 | uint_t stride_shape[1] = {1}; |
---|
71 | aubio_conv1d_set_stride(c, stride_shape); |
---|
72 | |
---|
73 | return c; |
---|
74 | |
---|
75 | failure: |
---|
76 | del_aubio_conv1d(c); |
---|
77 | return NULL; |
---|
78 | } |
---|
79 | |
---|
80 | void del_aubio_conv1d(aubio_conv1d_t *c) |
---|
81 | { |
---|
82 | AUBIO_ASSERT(c); |
---|
83 | // destroy internals here |
---|
84 | if (c->kernel) { |
---|
85 | del_aubio_tensor(c->kernel); |
---|
86 | } |
---|
87 | if (c->bias) |
---|
88 | del_fvec(c->bias); |
---|
89 | #if defined(HAVE_BLAS) |
---|
90 | if (c->padded_input) del_aubio_tensor(c->padded_input); |
---|
91 | #endif |
---|
92 | AUBIO_FREE(c); |
---|
93 | } |
---|
94 | |
---|
95 | |
---|
96 | uint_t aubio_conv1d_set_stride(aubio_conv1d_t *c, uint_t stride[1]) |
---|
97 | { |
---|
98 | if ((sint_t)stride[0] < 1) return AUBIO_FAIL; |
---|
99 | c->stride_shape = stride[0]; |
---|
100 | return AUBIO_OK; |
---|
101 | } |
---|
102 | |
---|
103 | uint_t aubio_conv1d_get_stride(aubio_conv1d_t *c) |
---|
104 | { |
---|
105 | return c->stride_shape; |
---|
106 | } |
---|
107 | |
---|
108 | uint_t aubio_conv1d_get_output_shape(aubio_conv1d_t *c, |
---|
109 | aubio_tensor_t *input_tensor, |
---|
110 | uint_t *shape) |
---|
111 | { |
---|
112 | uint_t output_shape[2] = {0, c->n_filters}; |
---|
113 | uint_t padding_shape = 0; // total amount of padding |
---|
114 | uint_t padding_start = 0; |
---|
115 | |
---|
116 | // check input parameters |
---|
117 | AUBIO_ASSERT(input_tensor); |
---|
118 | AUBIO_ASSERT(shape); |
---|
119 | |
---|
120 | // reset output array |
---|
121 | shape[0] = 0; |
---|
122 | shape[1] = 0; |
---|
123 | |
---|
124 | switch (c->padding_mode) { |
---|
125 | case PAD_SAME: |
---|
126 | // compute output shape |
---|
127 | output_shape[0] = (uint_t)CEIL(input_tensor->shape[0] |
---|
128 | / (smpl_t)c->stride_shape); |
---|
129 | |
---|
130 | padding_shape = (output_shape[0] - 1) * c->stride_shape + |
---|
131 | c->kernel_shape - input_tensor->shape[0]; |
---|
132 | |
---|
133 | padding_start = FLOOR(padding_shape / 2); |
---|
134 | break; |
---|
135 | case PAD_VALID: |
---|
136 | output_shape[0] = (input_tensor->shape[0] - c->kernel_shape + 1) |
---|
137 | / c->stride_shape; |
---|
138 | |
---|
139 | padding_start = 0; |
---|
140 | break; |
---|
141 | case PAD_CAUSAL: |
---|
142 | // TODO |
---|
143 | return AUBIO_FAIL; |
---|
144 | default: |
---|
145 | return AUBIO_FAIL; |
---|
146 | } |
---|
147 | |
---|
148 | uint_t kernel_shape[3]; |
---|
149 | kernel_shape[0] = c->kernel_shape; // filter length |
---|
150 | kernel_shape[1] = input_tensor->shape[1]; // channels |
---|
151 | kernel_shape[2] = c->n_filters; // outputs |
---|
152 | |
---|
153 | if (c->kernel) del_aubio_tensor(c->kernel); |
---|
154 | if (c->bias) del_fvec(c->bias); |
---|
155 | |
---|
156 | c->kernel = new_aubio_tensor(3, kernel_shape); |
---|
157 | if (!c->kernel) return AUBIO_FAIL; |
---|
158 | c->bias = new_fvec(c->n_filters); |
---|
159 | |
---|
160 | // set internals upon success |
---|
161 | c->output_shape[0] = output_shape[0]; |
---|
162 | c->output_shape[1] = output_shape[1]; |
---|
163 | |
---|
164 | #if defined(HAVE_BLAS) |
---|
165 | if (c->padded_input) del_aubio_tensor(c->padded_input); |
---|
166 | uint_t padded_shape[2] = {input_tensor->shape[0] + padding_shape, |
---|
167 | input_tensor->shape[1]}; |
---|
168 | c->padded_input = new_aubio_tensor(2, padded_shape); |
---|
169 | #endif |
---|
170 | |
