def run(host_name='local', pipeline=''): if pipeline != '': return exp_shared.load_exp(pipeline) # Create and run new experiment queue = exp_shared.create_queue(host_name) queue.sync('.', '.', exclude=['pipelines/*', 'fig/*', 'old/*', 'cogsci/*'], sync_to=sge.SyncTo.REMOTE, recursive=True) exp = Experiment(exp_name='rl', fixed_params=[('loss_type', 'REINFORCE'), ('bw_boost', 1), ('env', 'wcs'), ('max_epochs', 25000), # 10000 ('hidden_dim', 20), ('batch_size', 2048), ('perception_dim', 3), ('target_dim', 330), ('print_interval', 1000), ('msg_dim', 15)], param_ranges=[('avg_over', range(1)), # 50 ('perception_noise',[0, 0.1, 0.3, 0.5, 1, 2, 5]), # np.logspace(0, 9, num=10, base=2)) [0, 10, 20, 30, 40, 50, 80, 120, 160, 320]), [0, 25, 50, 100],[0, 10, 20, 40, 80, 160, 320] ('com_noise', [0, 0.1, 0.3, 0.5, 1, 2, 5])], # np.logspace(-3, 6, num=10, base=2) [0, 0.1, 0.3, 0.5, 1] [0, 0.5, 3, 10, 20, 50] queue=queue) queue.sync(exp.pipeline_path, exp.pipeline_path, sync_to=sge.SyncTo.REMOTE, recursive=True) env = exp.run(com_enviroments.make, exp.fixed_params['env']).result() exp_i = 0 for (params_i, params_v) in exp: print('Scheduled %d experiments out of %d' % (exp_i, len(list(exp)))) exp_i += 1 agent_a = agents.SoftmaxAgent(msg_dim=exp.fixed_params['msg_dim'], hidden_dim=exp.fixed_params['hidden_dim'], color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) agent_b = agents.SoftmaxAgent(msg_dim=exp.fixed_params['msg_dim'], hidden_dim=exp.fixed_params['hidden_dim'], color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) game = com_game.NoisyChannelGame(com_noise=params_v[exp.axes['com_noise']], msg_dim=exp.fixed_params['msg_dim'], max_epochs=exp.fixed_params['max_epochs'], perception_noise=params_v[exp.axes['perception_noise']], batch_size=exp.fixed_params['batch_size'], print_interval=exp.fixed_params['print_interval'], loss_type=exp.fixed_params['loss_type'], bw_boost=exp.fixed_params['bw_boost']) agent_a_trained = exp.run(game.play, env, agent_a, agent_b).result() exp.set_result('agent_a', params_i, agent_a_trained) return exp
def run(): exp = Experiment(exp_name='local_experiment', fixed_params=[('env', 'wgs'), ('max_epochs', 10000), #10000 ('hidden_dim', 20), ('batch_size', 100), ('perception_dim', 3), ('target_dim', 330), ('print_interval', 1000)], param_ranges=[('avg_over', range(2)), # 50 ('perception_noise', [0, 25]), # [0, 25, 50, 100], ('msg_dim', range(9, 11)), #3, 12 ('com_noise', np.linspace(start=0, stop=0.5, num=2))]) env = com_enviroments.make(exp.fixed_params['env']) exp_i = 0 for (params_i, params_v) in exp: print('Scheduled %d experiments out of %d' % (exp_i, len(list(exp)))) exp_i += 1 agent_a = agent_b = agents.SoftmaxAgent(msg_dim=params_v[exp.axes['msg_dim']], hidden_dim=exp.fixed_params['hidden_dim'], color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) game = com_game.NoisyChannelGame(com_noise=params_v[exp.axes['com_noise']], msg_dim=params_v[exp.axes['msg_dim']], max_epochs=exp.fixed_params['max_epochs'], perception_noise=params_v[exp.axes['perception_noise']], batch_size=exp.fixed_params['batch_size'], print_interval=exp.fixed_params['print_interval']) game_outcome = game.play(env, agent_a, agent_b) V = evaluate.agent_language_map(env, a=game_outcome) exp.set_result('gibson_cost', params_i, game.compute_gibson_cost(env, a=game_outcome)[1]) exp.set_result('regier_cost', params_i, evaluate.communication_cost_regier(env, V=V)) exp.set_result('wellformedness', params_i, evaluate.wellformedness(env, V=V)) exp.set_result('term_usage', params_i, evaluate.compute_term_usage(V=V)) print("\nAll tasks queued to clusters") # wait for all tasks to complete exp.save() return exp.pipeline_name
def run(host_name): # Create and run new experiment queue = exp_shared.create_queue(host_name) queue.sync('.', '.', exclude=['pipelines/*', 'fig/*', 'old/*', 'cogsci/*'], sync_to=sge.SyncTo.REMOTE, recursive=True) exp = Experiment( exp_name='num_b', fixed_params=[ ('env', 'numbers'), ('max_epochs', 1000), #10000 ('hidden_dim', 3), ('batch_size', 100), ('perception_dim', 1), ('target_dim', 100), ('print_interval', 10) ], param_ranges=[ ('avg_over', [0]), # 50 ('perception_noise', [0]), # [0, 25, 50, 100], ('msg_dim', [3]), #3, 12 ('com_noise', [0]) ], queue=queue) queue.sync(exp.pipeline_path, exp.pipeline_path, sync_to=sge.SyncTo.REMOTE, recursive=True) env = exp.run(com_enviroments.make, exp.fixed_params['env']).result() exp_i = 0 for (params_i, params_v) in exp: print('Scheduled %d experiments out of %d' % (exp_i, len(list(exp)))) exp_i += 1 #print('Param epoch %d of %d' % (params_i[exp.axes['avg_over']], exp.shape[exp.axes['avg_over']])) agent_a = agents.SoftmaxAgent( msg_dim=params_v[exp.axes['msg_dim']], hidden_dim=exp.fixed_params['hidden_dim'], # shared_dim=3, color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) agent_b = agents.SoftmaxAgent( msg_dim=params_v[exp.axes['msg_dim']], hidden_dim=exp.fixed_params['hidden_dim'], # shared_dim=3, color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) game = com_game.NoisyChannelGame( reward_func='abs_dist', com_noise=params_v[exp.axes['com_noise']], msg_dim=params_v[exp.axes['msg_dim']], max_epochs=exp.fixed_params['max_epochs'], perception_noise=params_v[exp.axes['perception_noise']], batch_size=exp.fixed_params['batch_size'], print_interval=exp.fixed_params['print_interval'], perception_dim=exp.fixed_params['perception_dim'], loss_type='REINFORCE') game_outcome = exp.run(game.play, env, agent_a, agent_b).result() V = exp.run(evaluate.agent_language_map, env, a=game_outcome).result() exp.set_result('agent_language_map', params_i, V) exp.set_result( 'gibson_cost', params_i, exp.run(game.compute_gibson_cost, env, a=game_outcome).result(1)) exp.set_result( 'regier_cost', params_i, exp.run(evaluate.communication_cost_regier, env, V=V).result()) exp.set_result('wellformedness', params_i, exp.run(evaluate.wellformedness, env, V=V).result()) exp.set_result('term_usage', params_i, exp.run(evaluate.compute_term_usage, V=V).result()) print("\nAll tasks queued to clusters") # wait for all tasks to complete exp.save() exp.wait(retry_interval=5) queue.sync(exp.pipeline_path, exp.pipeline_path, sync_to=sge.SyncTo.LOCAL, recursive=True) return exp
def run(host_name='local', pipeline=''): if pipeline != '': return exp_shared.load_exp(pipeline) # Create and run new experiment queue = exp_shared.create_queue(host_name) queue.sync('.', '.', exclude=['pipelines/*', 'fig/*', 'old/*', 'cogsci/*'], sync_to=sge.SyncTo.REMOTE, recursive=True) exp = Experiment( exp_name='rl_evo_dev', fixed_params=[ ('loss_type', 'REINFORCE'), ('bw_boost', 1), ('env', 'wcs'), ('max_epochs', 50000), # 10000 ('hidden_dim', 20), ('batch_size', 100), ('perception_dim', 3), ('target_dim', 330), ('print_interval', 1000), ('evaluate_interval', 100), ('msg_dim', 15), ('com_noise', 0.1), ('perception_noise', 40) ], #[0, 10, 20, 30, 40, 50, 80, 120, 160, 320] queue=queue) queue.sync(exp.pipeline_path, exp.pipeline_path, sync_to=sge.SyncTo.REMOTE, recursive=True) env = exp.run(com_enviroments.make, exp.fixed_params['env']).result() agent_a = agents.SoftmaxAgent( msg_dim=exp.fixed_params['msg_dim'], hidden_dim=exp.fixed_params['hidden_dim'], color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) agent_b = agents.SoftmaxAgent( msg_dim=exp.fixed_params['msg_dim'], hidden_dim=exp.fixed_params['hidden_dim'], color_dim=exp.fixed_params['target_dim'], perception_dim=exp.fixed_params['perception_dim']) game = com_game.NoisyChannelGame( com_noise=exp.fixed_params['com_noise'], msg_dim=exp.fixed_params['msg_dim'], max_epochs=exp.fixed_params['max_epochs'], perception_noise=exp.fixed_params['perception_noise'], batch_size=exp.fixed_params['batch_size'], print_interval=exp.fixed_params['print_interval'], evaluate_interval=exp.fixed_params['evaluate_interval'], loss_type=exp.fixed_params['loss_type'], bw_boost=exp.fixed_params['bw_boost'], log_path=exp.pipeline_path) game_outcome = exp.run(game.play, env, agent_a, agent_b).result() V = exp.run(evaluate.agent_language_map, env, a=game_outcome).result() # exp.set_result('agent_language_map', V) # exp.set_result('gibson_cost', exp.run(game.compute_gibson_cost, env, a=game_outcome).result(1)) # exp.set_result('regier_cost', exp.run(evaluate.communication_cost_regier, env, V=V).result()) # exp.set_result('wellformedness', exp.run(evaluate.wellformedness, env, V=V).result()) # exp.set_result('term_usage', exp.run(evaluate.compute_term_usage, V=V).result()) exp.save() print("\nAll tasks queued to clusters") # wait for all tasks to complete exp.wait(retry_interval=5) queue.sync(exp.pipeline_path, exp.pipeline_path, sync_to=sge.SyncTo.LOCAL, recursive=True) return exp