From 34ee2fdff2a212658f0a9052d1d4ca0f3fa1ea3f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?David=20Lanzend=C3=B6rfer?= <leviathan@libresilicon.com>
Date: Mon, 1 Jul 2024 07:57:43 +0100
Subject: [PATCH] Try single epoch training for now
---
firmware/firmware.c | 19 +++++++--
firmware/include/rnn.h | 19 +++++++++
firmware/rnn.c | 89 +++++++++++++++++++++++++++++-------------
src/py/tty3.py | 17 ++++----
src/rtl/neuron.v | 1 -
5 files changed, 106 insertions(+), 39 deletions(-)
diff --git a/firmware/firmware.c b/firmware/firmware.c
index bf107ce..8b66b95 100644
--- a/firmware/firmware.c
+++ b/firmware/firmware.c
@@ -78,7 +78,8 @@ void main()
TRAIN_STORE_TOKENS,
TRAIN_STORE_LEARNING_RATE,
TRAIN_STORE_DECAY_RATE,
- TRAIN_RUN_EPOCHS
+ TRAIN_RUN_EPOCHS,
+ TRAIN_RUN_SINGLE_EPOCH
} command_mode;
command_mode = START;
@@ -97,6 +98,8 @@ void main()
int token_counter = 0;
uint32_t learning_rate = 0;
uint32_t decay_rate = 0;
+ uint32_t epoch_value = 0;
+
while(true) {
@@ -277,6 +280,10 @@ void main()
command_mode = TRAIN_RUN_EPOCHS;
response = "OK";
}
+ else if(!strcmp(msg,"RUN_SINGLE_EPOCH")) {
+ command_mode = TRAIN_RUN_SINGLE_EPOCH;
+ response = "OK";
+ }
break;
case TRAIN_STORE_TOKENS:
@@ -303,11 +310,17 @@ void main()
break;
case TRAIN_RUN_EPOCHS:
- uint32_t num_epochs = atoi(numstr);
- response = run_training(response, num_epochs, learning_rate, decay_rate, token_series, token_counter);
+ epoch_value = atoi(numstr);
+ response = run_training(response, epoch_value, learning_rate, decay_rate, token_series, token_counter);
command_mode = START;
break;
+ case TRAIN_RUN_SINGLE_EPOCH:
+ epoch_value = atoi(numstr);
+ new_value = run_training_single_epoch(epoch_value, learning_rate, decay_rate, token_series, token_counter);
+ response = new_value?"SUCCESS":"FAILURE";
+ break;
+
}
write_response(response);
}
diff --git a/firmware/include/rnn.h b/firmware/include/rnn.h
index 48ede09..9fc064b 100644
--- a/firmware/include/rnn.h
+++ b/firmware/include/rnn.h
@@ -152,3 +152,22 @@ char* run_training(
uint32_t *xp,
uint32_t xn
);
+
+/*
+ * Run training cycle for 1 epoch
+ * epoch: The epoch we're in
+ * learning_rate_zero: initial learning rate
+ * decay_rate: the decay rate for gradient decay
+ * xp: Pointer to array of values
+ * xi: Index of y
+ *
+ * Returns 0 when successful
+ * Returns 1 when unsuccessful
+ */
+int run_training_single_epoch(
+ int epoch,
+ uint32_t learning_rate_zero,
+ uint32_t decay_rate,
+ uint32_t *xp,
+ uint32_t xi
+);
diff --git a/firmware/rnn.c b/firmware/rnn.c
index 05c7b69..921a1ec 100644
--- a/firmware/rnn.c
+++ b/firmware/rnn.c
@@ -387,6 +387,59 @@ void set_bias_values(int bias)
}
}
+/*
+ * Run training cycle for 1 epoch
+ * epoch: The epoch we're in
+ * learning_rate_zero: initial learning rate
+ * decay_rate: the decay rate for gradient decay
+ * xp: Pointer to array of values
+ * xi: Index of y
+ *
+ * Returns 0 when successful
+ * Returns 1 when unsuccessful
+ */
+int run_training_single_epoch(
+ int epoch,
+ uint32_t learning_rate_zero,
+ uint32_t decay_rate,
+ uint32_t *xp,
+ uint32_t xi
+)
+{
+ int last_val;
+ uint32_t train_mask;
+ uint32_t y = xp[xi];;
+
+ /*
+ * Revert the LSTM states
+ */
+ reset_network();
+ /*
+ * 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++) {
+ 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
