From 5cae610328befb6fc1cfe851ace43231ad42faba Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?David=20Lanzend=C3=B6rfer?= <leviathan@libresilicon.com>
Date: Thu, 4 Jul 2024 22:41:34 +0100
Subject: [PATCH] Updat firmware
---
firmware/firmware.c | 29 ++++++++++++++++++++------
firmware/include/rnn.h | 17 ++++++++--------
firmware/rnn.c | 46 +++++++++++++++++++++++-------------------
3 files changed, 57 insertions(+), 35 deletions(-)
diff --git a/firmware/firmware.c b/firmware/firmware.c
index bc8835f..29893e1 100644
--- a/firmware/firmware.c
+++ b/firmware/firmware.c
@@ -79,7 +79,9 @@ void main()
TRAIN_STORE_LEARNING_RATE,
TRAIN_STORE_DECAY_RATE,
TRAIN_RUN_EPOCHS,
- TRAIN_RUN_SINGLE_EPOCH
+ TRAIN_RUN_SINGLE_EPOCH,
+ // The fun part: Inference
+ PREDICT
} command_mode;
command_mode = START;
@@ -93,13 +95,12 @@ void main()
int new_value;
// Training
- uint32_t new_token;
- uint32_t token_series[MAX_NUM_TOKENS];
+ uint16_t new_token;
+ uint16_t token_series[MAX_NUM_TOKENS];
int token_counter = 0;
uint32_t learning_rate = 0;
uint32_t decay_rate = 0;
uint32_t epoch_value = 0;
-
while(true) {
@@ -145,6 +146,9 @@ void main()
else if(!strcmp(msg,"TRAIN")) {
command_mode = TRAIN;
}
+ else if(!strcmp(msg,"PREDICT")) {
+ command_mode = PREDICT;
+ }
break;
case INIT:
@@ -249,11 +253,15 @@ void main()
case WRITE_LAYER_WEIGHTS:
response = "END";
- if(value_write_counter<current_max_num_values) {
+ if((value_write_counter<current_max_num_values) && (current_layer_value == LAYER_VALUE_TYPE_WEIGHTS) ) {
response = numstr;
new_value = atoi(numstr);
set_layer_weight(current_layer, neuron_idx, value_write_counter, new_value);
value_write_counter++;
+ } else if (current_layer_value == LAYER_VALUE_TYPE_BIAS) {
+ response = numstr;
+ new_value = atoi(numstr);
+ set_layer_bias(current_layer, neuron_idx, new_value);
}
break;
@@ -286,7 +294,7 @@ void main()
new_token = atoi(numstr);
token_series[token_counter] = new_token;
token_counter++;
- response = numstr;
+ response = itoa(token_counter, numstr, 10);
} else {
response = "END";
}
@@ -316,6 +324,15 @@ void main()
response = itoa(new_value, numstr, 10);
break;
+ case PREDICT:
+ new_token = atoi(numstr);
+ for(int i=0;i<MSGBUF;i++) {
+ numstr[i]=0;
+ }
+ new_token = predict_next_token(new_token);
+ response = itoa(new_token, numstr, 10);
+ break;
+
}
write_response(response);
}
diff --git a/firmware/include/rnn.h b/firmware/include/rnn.h
index 9f8fa6d..943e91a 100644
--- a/firmware/include/rnn.h
+++ b/firmware/include/rnn.h
@@ -22,7 +22,8 @@
#define TRAINING_SINGLE_SHOT (1<<8)
-#define ENCODER_WEIGHT_COUNT (NUM_INPUT_SYNAPSES+NUM_INPUT_NEURONS)
+//#define ENCODER_WEIGHT_COUNT (NUM_INPUT_SYNAPSES+NUM_INPUT_NEURONS)
+#define ENCODER_WEIGHT_COUNT (NUM_INPUT_SYNAPSES)
#define HIDDEN_WEIGHT_COUNT (NUM_INPUT_NEURONS+NUM_HIDDEN_NEURONS_H)
#define DECODER_WEIGHT_COUNT (NUM_HIDDEN_NEURONS_H)
@@ -113,7 +114,7 @@ void set_layer_alpha(
/*
* Run inference
*/
-uint32_t predict_next_token(uint32_t token);
+uint16_t predict_next_token(uint16_t token);
/*
* Run trainingg
@@ -123,7 +124,7 @@ uint32_t predict_next_token(uint32_t token);
* The mask is an XOR value of the bits of the supposed
* output and what the network actually delivered
*/
-uint32_t mask_back_propgatation(uint32_t mask);
+uint16_t mask_back_propgatation(uint16_t mask);
/*
* Access the LSFR output inside the ANN controller