---|
171 | c->padding_start = padding_start; |
---|
172 | |
---|
173 | // set output |
---|
174 | shape[0] = output_shape[0]; |
---|
175 | shape[1] = output_shape[1]; |
---|
176 | |
---|
177 | aubio_conv1d_debug(c, input_tensor); |
---|
178 | |
---|
179 | return AUBIO_OK; |
---|
180 | } |
---|
181 | |
---|
182 | void aubio_conv1d_debug(aubio_conv1d_t *c, aubio_tensor_t *input_tensor) |
---|
183 | { |
---|
184 | // print some info |
---|
185 | AUBIO_ASSERT(c); |
---|
186 | uint_t n_params = (c->kernel->shape[0] * c->kernel->shape[2] + 1) |
---|
187 | * c->kernel->shape[1]; |
---|
188 | AUBIO_DBG("conv1d: %15s -> (%d, %d) (%d params)" |
---|
189 | " (weigths=(%d, %d, %d), stride=(%d,), pad_start=(%d,))\n", |
---|
190 | aubio_tensor_get_shape_string(input_tensor), |
---|
191 | c->output_shape[0], c->output_shape[1], |
---|
192 | n_params, |
---|
193 | c->kernel->shape[0], c->kernel->shape[1], c->kernel->shape[2], |
---|
194 | c->stride_shape, |
---|
195 | -c->padding_start); |
---|
196 | } |
---|
197 | |
---|
198 | uint_t aubio_conv1d_check_output_shape(aubio_conv1d_t *c, |
---|
199 | aubio_tensor_t *input_tensor, |
---|
200 | aubio_tensor_t *activations) |
---|
201 | { |
---|
202 | // fetch output_shape if it hasn't been done before |
---|
203 | if (c->output_shape[0] == 0 || |
---|
204 | c->output_shape[1] == 0) { |
---|
205 | if (!aubio_conv1d_get_output_shape(c, input_tensor, c->output_shape)) { |
---|
206 | return AUBIO_FAIL; |
---|
207 | } |
---|
208 | } |
---|
209 | |
---|
210 | // check we have as many filters as expected activation outputs |
---|
211 | if (activations->shape[1] != c->n_filters) return AUBIO_FAIL; |
---|
212 | if (activations->shape[1] != c->kernel->shape[2]) return AUBIO_FAIL; |
---|
213 | if (input_tensor->shape[1] != c->kernel->shape[1]) return AUBIO_FAIL; |
---|
214 | |
---|
215 | // check tensor activations has the expected sizes |
---|
216 | if (c->output_shape[0] != activations->shape[0]) return AUBIO_FAIL; |
---|
217 | if (c->output_shape[1] != activations->shape[1]) return AUBIO_FAIL; |
---|
218 | return AUBIO_OK; |
---|
219 | } |
---|
220 | |
---|
221 | #if !defined(HAVE_BLAS) |
---|
222 | void aubio_conv1d_do(aubio_conv1d_t *c, aubio_tensor_t *input_tensor, |
---|
223 | aubio_tensor_t *activations) |
---|
224 | { |
---|
225 | uint_t i, j, k, a; |
---|
226 | uint_t stride_a, kk; |
---|
227 | sint_t x; |
---|
228 | smpl_t s, w, bias, acc; |
---|
229 | |
---|
230 | AUBIO_ASSERT(c && input_tensor && activations); |
---|
231 | // check we have the correct output activation sizes |
---|
232 | if (aubio_conv1d_check_output_shape(c, input_tensor, activations)) |
---|
233 | { |
---|
234 | AUBIO_ERR("conv1d: check_output_shape failed\n"); |
---|
235 | return; |
---|
236 | } |
---|
237 | |
---|
238 | // for each kernel filter k |
---|
239 | for (i = 0; i < activations->shape[1]; i++) { |
---|
240 | // get bias |
---|
241 | bias = c->bias->data[i]; |
---|
242 | stride_a = 0; // k * c->stride_shape |
---|
243 | // for each output |
---|
244 | for (j = 0; j < activations->shape[0]; j++) { |
---|
245 | // reset output |
---|
246 | acc = 0; |
---|
247 | // compute convolution for one kernel |
---|
248 | for (a = 0; a < c->kernel_shape; a++) { |
---|
249 | x = stride_a + a - c->padding_start; |
---|
250 | if ((x > -1) && (x < (sint_t)input_tensor->shape[0])) { |
---|
251 | kk = 0; |
---|
252 | // for each input channel |
---|
253 | for (k = 0; k < input_tensor->shape[1]; k++) { |
---|
