mawei/StarPilot
1
0
Fork 0
mirror of https://github.com/firestar5683/StarPilot.git synced 2026-08-21 08:14:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Rocket Launcher Model (#25963)

Browse Source 
* 1456d261-d232-4654-8885-4d9fde883894/440 6b7d7cec-ead8-40f3-86cc-86d52c9b03fe/300

* compute only 9 tokens: 1456d261-d232-4654-8885-4d9fde883894/440 6b7d7cec-ead8-40f3-86cc-86d52c9b03fe/300

* tinygrad: cleanup gather

* 1456d261-d232-4654-8885-4d9fde883894/440 6b7d7cec-ead8-40f3-86cc-86d52c9b03fe/700

* empty commit for tests

* bump tinygrad

* dont use tinygrad matmul for now

* bump tinygrad

* 1456d261-d232-4654-8885-4d9fde883894/440 e63ab895-2222-4abd-a9a5-af86bb70e260/700

* float16 1456d261-d232-4654-8885-4d9fde883894/440 e63ab895-2222-4abd-a9a5-af86bb70e260/700

* increase steer rate cost

* Revert "increase steer rate cost"

This reverts commit 74ce9ab9be7ef17ecfec931f96851b12f37f2336.

* fork tinygrad

* empty commit for tests

* basics

* Kinda works

* new lat

* new tuning

* Move LATMPCN so scons compiles

* Update long weights

* Add tinygrad optim

* Update model ref

* update weights

* Update ref

* Try

* Error message for field ignore

* update model regf

* ref commit

* Fix onnx test

Co-authored-by: Yassine Yousfi <yyousfi1@binghamton.edu>
old-commit-hash: cb0b7375b728d1b6e92db68c9ba55f0f54c09a3f
This commit is contained in:
HaraldSchafer
2022-10-05 21:43:38 -07:00
committed by GitHub
parent f93f4e9f9b
commit a8ec6f4cfd
19 changed files with 93 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
selfdrive/modeld/SConscript
+2 -2
View File
@@ -71,9 +71,9 @@ if use_thneed and arch == "larch64" or GetOption('pc_thneed'):
fn = File("models/supercombo").abspath
if GetOption('pc_thneed'):
cmd = f"cd {Dir('#').abspath}/tinygrad_repo && NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
cmd = f"cd {Dir('#').abspath}/tinygrad_repo && GPU=1 NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
else:
cmd = f"cd {Dir('#').abspath}/tinygrad_repo && FLOAT16=1 PYOPENCL_NO_CACHE=1 MATMUL=1 NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
cmd = f"cd {Dir('#').abspath}/tinygrad_repo && FLOAT16=1 MATMUL=1 PYOPENCL_NO_CACHE=1 NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
# is there a better way then listing all of tinygrad?
lenv.Command(fn + ".thneed", [fn + ".onnx",
selfdrive/modeld/models/driving.cc
+12 -4
View File
@@ -41,11 +41,11 @@ void model_init(ModelState* s, cl_device_id device_id, cl_context context) {
&s->output[0], NET_OUTPUT_SIZE, USE_GPU_RUNTIME, true, false, context);
#ifdef TEMPORAL
s->m->addRecurrent(&s->output[OUTPUT_SIZE], TEMPORAL_SIZE);
s->m->addRecurrent(&s->feature_buffer[0], TEMPORAL_SIZE);
#endif
#ifdef DESIRE
s->m->addDesire(s->pulse_desire, DESIRE_LEN);
s->m->addDesire(s->pulse_desire, DESIRE_LEN*(HISTORY_BUFFER_LEN+1));
#endif
#ifdef TRAFFIC_CONVENTION
@@ -56,18 +56,20 @@ void model_init(ModelState* s, cl_device_id device_id, cl_context context) {
ModelOutput* model_eval_frame(ModelState* s, VisionBuf* buf, VisionBuf* wbuf,
const mat3 &transform, const mat3 &transform_wide, float *desire_in, bool is_rhd, bool prepare_only) {
#ifdef DESIRE
std::memmove(&s->pulse_desire[0], &s->pulse_desire[DESIRE_LEN], sizeof(float) * DESIRE_LEN*HISTORY_BUFFER_LEN);
if (desire_in != NULL) {
for (int i = 1; i < DESIRE_LEN; i++) {
// Model decides when action is completed
// so desire input is just a pulse triggered on rising edge
if (desire_in[i] - s->prev_desire[i] > .99) {
s->pulse_desire[i] = desire_in[i];
s->pulse_desire[DESIRE_LEN*(HISTORY_BUFFER_LEN-1)+i] = desire_in[i];
} else {
s->pulse_desire[i] = 0.0;
s->pulse_desire[DESIRE_LEN*(HISTORY_BUFFER_LEN-1)+i] = 0.0;
}
s->prev_desire[i] = desire_in[i];
}
}
LOGT("Desire enqueued");
#endif
int rhd_idx = is_rhd;
@@ -92,6 +94,12 @@ ModelOutput* model_eval_frame(ModelState* s, VisionBuf* buf, VisionBuf* wbuf,
s->m->execute();
LOGT("Execution finished");
#ifdef TEMPORAL
std::memmove(&s->feature_buffer[0], &s->feature_buffer[FEATURE_LEN], sizeof(float) * FEATURE_LEN*(HISTORY_BUFFER_LEN-1));
std::memcpy(&s->feature_buffer[FEATURE_LEN*(HISTORY_BUFFER_LEN-1)], &s->output[OUTPUT_SIZE], sizeof(float) * FEATURE_LEN);
