Instructions to use JackDH/microduck-crown-roll with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Microduck
How to use JackDH/microduck-crown-roll with Microduck:
# Replace SLOT with the slot specified in the model card (walk, stand, sitstand, ground_pick, kick_left, kick_right, roulade). sudo robotctl policy load SLOT JackDH/microduck-crown-roll
- Notebooks
- Google Colab
- Kaggle
Download training/export.py from JackDH/microduck-crown-roll: direct link, hf CLI and curl.
- Browser
- Download file 3.21 kB
-
https://huggingface.co/JackDH/microduck-crown-roll/resolve/main/training/export.py
- Command line
-
hf download hf://JackDH/microduck-crown-roll/training/export.py
-
curl -L -o export.py https://huggingface.co/JackDH/microduck-crown-roll/resolve/main/training/export.py
3.21 kB
| """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) | |