| 3222 | // ------------------------------------------------------------------------------------ |
| 3223 | |
| 3224 | void compute_act_halt_probabilities( |
| 3225 | resizable_tensor& halt_probs, |
| 3226 | resizable_tensor& logits, |
| 3227 | const tensor& input_data, |
| 3228 | const tensor& halt_params, |
| 3229 | long batch_size, |
| 3230 | long seq_len, |
| 3231 | long feature_dim |
| 3232 | ) |
| 3233 | { |
| 3234 | const float* in_ptr = input_data.host(); |
| 3235 | const float* W_halt = halt_params.host(); |
| 3236 | const float b_halt = halt_params.host()[feature_dim]; |
| 3237 | float* logits_ptr = logits.host(); |
| 3238 | float* halt_probs_ptr = halt_probs.host(); |
| 3239 | |
| 3240 | const long d_model = feature_dim / input_data.k(); |
| 3241 | const long num_channels = input_data.k(); |
| 3242 | |
| 3243 | for (long pos = 0; pos < batch_size * seq_len; ++pos) { |
| 3244 | const long n = pos / seq_len; |
| 3245 | const long s = pos % seq_len; |
| 3246 | |
| 3247 | float logit = b_halt; |
| 3248 | |
| 3249 | for (long c = 0; c < num_channels; ++c) { |
| 3250 | for (long d = 0; d < d_model; ++d) { |
| 3251 | const long in_idx = ((n * num_channels + c) * seq_len + s) * d_model + d; |
| 3252 | const long weight_idx = c * d_model + d; |
| 3253 | logit += in_ptr[in_idx] * W_halt[weight_idx]; |
| 3254 | } |
| 3255 | } |
| 3256 | |
| 3257 | logits_ptr[pos] = logit; |
| 3258 | |
| 3259 | halt_probs_ptr[pos] = 1.0f / (1.0f + std::exp(-logit)); |
| 3260 | } |
| 3261 | } |
| 3262 | |
| 3263 | void update_act_state( |
| 3264 | resizable_tensor& output, |