From 868cbd8815edff7a7c6fd1d2a48092ccce2f009b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?David=20Lanzend=C3=B6rfer?= <leviathan@libresilicon.com>
Date: Fri, 28 Jun 2024 16:13:06 +0100
Subject: [PATCH] First attempt of running training
---
Makefile | 46 +++++------------------------
firmware/firmware.c | 34 ++++++++++++++++-----
firmware/include/rnn.h | 18 ++++++++++++
firmware/rnn.c | 67 +++++++++++++++++++++---------------------
src/py/tty3.py | 52 ++++++++++++++++++++++++++++----
5 files changed, 132 insertions(+), 85 deletions(-)
diff --git a/Makefile b/Makefile
index 34dc6d0..6e53018 100644
--- a/Makefile
+++ b/Makefile
@@ -8,30 +8,6 @@ NETWORK_PARAMS = \
TARGET_BOARD?=ecp5_minifpga
-RISCV_GNU_TOOLCHAIN_INSTALL_PREFIX ?= /opt/riscv
-PYTHON = python3
-GCC_WARNS = -Werror -Wall -Wextra -Wshadow -Wundef -Wpointer-arith -Wcast-qual -Wcast-align -Wwrite-strings
-GCC_WARNS += -Wredundant-decls -Wstrict-prototypes -Wmissing-prototypes -pedantic # -Wconversion
-TOOLCHAIN_PREFIX = $(RISCV_GNU_TOOLCHAIN_INSTALL_PREFIX)/bin/riscv32-unknown-elf-
-COMPRESSED_ISA = C
-
-FIRMWARE_OBJS= \
-firmware/start.o \
-firmware/main.o \
-firmware/print.o
-
-FIRMWARE_FILES= \
-firmware/start.s \
-firmware/rnn.c \
-firmware/firmware.c \
-firmware/io.c \
-firmware/string.c
-
-FIRMWARE_CFLAGS=
-FIRMWARE_CFLAGS+=-mabi=ilp32 -march=rv32imc
-FIRMWARE_CFLAGS+=-Wl,--build-id=none,-Bstatic,-T,firmware/sections.lds,--strip-debug
-FIRMWARE_CFLAGS+=-ffreestanding -nostdlib
-
BENCHES=
VERILATOR_DIR?=/usr/share/verilator/include
@@ -78,7 +54,7 @@ src/cpptb/params.h:
echo "#define $(PARAM) $(VALUE)" >> $@ ; \
)
-firmware/defines.h:
+firmware/include/defines.h:
@ $(foreach PARAMS,$(NETWORK_PARAMS), \
$(eval PARAM = $(word 1,$(subst :, ,$(PARAMS)))) \
$(eval VALUE = $(word 2,$(subst :, ,$(PARAMS)))) \
@@ -100,20 +76,14 @@ result/soc.json: result/firmware.hex src/rtl/params.vh result
result/soc_out.config: $(LPFILE) # result/soc.json
nextpnr-ecp5 --json result/soc.json --lpf $(LPFILE) --textcfg $@ --freq 50 --package CABGA256 --lpf-allow-unconstrained
-result/firmware.hex: firmware/firmware.bin submodules/picorv32/firmware/makehex.py result
- $(TOOLCHAIN_PREFIX)objcopy -O verilog firmware/firmware.elf result/firmware.hex
- cp result/firmware.hex firmware.hex
-
-firmware/firmware.bin: firmware/firmware.elf
- $(TOOLCHAIN_PREFIX)objcopy -O binary $< $@
-
-firmware/firmware.elf: firmware/sections.lds firmware/defines.h $(FIRMWARE_FILES)
- $(TOOLCHAIN_PREFIX)gcc $(CFLAGS) $(FIRMWARE_CFLAGS) -o $@ $(FIRMWARE_FILES)
+result/firmware.hex: result firmware/include/defines.h
+ make -C firmware firmware.hex
+ cp firmware/firmware.hex firmware.hex
+ cp firmware/firmware.hex result/firmware.hex
result:
mkdir -p result
clean:
- rm -rf firmware/firmware.elf result/firmware.hex \
- firmware/firmware.bin firmware/*.o firmware/firmware.map \
- *.bin *.vcd obj_dir src/cpptb/*.o soc.json src/rtl/params.vh \
- src/cpptb/params.h firmware/defines.h
+ rm -rf *.bin *.vcd obj_dir src/cpptb/*.o soc.json src/rtl/params.vh \
+ src/cpptb/params.h firmware/include/defines.h result/firmware.hex
+ make -C firmware clean
diff --git a/firmware/firmware.c b/firmware/firmware.c
index aaf0253..8558656 100644
--- a/firmware/firmware.c
+++ b/firmware/firmware.c
@@ -77,7 +77,8 @@ void main()
TRAIN,
TRAIN_STORE_TOKENS,
TRAIN_STORE_LEARNING_RATE,
- TRAIN_PROCESS
+ TRAIN_STORE_DECAY_RATE,
+ TRAIN_RUN_EPOCHS
