def _cmd_tune(args):
transport_launcher = device_util.DeviceTransportLauncher(
{'use_tracker': True})
log_util.config(['autotune', args.model_spec], logging.INFO)
model_inst, _ = model.instantiate_from_spec(args.model_spec)
tasks, log_file = tune_model(args, transport_launcher, model_inst)
analyze(args, model_inst, tasks, log_file, promote=True)
def _cmd_launch_transport(args):
log_util.config(['autotune', args.model_spec])
transport_launcher = device_util.DeviceTransportLauncher(
{'use_tracker': True})
generate_config = not args.skip_writing_transport_config
launch_kw = {'generate_config': generate_config}
if generate_config:
model_inst, _ = model.instantiate_from_spec(args.model_spec)
index_and_task = None
if args.task_index is not None:
tasks = model_inst.extract_tunable_tasks(model_inst.build_model())
index_and_task = (args.task_index, tasks[args.task_index])
launch_kw[
'generate_config_func'] = tvm.micro.device.arm.stm32f746xx.generate_config
launch_kw['generate_config_kw'] = {
'section_constraints':
model_inst.section_constraints(index_and_task)
}
with transport_launcher.launch(**launch_kw):
print('Transport launched. Press Ctrl+C to terminate.')
try:
while True:
time.sleep(10)
except KeyboardInterrupt:
print('Caught SIGINT; shutting down')
def _cmd_rpc_dev_config(args):
log_util.config([], logging.INFO, console_only=True)
transport_launcher = device_util.DeviceTransportLauncher(
{'use_tracker': True})
model_inst, _ = model.instantiate_from_spec(args.model_spec)
index_and_task = None
if args.task_index is not None:
tasks = model_inst.extract_tunable_tasks(model_inst.build_model())
index_and_task = (args.task_index, tasks[args.task_index])
transport_launcher.generate_rpc_server_configs(
tvm.micro.device.arm.stm32f746xx.generate_config, {
'section_constraints':
model_inst.section_constraints(index_and_task)
})
transport_launcher.generate_openocd_configs()
print(
f'Wrote OpenOCD and RPC server configs underneath {transport_launcher.work_dirtree_root}'
)
def main():
args = parse_args()
log_util.config(['eval', '-'.join(args.model_specs)], level=args.log_level)
model_inst_setting = collections.OrderedDict()
for spec in args.model_specs:
assert spec not in model_inst_setting, f'spec {spec} given twice'
model_inst_setting[spec] = model.instantiate_from_spec(spec)
validate_against = None
if args.validate_against:
assert args.validate_against not in model_inst_setting, (
f'--validate-against={args.validate_against} also given in model_specs '
'command-line argument')
validate_against = model.instantiate_from_spec(args.validate_against)
model_inst_setting[args.validate_against] = validate_against
dataset_generator_name = next(iter(
model_inst_setting.values()))[0].dataset_generator_name()
for spec, (m, _) in model_inst_setting.items():
m.dataset_generator_name() == dataset_generator_name, (
f'expected all models to share the same dataset, but {spec} has '
f'{m.dataset_generator_name()}')
dataset_gen = dataset.DatasetGenerator.instantiate(
dataset_generator_name, {'shuffle': not validate_against})
util.DEBUG_MODE = args.use_debug_executor
samples = dataset_gen.generate(args.num_samples)
results = {}
with contextlib.ExitStack() as all_models_stack:
if args.debug_micro_execution:
_LOG.warn(
'NOTE: to debug micro execution, compiled object files will be left in your '
'temp directory at: %s',
contrib_utils.TempDirectory._DEBUG_PARENT_DIR)
_LOG.warn(
'This is a limitation of the current microTVM compilation structure'
)
all_models_stack.enter_context(
contrib_utils.TempDirectory.set_keep_for_debug())
for spec, (model_inst, setting) in model_inst_setting.items():
with contextlib.ExitStack() as model_stack:
if args.use_tuned_schedule:
if args.use_tuned_schedule == USE_DEFAULT_TUNED_SCHEDULE:
tuned_schedule = autotvm_log_util.get_default_log_path(
autotvm_log_util.compute_job_name(
spec, model_inst))
if not os.path.exists(tuned_schedule):
_LOG.warning(
'Tuned schedule for %s not found; proceeding without: %s',
spec, tuned_schedule)
tuned_schedule = None
else:
tuned_schedule = args.use_tuned_schedule
if tuned_schedule is not None:
model_stack.enter_context(
autotvm.apply_history_best(tuned_schedule))
compiled = model_inst.build_model()
results[spec] = globals()[f'eval_{setting}'](args, model_inst,
compiled, samples)
if args.validate_against:
for i in range(args.num_samples):
allclose = {}
for model_spec in args.model_specs:
allclose[model_spec] = np.allclose(
results[model_spec][i]['label'].astype('float32'), results[
args.validate_against][i]['label'].astype('float32'))
_LOG.info(f'Sample {i} ---->')
rows = []
rows.append([['model_name', 'setting', 'config']] +
[x for x in range(10)])
def _add_row(model_spec, values):
model_spec_parts = model_spec.split(':', 3)
if len(model_spec_parts) == 2:
model_spec_parts.append('')
rows.append([model_spec_parts] + values)
for model_spec in args.model_specs:
color = ''
if model_spec != args.validate_against:
if not allclose[model_spec]:
level = logging.ERROR
else:
level = logging.INFO
_add_row(model_spec, list(results[model_spec][i]['label']))
_add_row(args.validate_against,
results[args.validate_against][i]['label'].tolist())
spacings = []
for c in range(0, 3):
spacing = max(len(r[0][c]) + 1 for r in rows)
spacings.append(f'{{0:{spacing}s}}')
_LOG.info(''.join(
[spacings[c].format(rows[0][0][c]) for c in range(0, 3)] +
['{0:5d}'.format(c) for c in rows[0][1:]]))
format_string = f'{{0:{spacing}s}}'
for r in rows[1:]:
model_spec_parts = ''.join(
[spacings[c].format(r[0][c]) for c in range(0, 3)])
color = r[1]
result_str = ''.join([' {0:+04d}'.format(y) for y in r[1:]])
_LOG.log(level, '%s%s', model_spec_parts, result_str)
def _cmd_analyze(args):
log_util.config([], logging.INFO, console_only=True)
model_inst, _ = model.instantiate_from_spec(args.model_spec)
tasks = model_inst.extract_tunable_tasks(model_inst.build_model())
analyze(args, model_inst, tasks, args.log_file, promote=args.promote)
def main():
args = parse_args()
model_inst, _ = model.instantiate_from_spec(args.model_specs[0])
generate_project(model_inst, args.project_dir)