@@ -165,8 +166,8 @@ char* run_training(
int num_epochs,
uint32_t learning_rate_zero,
uint32_t decay_rate,
- uint32_t *xp,
- uint32_t xn
+ uint16_t *xp,
+ int xi
);
/*
@@ -180,10 +181,10 @@ char* run_training(
* Returns 0 when successful
* Returns 1 when unsuccessful
*/
-uint32_t run_training_single_epoch(
+int run_training_single_epoch(
int epoch,
uint32_t learning_rate_zero,
uint32_t decay_rate,
- uint32_t *xp,
- uint32_t xi
+ uint16_t *xp,
+ int xi
);
diff --git a/firmware/rnn.c b/firmware/rnn.c
index b1e5f02..810774d 100644
--- a/firmware/rnn.c
+++ b/firmware/rnn.c
@@ -131,9 +131,9 @@ void set_layer_alpha(
/*
* Run inference
*/
-uint32_t predict_next_token(uint32_t token)
+uint16_t predict_next_token(uint16_t token)
{
- uint32_t ret;
+ uint16_t ret;
ann_cmd = RUN_INFERENCE|INFERENCE_START;
ann_value1 = token;
ann_cmd = RUN_INFERENCE|INFERENCE_FETCH;
@@ -149,7 +149,7 @@ uint32_t predict_next_token(uint32_t token)
* The mask is an XOR value of the bits of the supposed
* output and what the network actually delivered
*/
-uint32_t mask_back_propgatation(uint32_t mask)
+uint16_t mask_back_propgatation(uint16_t mask)
{
ann_cmd = RUN_TRAINING|TRAINING_SINGLE_SHOT;
ann_value1 = mask;
@@ -256,17 +256,18 @@ void set_bias_values(int bias)
* Returns 0 when successful
* Returns 1 when unsuccessful
*/
-uint32_t run_training_single_epoch(
+int run_training_single_epoch(
int epoch,
uint32_t learning_rate_zero,
uint32_t decay_rate,
- uint32_t *xp,
- uint32_t xi
+ uint16_t *xp,
+ int xi
)
{
+ int error;
int last_val;
- uint32_t train_mask;
- uint32_t y = xp[xi];;
+ uint16_t train_mask;
+ uint16_t y = xp[xi-1];;
/*
* Revert the LSTM states
@@ -276,16 +277,17 @@ uint32_t run_training_single_epoch(
* Put all the prior tokens of the time series
* into the LSTM for reaching the state we wish to train on
*/
- for(int xpi=0; xpi<xi; xpi++) {
+ for(int xpi=0; xpi<xi-1; xpi++) {
last_val = predict_next_token(xp[xpi]);
}
+
/*
* Check whether the network has already been trained
* Set return value to zero and exit the loop if so
*/
train_mask = last_val ^ y;
- if(!train_mask) {
- return train_mask;
+ if(last_val==y) {
+ return 0;
}
/*
* Determine the training mask my finding out which
@@ -293,8 +295,11 @@ uint32_t run_training_single_epoch(
*/
set_alpha(learning_rate_zero/(1+(decay_rate*epoch)));
mask_back_propgatation(train_mask);
-
- return train_mask;
+
+ error = 0;
+ for(int i=0;i<32;i++)
+ error+=(train_mask>>i)&1;
+ return error;
}
@@ -313,13 +318,13 @@ int run_training_epochs(
int num_epochs,
uint32_t learning_rate_zero,
uint32_t decay_rate,
- uint32_t *xp,
- uint32_t xi
+ uint16_t *xp,
+ int xi
)
{
int last_val;
- uint32_t train_mask;
- uint32_t y = xp[xi];;
+ uint16_t train_mask;
+ uint16_t y = xp[xi];;
for(int epoch=0; epoch<num_epochs;epoch++) {
if(run_training_single_epoch(
@@ -349,8 +354,8 @@ char* run_training(
int num_epochs,
uint32_t learning_rate_zero,
uint32_t decay_rate,
- uint32_t *xp,
- uint32_t xn
+ uint16_t *xp,
+ int xn
)
{
int last_val;
@@ -367,7 +372,7 @@ char* run_training(
* xi tells the functions where y is located in the array (x -> y)
* xi must be xn-1 at most
*/
- for(uint32_t xi=1; xi<xn; xi++) {
+ for(int xi=1; xi<xn; xi++) {
ret = run_training_epochs(
num_epochs,
learning_rate_zero,
@@ -391,4 +396,3 @@ char* run_training(
return msgbuf;
}
-
--
GitLab