} command_mode;
command_mode = START;
@@ -94,6 +95,8 @@ void main()
uint32_t new_token;
uint32_t token_series[MAX_NUM_TOKENS];
int token_counter;
+ uint32_t learning_rate;
+ uint32_t decay_rate;
while(true) {
@@ -266,10 +269,14 @@ void main()
response = "OK";
command_mode = TRAIN_STORE_LEARNING_RATE;
}
- /*else if(!strcmp(msg,"RUN")) {
- command_mode = TRAIN_PROCESS;
+ else if(!strcmp(msg,"DECAY_RATE")) {
response = "OK";
- }*/
+ command_mode = TRAIN_STORE_DECAY_RATE;
+ }
+ else if(!strcmp(msg,"RUN_EPOCHS")) {
+ command_mode = TRAIN_RUN_EPOCHS;
+ response = "OK";
+ }
break;
case TRAIN_STORE_TOKENS:
@@ -277,15 +284,28 @@ void main()
new_token = atoi(numstr);
token_series[token_counter] = new_token;
token_counter++;
- response = "OK";
+ response = numstr;
} else {
response = "END";
}
break;
case TRAIN_STORE_LEARNING_RATE:
- new_token = atoi(numstr);
- response = "OK";
+ learning_rate = atoi(numstr);
+ response = numstr;
+ command_mode = START;
+ break;
+
+ case TRAIN_STORE_DECAY_RATE:
+ decay_rate = atoi(numstr);
+ response = numstr;
+ command_mode = START;
+ break;
+
+ case TRAIN_RUN_EPOCHS:
+ uint32_t num_epochs = atoi(numstr);
+ response = run_training(response, num_epochs, learning_rate, decay_rate, token_series[0], token_series[1]);
+ command_mode = START;
break;
}
diff --git a/firmware/include/rnn.h b/firmware/include/rnn.h
index ad36ee7..bfa3461 100644
--- a/firmware/include/rnn.h
+++ b/firmware/include/rnn.h
@@ -134,3 +134,21 @@ void set_bias_values(int bias);
* Resetting the network: Clear LTSM
*/
void reset_network();
+
+/*
+ * Run training cycle for N epochs
+ * msgbuf: buffer for infos
+ * num_epochs: Amount of epochs
+ * learning_rate_zero: initial learning rate
+ * decay_rate: the decay rate for gradient decay
+ * x: input token
+ * y: output token
+ */
+char* run_training(
+ char *msgbuf,
+ int num_epochs,
+ uint32_t learning_rate_zero,
+ uint32_t decay_rate,
+ uint32_t x,
+ uint32_t y
+);
diff --git a/firmware/rnn.c b/firmware/rnn.c
index af34057..7bf595a 100644
--- a/firmware/rnn.c
+++ b/firmware/rnn.c
@@ -318,28 +318,21 @@ void reset_network()
/*
* Setting the delta for the weights
* We try to achieve something like a gradients
- * with random noise added
*/
-void set_dws(int dw, bool randomize)
+void set_alpha(int alpha)
{
int i, x, y;
- int dwd;
-
- dwd = (dw/NUM_HIDDEN_NEURONS_W);
+
for(i=0;i<NUM_INPUT_NEURONS;i++) {
- if(randomize) set_layer_weight(LAYER_TYPE_ENCODER, LAYER_VALUE_TYPE_DELTA_W, i, 0, dwd+get_random_char()/NUM_HIDDEN_NEURONS_W);
- else set_layer_weight(LAYER_TYPE_ENCODER, LAYER_VALUE_TYPE_DELTA_W, i, 0, dwd);
+ set_layer_weight(LAYER_TYPE_ENCODER, LAYER_VALUE_TYPE_DELTA_W, i, 0, alpha);
}
for(x=0;x<NUM_HIDDEN_NEURONS_W;x++) {
- dwd = (x+1)*(dw/NUM_HIDDEN_NEURONS_W);
for(y=0;y<NUM_HIDDEN_NEURONS_H;y++) {
- if(randomize) set_layer_weight(LAYER_TYPE_HIDDEN,LAYER_VALUE_TYPE_DELTA_W, x*NUM_HIDDEN_NEURONS_H+y, 0, dwd+get_random_char()/NUM_HIDDEN_NEURONS_W);
- else set_layer_weight(LAYER_TYPE_HIDDEN,LAYER_VALUE_TYPE_DELTA_W, x*NUM_HIDDEN_NEURONS_H+y, 0, dwd);
+ set_layer_weight(LAYER_TYPE_HIDDEN,LAYER_VALUE_TYPE_DELTA_W, x*NUM_HIDDEN_NEURONS_H+y, 0, alpha);
}
}
for(i=0;i<NUM_OUTPUT_NEURONS;i++) {
- if(randomize) set_layer_weight(LAYER_TYPE_DECODER,LAYER_VALUE_TYPE_DELTA_W, i, 0, dw+get_random_char());
- else set_layer_weight(LAYER_TYPE_DECODER,LAYER_VALUE_TYPE_DELTA_W, i, 0, dw);
