File size: 16,024 Bytes
00e6582
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#include <string.h>
#include <time.h>
#include "lstm.h"
// Activation Functions & Derivatives
static inline float sigmoid(float x) { return 1.0f / (1.0f + expf(-x)); }
static inline float sigmoid_deriv(float s) { return s * (1.0f - s); }
static inline float tanh_deriv(float t) { return 1.0f - t * t; }

// --- INITIALIZATION AND MEMORY MANAGEMENT ---

void init_weight(float *w, int rows, int cols) {
    float scale = sqrtf(6.0f / (rows + cols));
    for (int i = 0; i < rows * cols; i++) {
        w[i] = ((float)rand() / RAND_MAX) * 2.0f * scale - scale;
    }
}

LSTMNetwork* lstm_create(int input_dim, int hidden_dim) {
    LSTMNetwork *nn = (LSTMNetwork*)malloc(sizeof(LSTMNetwork));
    nn->input_dim = input_dim;
    nn->hidden_dim = hidden_dim;

    int w_size = hidden_dim * input_dim;
    int u_size = hidden_dim * hidden_dim;
    int b_size = hidden_dim;

    // Allocate parameters with explicit casts
    nn->W_i = (float*)malloc(w_size * sizeof(float)); nn->W_f = (float*)malloc(w_size * sizeof(float));
    nn->W_c = (float*)malloc(w_size * sizeof(float)); nn->W_o = (float*)malloc(w_size * sizeof(float));
    nn->U_i = (float*)malloc(u_size * sizeof(float)); nn->U_f = (float*)malloc(u_size * sizeof(float));
    nn->U_c = (float*)malloc(u_size * sizeof(float)); nn->U_o = (float*)malloc(u_size * sizeof(float));
    nn->b_i = (float*)calloc(b_size, sizeof(float));  nn->b_f = (float*)calloc(b_size, sizeof(float));
    nn->b_c = (float*)calloc(b_size, sizeof(float));  nn->b_o = (float*)calloc(b_size, sizeof(float));

    // Allocate gradients with explicit casts
    nn->dW_i = (float*)malloc(w_size * sizeof(float)); nn->dW_f = (float*)malloc(w_size * sizeof(float));
    nn->dW_c = (float*)malloc(w_size * sizeof(float)); nn->dW_o = (float*)malloc(w_size * sizeof(float));
    nn->dU_i = (float*)malloc(u_size * sizeof(float)); nn->dU_f = (float*)malloc(u_size * sizeof(float));
    nn->dU_c = (float*)malloc(u_size * sizeof(float)); nn->dU_o = (float*)malloc(u_size * sizeof(float));
    nn->db_i = (float*)malloc(b_size * sizeof(float)); nn->db_f = (float*)malloc(b_size * sizeof(float));
    nn->db_c = (float*)malloc(b_size * sizeof(float)); nn->db_o = (float*)malloc(b_size * sizeof(float));

    // Initialize weights
    init_weight(nn->W_i, hidden_dim, input_dim); init_weight(nn->W_f, hidden_dim, input_dim);
    init_weight(nn->W_c, hidden_dim, input_dim); init_weight(nn->W_o, hidden_dim, input_dim);
    init_weight(nn->U_i, hidden_dim, hidden_dim); init_weight(nn->U_f, hidden_dim, hidden_dim);
    init_weight(nn->U_c, hidden_dim, hidden_dim); init_weight(nn->U_o, hidden_dim, hidden_dim);

    // Forget gate bias initialization trick
    for (int i = 0; i < hidden_dim; i++) nn->b_f[i] = 1.0f;

    return nn;
}

LSTMHistory* lstm_create_history(int seq_len, int input_dim, int hidden_dim) {
    LSTMHistory *hist = (LSTMHistory*)malloc(sizeof(LSTMHistory));
    hist->seq_len = seq_len;
    hist->input_dim = input_dim;
    hist->hidden_dim = hidden_dim;

    hist->x = (float*)calloc(seq_len * input_dim, sizeof(float));
    hist->i_gate = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    hist->f_gate = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    hist->c_cand = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    hist->c = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    hist->o_gate = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    hist->h = (float*)calloc(seq_len * hidden_dim, sizeof(float));
    return hist;
}

