"""Build the single 61-input, 14-output Arena policy plus its timed roll cue.""" import argparse import json from pathlib import Path import numpy as np import onnx from onnx import helper as h, numpy_helper, TensorProto as T import policy as P def export(checkpoint,folder): theta=json.loads(Path(checkpoint).read_text())['weights'] model=onnx.load(str(P.base_path())) nodes=[] constants=list(model.graph.initializer) source=model.graph.input[0].name skate_output=model.graph.output[0].name nodes.extend(model.graph.node) def constant(name,values,dtype=np.float32): constants.append(numpy_helper.from_array(np.asarray(values,dtype=dtype),name)) return name def op(kind,name,*args,**kwargs): nodes.append(h.make_node(kind,list(args),[name],name=name,**kwargs)) return name cue=op('Gather','roll_cue',source,constant('cue_index',[60],np.int64),axis=1) raw=op('Mul','scaled_cue',cue,constant('sixteen',16)) raw=op('Sub','one_based',raw,constant('one',1)) raw=op('Cast','index64',op('Round','rounded',raw),to=T.INT64) raw=op('Reshape','frame',raw,constant('scalar_shape',[],np.int64)) zero=constant('index_zero',0,np.int64) limit=constant('last_index',len(P.ACTION)-1,np.int64) idx=op('Clip','safe_index',raw,zero,limit) use=op('And','is_trick',op('GreaterOrEqual','after_start',raw,zero),op('LessOrEqual','before_end',raw,limit)) obs=op('Gather','motor_sensors',source,constant('sensor_indices',P.INDICES,np.int64),axis=1) def table(name,arr): return op('Gather',name,constant(name+'_table',arr),idx,axis=0) delta=op('Sub','sensor_error',obs,table('expected',P.INPUT)) correction=op('MatMul','feedback',delta,table('gains',P.feedback(theta))) clipped=op('Clip','bounded_feedback',correction,constant('feedback_min',-.8),constant('feedback_max',.8)) joints=op('Add','roll_joints',table('trained_joints',P.ACTION),clipped) op('Where','motor_targets',use,joints,skate_output) graph=h.make_graph(nodes,'crown_roll',list(model.graph.input), [h.make_tensor_value_info('motor_targets',T.FLOAT,[1,14])],constants) built=h.make_model(graph,ir_version=max(8,model.ir_version),opset_imports=[h.make_opsetid('',18)]) onnx.checker.check_model(built, full_check=True) folder=Path(folder) folder.mkdir(parents=True,exist_ok=True) onnx.save(built,str(folder/'policy.onnx')) frames=[] previous=None for i in range(P.STEPS): cmd=P.command(i) if previous is None or not np.array_equal(previous,cmd): frames.append(dict(t=round(i*P.DT,2),twist=[.65 if cmd[0] else 0,0,0], body=[0,0,0,0,0,float(cmd[-1])])) previous=cmd act=dict(timeline_version=2,duration_s=P.STEPS*P.DT, ducks=[dict(policy='self',start=dict(x=-.85,y=0,yaw=0),keyframes=frames)]) (folder/'timeline.json').write_text(json.dumps(act,indent=2)+'\n') if __name__=='__main__': parser=argparse.ArgumentParser() parser.add_argument('checkpoint',type=Path) parser.add_argument('folder',type=Path) args=parser.parse_args() export(args.checkpoint,args.folder)