T=args.T)
return utils_elf.GCWrapper(GC, inputs, replies, name2idx, params)
nIter = 5000
elapsed_wait_only = 0
import pickle
import argparse
if __name__ == '__main__':
parser = argparse.ArgumentParser()
loader = Loader()
args = args_loader(parser, [loader])
def actor(sel, sel_gpu, reply):
'''
import pdb
pdb.set_trace()
pickle.dump(utils_elf.to_numpy(sel), open("tmp%d.bin" % k, "wb"), protocol=2)
'''
reply[0]["a"][:] = 0
GC = loader.initialize()
GC.reg_callback("actor", actor)
before = datetime.now()
GC.Start()
params["hist_len"] = args.hist_len
params["T"] = args.T
return utils_elf.GCWrapper(GC, inputs, replies, name2idx, params)
cmd_line = "--num_games 16 --batchsize 4 --hist_len 1 --frame_skip 4 --actor_only"
nIter = 5000
elapsed_wait_only = 0
if __name__ == '__main__':
parser = argparse.ArgumentParser()
loader = Loader()
args = args_loader(parser, [loader], cmd_line=cmd_line.split(" "))
GC = loader.initialize()
def actor(sel, sel_gpu, reply):
# pickle.dump(to_numpy(sel), open("tmp%d.bin" % k, "wb"), protocol=2)
reply[0]["a"][:] = 0
GC.reg_callback("actor", actor)
reward_dist = Counter()
before = datetime.now()
GC.Start()
import tqdm
method = method_class()
args_providers = [sampler, trainer, game, runner, model_loader, method]
eval_only = os.environ.get("eval_only", False)
has_eval_process = os.environ.get("eval_process", False)
if has_eval_process or eval_only:
eval_process = EvaluationProcess()
evaluator = Eval()
args_providers.append(eval_process)
args_providers.append(evaluator)
else:
eval_process = None
all_args = args_loader(parser, args_providers)
GC = game.initialize()
GC.setup_gpu(0)
all_args.method_class = method_class
model = model_loader.load_model(GC.params)
mi = ModelInterface()
mi.add_model("model", model, optim_params={ "lr" : 0.001})
mi.add_model("actor", model, copy=True, cuda=True)
method.set_model_interface(mi)
trainer.setup(sampler=sampler, mi=mi, rl_method=method)
if use_multi_process:
GC.reg_callback("actor", trainer.actor)