LSTMDenseLayer* dense_create(int input_dim, int output_dim) {
    LSTMDenseLayer *layer = (LSTMDenseLayer*)malloc(sizeof(LSTMDenseLayer));
    layer->input_dim = input_dim;
    layer->output_dim = output_dim;

    layer->W = (float*)malloc(output_dim * input_dim * sizeof(float));
    layer->b = (float*)calloc(output_dim, sizeof(float));
    layer->dW = (float*)calloc(output_dim * input_dim, sizeof(float));
    layer->db = (float*)calloc(output_dim, sizeof(float));

    init_weight(layer->W, output_dim, input_dim);
    return layer;
}

void lstm_zero_grads(LSTMNetwork *nn) {
    int w_size = nn->hidden_dim * nn->input_dim;
    int u_size = nn->hidden_dim * nn->hidden_dim;
    int b_size = nn->hidden_dim;

    memset(nn->dW_i, 0, w_size * sizeof(float)); memset(nn->dW_f, 0, w_size * sizeof(float));
    memset(nn->dW_c, 0, w_size * sizeof(float)); memset(nn->dW_o, 0, w_size * sizeof(float));
    memset(nn->dU_i, 0, u_size * sizeof(float)); memset(nn->dU_f, 0, u_size * sizeof(float));
    memset(nn->dU_c, 0, u_size * sizeof(float)); memset(nn->dU_o, 0, u_size * sizeof(float));
    memset(nn->db_i, 0, b_size * sizeof(float)); memset(nn->db_f, 0, b_size * sizeof(float));
    memset(nn->db_c, 0, b_size * sizeof(float)); memset(nn->db_o, 0, b_size * sizeof(float));
}

// --- ALGORITHMIC EXECUTION PIPELINES ---

void lstm_forward(LSTMNetwork *nn, LSTMHistory *hist, const float *sequence_inputs) {
    int H = nn->hidden_dim;
    int I = nn->input_dim;
    memcpy(hist->x, sequence_inputs, hist->seq_len * I * sizeof(float));

    for (int t = 0; t < hist->seq_len; t++) {
        float *x_t = &hist->x[t * I];
        float *h_prev = (t == 0) ? NULL : &hist->h[(t - 1) * H];
        float *c_prev = (t == 0) ? NULL : &hist->c[(t - 1) * H];

        float *i_t = &hist->i_gate[t * H];
        float *f_t = &hist->f_gate[t * H];
        float *cc_t = &hist->c_cand[t * H];
        float *c_t = &hist->c[t * H];
        float *o_t = &hist->o_gate[t * H];
        float *h_t = &hist->h[t * H];

        for (int j = 0; j < H; j++) {
            float net_i = nn->b_i[j]; float net_f = nn->b_f[j];
            float net_c = nn->b_c[j]; float net_o = nn->b_o[j];

            for (int k = 0; k < I; k++) {
                net_i += nn->W_i[j * I + k] * x_t[k];
                net_f += nn->W_f[j * I + k] * x_t[k];
                net_c += nn->W_c[j * I + k] * x_t[k];
                net_o += nn->W_o[j * I + k] * x_t[k];
            }

            if (h_prev) {
                for (int k = 0; k < H; k++) {
                    net_i += nn->U_i[j * H + k] * h_prev[k];
                    net_f += nn->U_f[j * H + k] * h_prev[k];
                    net_c += nn->U_c[j * H + k] * h_prev[k];
                    net_o += nn->U_o[j * H + k] * h_prev[k];
                }
            }

            i_t[j] = sigmoid(net_i);
            f_t[j] = sigmoid(net_f);
            cc_t[j] = tanhf(net_c);
            o_t[j] = sigmoid(net_o);

            c_t[j] = i_t[j] * cc_t[j];
            if (c_prev) {
                c_t[j] += f_t[j] * c_prev[j];
            }
            h_t[j] = o_t[j] * tanhf(c_t[j]);
        }
    }
}

void lstm_backward(LSTMNetwork *nn, LSTMHistory *hist, const float *dh_sequence, float *dx_sequence) {
    int H = nn->hidden_dim;
    int I = nn->input_dim;

    float *dh_next = (float*)calloc(H, sizeof(float));
    float *dc_next = (float*)calloc(H, sizeof(float));
    float *dh_prev_step = (float*)calloc(H, sizeof(float));
    float *zero_buf = (float*)calloc(H, sizeof(float));

    lstm_zero_grads(nn);
    if (dx_sequence) {
        memset(dx_sequence, 0, hist->seq_len * I * sizeof(float));
    }

    for (int t = hist->seq_len - 1; t >= 0; t--) {
        float *x_t = &hist->x[t * I];
        float *i_t = &hist->i_gate[t * H];
        float *f_t = &hist->f_gate[t * H];
        float *cc_t = &hist->c_cand[t * H];
        float *c_t = &hist->c[t * H];
        float *o_t = &hist->o_gate[t * H];

        float *c_prev = (t == 0) ? zero_buf : &hist->c[(t - 1) * H];
        float *h_prev = (t == 0) ? zero_buf : &hist->h[(t - 1) * H];

        memset(dh_prev_step, 0, H * sizeof(float));

        for (int j = 0; j < H; j++) {
            float dh_total = dh_sequence[t * H + j] + dh_next[j];
            float dc_curr = dh_total * o_t[j] * tanh_deriv(tanhf(c_t[j])) + dc_next[j];