+ set_layer_weight(LAYER_TYPE_DECODER,LAYER_VALUE_TYPE_DELTA_W, i, 0, alpha);
}
}
@@ -394,32 +387,38 @@ void set_bias_values(int bias)
}
}
-void run_training()
+/*
+ * Run training cycle for N epochs
+ * msgbuf: buffer for infos
+ * num_epochs: Amount of epochs
+ * learning_rate_zero: initial learning rate
+ * decay_rate: the decay rate for gradient decay
+ * x: input token
+ * y: output token
+ */
+char* run_training(
+ char *msgbuf,
+ int num_epochs,
+ uint32_t learning_rate_zero,
+ uint32_t decay_rate,
+ uint32_t x,
+ uint32_t y
+)
{
- int dw = 200000000;
- int bias = 0;
int last_val;
uint32_t train_mask;
- last_val = predict_next_token(27017);
-
- /* Compare values
- * The phrase "Taxation is theft"
- * tokenized with the GPT2 tokenizer
- * is tensor([[27017, 341, 318, 12402]])
- * this means the next value we want is 341
- */
- while(last_val!=341) {
-
- train_mask = last_val ^ 341;
-
- //set_dws(dw);
- set_dws(dw, false);
- dw=dw/100;
- dw=99*dw;
+ for(int epoch=0; epoch<num_epochs;epoch++) {
+ reset_network();
+ last_val = predict_next_token(x);
+ if(last_val==y) {
+ return "SUCCESS";
+ break;
+ }
+ train_mask = last_val ^ y;
+ set_alpha(learning_rate_zero/(1+(decay_rate*epoch)));
mask_back_propgatation(train_mask);
-
- last_val = predict_next_token(27017);
}
-
+ return "FAIL";
}
+
diff --git a/src/py/tty3.py b/src/py/tty3.py
index 7ff1660..a0df053 100644
--- a/src/py/tty3.py
+++ b/src/py/tty3.py
@@ -6,6 +6,12 @@ from fac_tools import load_weights_and_biases
from transformers import AutoTokenizer
+tokenizer = AutoTokenizer.from_pretrained("gpt2")
+prompt = "Taxation is theft"
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+tokens = input_ids[0].numpy()
+print(tokens)
+
server = get_fac_wrapper("telnet")
print("Reading in JSON")
@@ -13,12 +19,14 @@ with open("test_files/weights_and_biases.json", "r") as f:
weights_and_biases = json.loads(f.read())
f.close()
-#load_weights_and_biases(server, weights_and_biases)
+load_weights_and_biases(server, weights_and_biases)
+
+INTMAX=2147483647
+lr=0.95
+decay_rate=1
+
+run_command(server,"HELLO")
-tokenizer = AutoTokenizer.from_pretrained("gpt2")
-prompt = "Taxation is theft"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids
-tokens = input_ids[0].numpy()
run_command(server,"TRAIN")
run_command(server,"TOKENS")
for tok in tokens:
@@ -26,8 +34,40 @@ for tok in tokens:
run_command(server,"DONE")
run_command(server,"TRAIN")
-run_command(server,"SHOOT")
+run_command(server,"LEARNING_RATE")
+# PicoRV does not have a floating point unit
+# All weights are stored as signed 32 bit integers
+# this means we have to do the math here and multiply
+# the maximum integer value by our desired 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
+# Because delta W aka alpha is calculated
+# as the initial learning rate α0 multiplied
+# as shown below anything floaty doesn't make sense
+# anyway. Usually the decay rate is set to 1
+# α=(1/(1+decayRate×epochNumber))*​α0
+run_command(server,str(decay_rate))
+
+
+run_command(server,"TRAIN")
+run_command(server,"RUN_EPOCHS")
+run_command(server,str(100))
+
+weights_and_biases = dump_neural_network(server)
+
+j = json.dumps(weights_and_biases, indent=4, sort_keys=True)
+print(j)
run_command(server,"TERMINATE")
server.close()
+
+print("Writing out JSON")
+with open("result/weights_and_biases_trained.json", "w") as f:
+ f.write(j)
+ f.close()
--
GitLab