LOGT("Features enqueued");
#endif
return (ModelOutput*)&s->output;
}
selfdrive/modeld/models/driving.h
+12 -3
View File
@@ -16,6 +16,8 @@
#include "selfdrive/modeld/models/commonmodel.h"
#include "selfdrive/modeld/runners/run.h"
constexpr int FEATURE_LEN = 2048;
constexpr int HISTORY_BUFFER_LEN = 99;
constexpr int DESIRE_LEN = 8;
constexpr int DESIRE_PRED_LEN = 4;
constexpr int TRAFFIC_CONVENTION_LEN = 2;
@@ -233,6 +235,11 @@ struct ModelOutputMeta {
};
static_assert(sizeof(ModelOutputMeta) == sizeof(ModelOutputDesireProb) + sizeof(float) + (sizeof(ModelOutputDisengageProb)*DISENGAGE_LEN) + (sizeof(ModelOutputBlinkerProb)*BLINKER_LEN) + (sizeof(ModelOutputDesireProb)*DESIRE_PRED_LEN));
struct ModelOutputFeatures {
std::array<float, FEATURE_LEN> feature;
};
static_assert(sizeof(ModelOutputFeatures) == (sizeof(float)*FEATURE_LEN));
struct ModelOutput {
const ModelOutputPlans plans;
const ModelOutputLaneLines lane_lines;
@@ -244,22 +251,24 @@ struct ModelOutput {
};
constexpr int OUTPUT_SIZE = sizeof(ModelOutput) / sizeof(float);
#ifdef TEMPORAL
constexpr int TEMPORAL_SIZE = 512;
constexpr int TEMPORAL_SIZE = HISTORY_BUFFER_LEN * FEATURE_LEN;
#else
constexpr int TEMPORAL_SIZE = 0;
#endif
constexpr int NET_OUTPUT_SIZE = OUTPUT_SIZE + TEMPORAL_SIZE;
constexpr int NET_OUTPUT_SIZE = OUTPUT_SIZE + FEATURE_LEN;
// TODO: convert remaining arrays to std::array and update model runners
struct ModelState {
ModelFrame *frame = nullptr;
ModelFrame *wide_frame = nullptr;
std::array<float, HISTORY_BUFFER_LEN * FEATURE_LEN> feature_buffer = {};
std::array<float, NET_OUTPUT_SIZE> output = {};
std::unique_ptr<RunModel> m;
#ifdef DESIRE
float prev_desire[DESIRE_LEN] = {};
float pulse_desire[DESIRE_LEN] = {};
float pulse_desire[DESIRE_LEN*(HISTORY_BUFFER_LEN+1)] = {};
#endif
#ifdef TRAFFIC_CONVENTION
float traffic_convention[TRAFFIC_CONVENTION_LEN] = {};
selfdrive/modeld/models/supercombo.onnx
+2 -2
View File
@@ -1,3 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:50c7fc8565ac69a4b9a0de122e961326820e78bf13659255a89d0ed04be030d5
size 95167481
oid sha256:30f30bc1251c03db135564ecbf7dc0bc96cbb07be0ebd3691edd8d555dc087fa
size 58539693
selfdrive/modeld/runners/onnx_runner.py
+8 -5
View File
@@ -9,6 +9,8 @@ os.environ["OMP_WAIT_POLICY"] = "PASSIVE"
import onnxruntime as ort # pylint: disable=import-error
ORT_TYPES_TO_NP_TYPES = {'tensor(float16)': np.float16, 'tensor(float)': np.float32, 'tensor(uint8)': np.uint8}
def read(sz, tf8=False):
dd = []
gt = 0
@@ -18,7 +20,7 @@ def read(sz, tf8=False):
assert(len(st) > 0)
dd.append(st)
gt += len(st)
r = np.frombuffer(b''.join(dd), dtype=np.uint8 if tf8 else np.float32).astype(np.float32)
r = np.frombuffer(b''.join(dd), dtype=np.uint8 if tf8 else np.float32)
if tf8:
r = r / 255.
return r
@@ -29,22 +31,23 @@ def write(d):
def run_loop(m, tf8_input=False):
ishapes = [[1]+ii.shape[1:] for ii in m.get_inputs()]
keys = [x.name for x in m.get_inputs()]
itypes = [ORT_TYPES_TO_NP_TYPES[x.type] for x in m.get_inputs()]
# run once to initialize CUDA provider
if "CUDAExecutionProvider" in m.get_providers():
m.run(None, dict(zip(keys, [np.zeros(shp, dtype=np.float32) for shp in ishapes])))
m.run(None, dict(zip(keys, [np.zeros(shp, dtype=itp) for shp, itp in zip(ishapes, itypes)])))
print("ready to run onnx model", keys, ishapes, file=sys.stderr)
while 1:
inputs = []
for k, shp in zip(keys, ishapes):
for k, shp, itp in zip(keys, ishapes, itypes):
ts = np.product(shp)
#print("reshaping %s with offset %d" % (str(shp), offset), file=sys.stderr)
inputs.append(read(ts, (k=='input_img' and tf8_input)).reshape(shp))
inputs.append(read(ts, (k=='input_img' and tf8_input)).reshape(shp).astype(itp))
ret = m.run(None, dict(zip(keys, inputs)))
#print(ret, file=sys.stderr)
for r in ret:
write(r)
write(r.astype(np.float32))
if __name__ == "__main__":
selfdrive/modeld/thneed/thneed_common.cc
+1 -1
View File
@@ -12,7 +12,7 @@ map<pair<cl_kernel, int>, int> g_args_size;
map<cl_program, string> g_program_source;
void Thneed::stop() {
printf("Thneed::stop: recorded %lu commands\n", cmds.size());
//printf("Thneed::stop: recorded %lu commands\n", cmds.size());
record = false;
}