            float dnet_o = dh_total * tanhf(c_t[j]) * sigmoid_deriv(o_t[j]);
            float dnet_c = dc_curr * i_t[j] * tanh_deriv(cc_t[j]);
            float dnet_i = dc_curr * cc_t[j] * sigmoid_deriv(i_t[j]);
            float dnet_f = dc_curr * c_prev[j] * sigmoid_deriv(f_t[j]);

            nn->db_o[j] += dnet_o; nn->db_c[j] += dnet_c;
            nn->db_i[j] += dnet_i; nn->db_f[j] += dnet_f;

            for (int k = 0; k < I; k++) {
                nn->dW_o[j * I + k] += dnet_o * x_t[k];
                nn->dW_c[j * I + k] += dnet_c * x_t[k];
                nn->dW_i[j * I + k] += dnet_i * x_t[k];
                nn->dW_f[j * I + k] += dnet_f * x_t[k];

                if (dx_sequence) {
                    dx_sequence[t * I + k] += dnet_i * nn->W_i[j * I + k] +
                                              dnet_f * nn->W_f[j * I + k] +
                                              dnet_c * nn->W_c[j * I + k] +
                                              dnet_o * nn->W_o[j * I + k];
                }
            }

            for (int k = 0; k < H; k++) {
                nn->dU_o[j * H + k] += dnet_o * h_prev[k];
                nn->dU_c[j * H + k] += dnet_c * h_prev[k];
                nn->dU_i[j * H + k] += dnet_i * h_prev[k];
                nn->dU_f[j * H + k] += dnet_f * h_prev[k];

                dh_prev_step[k] += dnet_i * nn->U_i[j * H + k] +
                                   dnet_f * nn->U_f[j * H + k] +
                                   dnet_c * nn->U_c[j * H + k] +
                                   dnet_o * nn->U_o[j * H + k];
            }
            dc_next[j] = dc_curr * f_t[j];
        }
        memcpy(dh_next, dh_prev_step, H * sizeof(float));
    }

    free(dh_next); free(dc_next); free(dh_prev_step); free(zero_buf);
}

void lstm_update(LSTMNetwork *nn, float lr) {
    int w_size = nn->hidden_dim * nn->input_dim;
    int u_size = nn->hidden_dim * nn->hidden_dim;
    int b_size = nn->hidden_dim;

    for (int i = 0; i < w_size; i++) {
        nn->W_i[i] -= lr * nn->dW_i[i]; nn->W_f[i] -= lr * nn->dW_f[i];
        nn->W_c[i] -= lr * nn->dW_c[i]; nn->W_o[i] -= lr * nn->dW_o[i];
    }
    for (int i = 0; i < u_size; i++) {
        nn->U_i[i] -= lr * nn->dU_i[i]; nn->U_f[i] -= lr * nn->dU_f[i];
        nn->U_c[i] -= lr * nn->dU_c[i]; nn->U_o[i] -= lr * nn->dU_o[i];
    }
    for (int i = 0; i < b_size; i++) {
        nn->b_i[i] -= lr * nn->db_i[i]; nn->b_f[i] -= lr * nn->db_f[i];
        nn->b_c[i] -= lr * nn->db_c[i]; nn->b_o[i] -= lr * nn->db_o[i];
    }
}

// --- CLEANUP MANAGEMENT ---

void lstm_free_history(LSTMHistory *hist) {
    free(hist->x); free(hist->i_gate); free(hist->f_gate);
    free(hist->c_cand); free(hist->c); free(hist->o_gate); free(hist->h);
    free(hist);
}

void lstm_free(LSTMNetwork *nn) {
    free(nn->W_i); free(nn->W_f); free(nn->W_c); free(nn->W_o);
    free(nn->U_i); free(nn->U_f); free(nn->U_c); free(nn->U_o);
    free(nn->b_i); free(nn->b_f); free(nn->b_c); free(nn->b_o);
    free(nn->dW_i); free(nn->dW_f); free(nn->dW_c); free(nn->dW_o);
    free(nn->dU_i); free(nn->dU_f); free(nn->dU_c); free(nn->dU_o);
    free(nn->db_i); free(nn->db_f); free(nn->db_c); free(nn->db_o);
    free(nn);
}

void dense_free(LSTMDenseLayer *layer) {
    free(layer->W); free(layer->b);
    free(layer->dW); free(layer->db);
    free(layer);
}

// --- MAIN ARCHITECTURE ASSEMBLY ---

int CreateLSTM() {
    srand((unsigned int)time(NULL));

    // Structural Hyperparameters
    int seq_len    = 5;
    int input_dim  = 384;
    int hidden1    = 256;
    int hidden2    = 128;
    int output_dim = 16;
    float lr       = 0.01f;