+ */
+ if(last_val==y) {
+ return 1;
+ }
+ /*
+ * Determine the training mask my finding out which
+ * neurons misfired or didn't fire although they should have
+ */
+ train_mask = last_val ^ y;
+ set_alpha(learning_rate_zero/(1+(decay_rate*epoch)));
+ mask_back_propgatation(train_mask);
+
+ return 0;
+}
+
+
/*
* Run training cycle for N epochs
* num_epochs: Amount of epochs
@@ -406,41 +459,21 @@ int run_training_epochs(
uint32_t xi
)
{
- int ret = 1;
int last_val;
uint32_t train_mask;
uint32_t y = xp[xi];;
for(int epoch=0; epoch<num_epochs;epoch++) {
- /*
- * Revert the LSTM states
- */
- reset_network();
- /*
- * 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++) {
- 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
- */
- if(last_val==y) {
- ret = 0;
- break;
- }
- /*
- * Determine the training mask my finding out which
- * neurons misfired or didn't fire although they should have
- */
- train_mask = last_val ^ y;
- set_alpha(learning_rate_zero/(1+(decay_rate*epoch)));
- mask_back_propgatation(train_mask);
+ if(run_training_single_epoch(
+ epoch,
+ learning_rate_zero,
+ decay_rate,
+ xp,
+ xi
+ )) return 0;
}
- return ret;
+ return 1;
}
diff --git a/src/py/tty3.py b/src/py/tty3.py
index 0922d96..1078dcd 100644
--- a/src/py/tty3.py
+++ b/src/py/tty3.py
@@ -17,7 +17,7 @@ server = get_fac_wrapper("telnet")
INTMAX=2147483647
lr=0.95
-decay_rate=10
+decay_rate=100
run_command(server,"HELLO")
@@ -31,7 +31,6 @@ run_command(server,"RWEIGHTS")
run_command(server,"INIT")
run_command(server,"BIAS")
-
run_command(server,"TRAIN")
run_command(server,"LEARNING_RATE")
# PicoRV does not have a floating point unit
@@ -41,7 +40,6 @@ run_command(server,"LEARNING_RATE")
lr=lr*INTMAX
run_command(server,str(int(lr)))
-
run_command(server,"TRAIN")
run_command(server,"DECAY_RATE")
# The decay rate is a hyper parameter
@@ -62,9 +60,13 @@ tok=tokens[1]
run_command(server,str(tok))
run_command(server,"DONE")
run_command(server,"TRAIN")
-run_command(server,"RUN_EPOCHS")
-run_command(server,str(1000))
+run_command(server,"RUN_SINGLE_EPOCH")
+for i in range(0,500):
+ ret = run_command(server,str(i))
+ if "SUCCESS" in ret:
+ break
+'''
run_command(server,"TRAIN")
run_command(server,"TOKENS")
tok=tokens[1]
@@ -74,7 +76,7 @@ run_command(server,str(tok))
run_command(server,"DONE")
run_command(server,"TRAIN")
run_command(server,"RUN_EPOCHS")
-run_command(server,str(1000))
+run_command(server,str(500))
run_command(server,"TRAIN")
run_command(server,"TOKENS")
@@ -85,7 +87,7 @@ run_command(server,str(tok))
run_command(server,"DONE")
run_command(server,"TRAIN")
run_command(server,"RUN_EPOCHS")
-run_command(server,str(1000))
+run_command(server,str(500))
# Upload token series
run_command(server,"TRAIN")
@@ -96,6 +98,7 @@ run_command(server,"DONE")
run_command(server,"TRAIN")
run_command(server,"RUN_EPOCHS")
run_command(server,str(1000))
+'''
# Store the weights and biases
weights_and_biases = dump_neural_network(server)
diff --git a/src/rtl/neuron.v b/src/rtl/neuron.v
index d1d5d2c..6895e96 100644
--- a/src/rtl/neuron.v
+++ b/src/rtl/neuron.v
@@ -101,7 +101,6 @@ module neuron #(
else
if(!backprop_done && backprop_running)
begin
- $display("Backprop with dW=", $signed(dw));
// Adjust normal weights
if( regidx < NUMBER_SYNAPSES )
begin
--
GitLab