254 | // get kernel weight |
---|
255 | w = c->kernel->data[a][kk + i]; |
---|
256 | // get input sample |
---|
257 | s = input_tensor->data[x][k]; |
---|
258 | acc += w * s; |
---|
259 | kk += c->kernel->shape[2]; |
---|
260 | } |
---|
261 | } |
---|
262 | } |
---|
263 | stride_a += c->stride_shape; |
---|
264 | // apply bias |
---|
265 | acc += bias; |
---|
266 | // compute RELU |
---|
267 | activations->data[j][i] = MAX(acc, 0); |
---|
268 | } |
---|
269 | } |
---|
270 | } |
---|
271 | |
---|
272 | #else /* HAVE_BLAS */ |
---|
273 | |
---|
274 | // blas implementation |
---|
275 | // |
---|
276 | // uses gemv on the padded input to compute each output elements at once |
---|
277 | // |
---|
278 | // TODO |
---|
279 | // - avoid copy when padding_start == 0 |
---|
280 | // - optimize copying using tensor helpers |
---|
281 | |
---|
282 | void aubio_conv1d_do(aubio_conv1d_t *c, aubio_tensor_t *input_tensor, |
---|
283 | aubio_tensor_t *activations) |
---|
284 | { |
---|
285 | uint_t i, j; |
---|
286 | |
---|
287 | uint_t sdot_size = c->kernel->shape[0] * c->kernel->shape[1]; |
---|
288 | uint_t input_stride = c->stride_shape * c->padded_input->shape[1]; |
---|
289 | |
---|
290 | AUBIO_ASSERT(c && input_tensor && activations); |
---|
291 | if (aubio_conv1d_check_output_shape(c, input_tensor, activations)) |
---|
292 | { |
---|
293 | AUBIO_ERR("conv1d: check_output_shape failed\n"); |
---|
294 | return; |
---|
295 | } |
---|
296 | |
---|
297 | // copy input to padded version |
---|
298 | for (j = 0; j < input_tensor->shape[0]; j++) { |
---|
299 | for (i = 0; i < input_tensor->shape[1]; i++) { |
---|
300 | c->padded_input->data[j + c->padding_start][i] = |
---|
301 | input_tensor->data[j][i]; |
---|
302 | } |
---|
303 | } |
---|
304 | |
---|
305 | // for each output |
---|
306 | for (j = 0; j < activations->shape[0]; j++) { |
---|
307 | // for each row of activation output |
---|
308 | aubio_cblas__gemv(CblasRowMajor, CblasTrans, |
---|
309 | sdot_size, c->kernel->shape[2], 1., |
---|
310 | c->kernel->buffer, c->kernel->shape[2], |
---|
311 | c->padded_input->buffer + j * input_stride, 1, 0., |
---|
312 | activations->buffer + j * activations->shape[1], 1); |
---|
313 | } |
---|
314 | for (j = 0; j < activations->shape[0]; j++) { |
---|
315 | // for each kernel filter k |
---|
316 | for (i = 0; i < activations->shape[1]; i++) { |
---|
317 | activations->data[j][i] += c->bias->data[i]; |
---|
318 | activations->data[j][i] = MAX(activations->data[j][i], 0); |
---|
319 | } |
---|
320 | } |
---|
321 | } |
---|
322 | #endif /* HAVE_BLAS */ |
---|
323 | |
---|
324 | uint_t aubio_conv1d_set_padding_mode(aubio_conv1d_t *c, |
---|
325 | const char_t *padding_mode) |
---|
326 | { |
---|
327 | AUBIO_ASSERT(c && padding_mode); |
---|
328 | if (strncmp(padding_mode, "same", PATH_MAX) == 0) { |
---|
329 | c->padding_mode = PAD_SAME; |
---|
330 | } else if (strncmp(padding_mode, "valid", PATH_MAX) == 0) { |
---|
331 | c->padding_mode = PAD_VALID; |
---|
332 | } else { |
---|
333 | return AUBIO_FAIL; |
---|
334 | } |
---|
335 | return AUBIO_OK; |
---|
336 | } |
---|
337 | |
---|
338 | aubio_tensor_t *aubio_conv1d_get_kernel(aubio_conv1d_t* c) |
---|
339 | { |
---|
340 | AUBIO_ASSERT(c && c->kernel); |
---|
341 | return c->kernel; |
---|
342 | } |
---|
343 | |
---|
344 | fvec_t *aubio_conv1d_get_bias(aubio_conv1d_t* c) |
---|
345 | { |
---|
346 | AUBIO_ASSERT(c && c->bias); |
---|
347 | return c->bias; |
---|
348 | } |
---|