    // Layer Object Allocations
    LSTMNetwork *layer1 = lstm_create(input_dim, hidden1);
    LSTMNetwork *layer2 = lstm_create(hidden1, hidden2);
    LSTMDenseLayer  *dense  = dense_create(hidden2, output_dim);

    // Context Sequence Memory History Structures
    LSTMHistory *hist1 = lstm_create_history(seq_len, input_dim, hidden1);
    LSTMHistory *hist2 = lstm_create_history(seq_len, hidden1, hidden2);

    // Variable Buffers
    float *raw_inputs    = (float*)calloc(seq_len * input_dim, sizeof(float));
    float *dense_outputs = (float*)calloc(seq_len * output_dim, sizeof(float));
    float *target_outs   = (float*)calloc(seq_len * output_dim, sizeof(float));

    // Recurrent Backward Gradient Trackers
    float *d_dense_out   = (float*)calloc(seq_len * output_dim, sizeof(float));
    float *dh_layer2     = (float*)calloc(seq_len * hidden2, sizeof(float));
    float *dh_layer1     = (float*)calloc(seq_len * hidden1, sizeof(float));

    // Populate mock training sequence data
    for (int i = 0; i < seq_len * input_dim; i++)  raw_inputs[i] = ((float)rand() / RAND_MAX) * 0.5f;
    for (int i = 0; i < seq_len * output_dim; i++) target_outs[i] = ((float)rand() / RAND_MAX);

    printf("Executing Deep Stacked Network Architecture Training Loop...\n\n");

    for (int epoch = 0; epoch <= 100; epoch++) {
        // ----------------------------------------------------
        // A. FORWARD PIPELINE
        // ----------------------------------------------------
        lstm_forward(layer1, hist1, raw_inputs);
        lstm_forward(layer2, hist2, hist1->h);

        // Project Output from Hidden Layer 2 through Dense Weights
        for (int t = 0; t < seq_len; t++) {
            for (int i = 0; i < output_dim; i++) {
                float net = dense->b[i];
                for (int j = 0; j < hidden2; j++) {
                    net += dense->W[i * hidden2 + j] * hist2->h[t * hidden2 + j];
                }
                dense_outputs[t * output_dim + i] = net;
            }
        }

        // Calculate loss (MSE Evaluation)
        float total_loss = 0.0f;
        for (int i = 0; i < seq_len * output_dim; i++) {
            float diff = dense_outputs[i] - target_outs[i];
            total_loss += 0.5f * diff * diff;
            d_dense_out[i] = diff; // Backprop seeding derivative
        }

        if (epoch % 20 == 0) {
            printf("Epoch %3d | Aggregated Sequence MSE Loss: %f\n", epoch, total_loss / seq_len);
        }

        // ----------------------------------------------------
        // B. BACKPROPAGATION THROUGH TIME (BPTT) CHAIN
        // ----------------------------------------------------

        // Step 1: Initialize Dense Gradients
        memset(dense->db, 0, output_dim * sizeof(float));
        memset(dense->dW, 0, output_dim * hidden2 * sizeof(float));
        memset(dh_layer2, 0, seq_len * hidden2 * sizeof(float));

        // Step 2: Clear and push dense deviations back down to hidden states
        for (int t = 0; t < seq_len; t++) {
            for (int i = 0; i < output_dim; i++) {
                float d_out = d_dense_out[t * output_dim + i];
                dense->db[i] += d_out;
                for (int j = 0; j < hidden2; j++) {
                    dense->dW[i * hidden2 + j] += d_out * hist2->h[t * hidden2 + j];
                    dh_layer2[t * hidden2 + j] += d_out * dense->W[i * hidden2 + j];
                }
            }
        }

        // Step 3: Recurrent step extraction back through Layer 2
        lstm_backward(layer2, hist2, dh_layer2, dh_layer1);

        // Step 4: Recurrent step extraction back through Layer 1
        lstm_backward(layer1, hist1, dh_layer1, NULL);

        // ----------------------------------------------------
        // C. DESCENT WEIGHT PARAMETER OPTIMIZATION
        // ----------------------------------------------------
        lstm_update(layer1, lr);
        lstm_update(layer2, lr);

        for (int i = 0; i < output_dim * hidden2; i++) dense->W[i] -= lr * dense->dW[i];
        for (int i = 0; i < output_dim; i++)           dense->b[i] -= lr * dense->db[i];
    }

    // Free all assigned dynamic spaces safely
    free(raw_inputs); free(dense_outputs); free(target_outs);
    free(d_dense_out); free(dh_layer2); free(dh_layer1);

    lstm_free_history(hist1); lstm_free_history(hist2);
    lstm_free(layer1);        lstm_free(layer2);
    dense_free(dense);

    printf("\nTraining terminated. Free routines executed cleanly.\n");
    return 0;
}