def test(args): # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) exp_dir = utils.get_path_from_args( args) if not args.output_dir else args.output_dir path = "{}/results/{}".format(code_root, exp_dir) assert os.path.isdir(path) task_family_test = tasks_sine.RegressionTasksSinusoidal( "test", args.skew_task_distribution) best_valid_model = utils.load_obj(os.path.join(path, "logs")).best_valid_model k_shots = [5, 10, 20, 40] df = [] for k_shot in k_shots: losses = np.array( eval( args, copy.copy(best_valid_model), task_family=task_family_test, num_updates=10, lr_inner=0.01, n_tasks=1000, k_shot=k_shot, )) for grad_step, task_losses in enumerate(losses.T, 1): new_rows = [[k_shot, grad_step, tl] for tl in task_losses] df.extend(new_rows) df = pd.DataFrame(df, columns=["k_shot", "grad_steps", "loss"]) df.to_pickle(os.path.join(path, "res.pkl")) utils.plot_df(df, path)
def run(args, log_interval=5000, rerun=False): # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) exp_dir = utils.get_path_from_args( args) if not args.output_dir else args.output_dir path = "{}/results/{}".format(code_root, exp_dir) if not os.path.isdir(path): os.makedirs(path) if os.path.exists(os.path.join(path, "logs.pkl")) and not rerun: return utils.load_obj(os.path.join(path, "logs")) start_time = time.time() # correctly seed everything utils.set_seed(args.seed) # --- initialise everything --- task_family_train = tasks_sine.RegressionTasksSinusoidal( "train", args.skew_task_distribution) task_family_valid = tasks_sine.RegressionTasksSinusoidal( "valid", args.skew_task_distribution) # initialise network model_inner = MamlModel( task_family_train.num_inputs, task_family_train.num_outputs, n_weights=args.num_hidden_layers, device=args.device, ).to(args.device) model_outer = copy.deepcopy(model_inner) if args.detector == "minimax": task_sampler = TaskSampler( task_family_train.atoms // (2 if args.skew_task_distribution else 1)).to(args.device) elif args.detector == "neyman-pearson": constrainer = Constrainer( task_family_train.atoms // (2 if args.skew_task_distribution else 1)).to(args.device) # intitialise meta-optimiser meta_optimiser = optim.Adam(model_outer.weights + model_outer.biases, args.lr_meta) # initialise loggers logger = Logger() logger.best_valid_model = copy.deepcopy(model_outer) for i_iter in range(args.n_iter): # copy weights of network copy_weights = [w.clone() for w in model_outer.weights] copy_biases = [b.clone() for b in model_outer.biases] # get all shared parameters and initialise cumulative gradient meta_gradient = [ 0 for _ in range( len(copy_weights + copy_biases) + (2 if args.detector != "bayes" else 0)) ] # sample tasks if args.detector == "minimax": task_idxs, task_probs = task_sampler(args.tasks_per_metaupdate) elif args.detector == "neyman-pearson": amplitude_idxs = torch.randint( task_family_train.atoms // (2 if args.skew_task_distribution else 1), (args.tasks_per_metaupdate, ), ) phase_idxs = torch.randint( task_family_train.atoms // (2 if args.skew_task_distribution else 1), (args.tasks_per_metaupdate, ), ) task_idxs = amplitude_idxs, phase_idxs else: task_idxs = None target_functions = task_family_train.sample_tasks( args.tasks_per_metaupdate, task_idxs=task_idxs) for t in range(args.tasks_per_metaupdate): # reset network weights model_inner.weights = [w.clone() for w in copy_weights] model_inner.biases = [b.clone() for b in copy_biases] # get data for current task train_inputs = task_family_train.sample_inputs( args.k_meta_train).to(args.device) for _ in range(args.num_inner_updates): # make prediction using the current model outputs = model_inner(train_inputs) # get targets targets = target_functions[t](train_inputs) # ------------ update on current task ------------ # compute loss for current task loss_task = F.mse_loss(outputs, targets) # compute the gradient wrt current model params = [w for w in model_inner.weights ] + [b for b in model_inner.biases] grads = torch.autograd.grad(loss_task, params, create_graph=True, retain_graph=True) # make an update on the inner model using the current model (to build up computation graph) for i in range(len(model_inner.weights)): if not args.first_order: model_inner.weights[i] = (model_inner.weights[i] - args.lr_inner * grads[i]) else: model_inner.weights[i] = ( model_inner.weights[i] - args.lr_inner * grads[i].detach()) for j in range(len(model_inner.biases)): if not args.first_order: model_inner.biases[j] = ( model_inner.biases[j] - args.lr_inner * grads[i + j + 1]) else: model_inner.biases[j] = ( model_inner.biases[j] - args.lr_inner * grads[i + j + 1].detach()) # ------------ compute meta-gradient on test loss of current task ------------ # get test data test_inputs = task_family_train.sample_inputs(args.k_meta_test).to( args.device) # get outputs after update test_outputs = model_inner(test_inputs) # get the correct targets test_targets = target_functions[t](test_inputs) # compute loss (will backprop through inner loop) if args.detector == "minimax": importance = task_probs[t] else: importance = 1.0 / args.tasks_per_metaupdate loss_meta_raw = F.mse_loss(test_outputs, test_targets) loss_meta = loss_meta_raw * importance if args.detector == "neyman-pearson": amplitude_idxs, phase_idxs = task_idxs aux_loss = constrainer(amplitude_idxs[t], phase_idxs[t], loss_meta_raw) loss_meta = loss_meta + aux_loss # compute gradient w.r.t. *outer model* outer_params = model_outer.weights + model_outer.biases if args.detector == "minimax": outer_params += [ task_sampler.tau_amplitude, task_sampler.tau_phase ] elif args.detector == "neyman-pearson": outer_params += [ constrainer.tau_amplitude, constrainer.tau_phase ] task_grads = torch.autograd.grad( loss_meta, outer_params, retain_graph=(args.detector != "bayes")) for i in range(len(outer_params)): meta_gradient[i] += task_grads[i].detach() # ------------ meta update ------------ meta_optimiser.zero_grad() # print(meta_gradient) # assign meta-gradient for i in range(len(model_outer.weights)): model_outer.weights[i].grad = meta_gradient[i] meta_gradient[i] = 0 for j in range(len(model_outer.biases)): model_outer.biases[j].grad = meta_gradient[i + j + 1] meta_gradient[i + j + 1] = 0 if args.detector == "minimax": task_sampler.tau_amplitude.grad = -meta_gradient[i + j + 2] task_sampler.tau_phase.grad = -meta_gradient[i + j + 3] meta_gradient[i + j + 2] = 0 meta_gradient[i + j + 3] = 0 elif args.detector == "neyman-pearson": constrainer.tau_amplitude.grad = -meta_gradient[i + j + 2] constrainer.tau_phase.grad = -meta_gradient[i + j + 3] meta_gradient[i + j + 2] = 0 meta_gradient[i + j + 3] = 0 # do update step on outer model meta_optimiser.step() # ------------ logging ------------ if i_iter % log_interval == 0: # evaluate on training set losses = eval( args, copy.copy(model_outer), task_family=task_family_train, num_updates=args.num_inner_updates, ) loss_mean, loss_conf = utils.get_stats(np.array(losses)) logger.train_loss.append(loss_mean) logger.train_conf.append(loss_conf) # evaluate on valid set losses = eval( args, copy.copy(model_outer), task_family=task_family_valid, num_updates=args.num_inner_updates, ) loss_mean, loss_conf = utils.get_stats(np.array(losses)) logger.valid_loss.append(loss_mean) logger.valid_conf.append(loss_conf) # save best model if logger.valid_loss[-1] == np.min(logger.valid_loss): print("saving best model at iter", i_iter) logger.best_valid_model = copy.copy(model_outer) # save logging results utils.save_obj(logger, os.path.join(path, "logs")) # print current results logger.print_info(i_iter, start_time) start_time = time.time() return logger
def run(args, log_interval=5000, rerun=False): global temp assert not args.maml # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) if not os.path.isdir('{}/{}_result_files/'.format(code_root, args.task)): os.mkdir('{}/{}_result_files/'.format(code_root, args.task)) path = '{}/{}_result_files/'.format( code_root, args.task) + utils.get_path_from_args(args) if os.path.exists(path + '.pkl') and not rerun: return utils.load_obj(path) start_time = time.time() utils.set_seed(args.seed) # --- initialise everything --- # get the task family if args.task == 'sine': task_family_train = tasks_sine.RegressionTasksSinusoidal() task_family_valid = tasks_sine.RegressionTasksSinusoidal() task_family_test = tasks_sine.RegressionTasksSinusoidal() elif args.task == 'celeba': task_family_train = tasks_celebA.CelebADataset('train', device=args.device) task_family_valid = tasks_celebA.CelebADataset('valid', device=args.device) task_family_test = tasks_celebA.CelebADataset('test', device=args.device) elif args.task == 'multi': task_family_train = multi() task_family_valid = multi() task_family_test = multi() else: raise NotImplementedError # initialise network model = CaviaModel(n_in=task_family_train.num_inputs, n_out=task_family_train.num_outputs, num_context_params=args.num_context_params, n_hidden=args.num_hidden_layers, device=args.device).to(args.device) # intitialise meta-optimiser # (only on shared params - context parameters are *not* registered parameters of the model) meta_optimiser = optim.Adam(model.parameters(), args.lr_meta) encoder = pool_encoder().to(args.device) encoder_optimiser = optim.Adam(encoder.parameters(), lr=1e-3) decoder = pool_decoder().to(args.device) decoder_optimiser = optim.Adam(decoder.parameters(), lr=1e-3) #encoder.load_state_dict(torch.load('./model/encoder')) p_encoder = place().to(args.device) p_optimiser = optim.Adam(p_encoder.parameters(), lr=1e-3) # initialise loggers logger = Logger() logger.best_valid_model = copy.deepcopy(model) # --- main training loop --- for i_iter in range(args.n_iter): # initialise meta-gradient meta_gradient = [0 for _ in range(len(model.state_dict()))] place_gradient = [0 for _ in range(len(p_encoder.state_dict()))] encoder_gradient = [0 for _ in range(len(encoder.state_dict()))] #print(meta_gradient) # sample tasks target_functions, ty = task_family_train.sample_tasks( args.tasks_per_metaupdate, True) # --- inner loop --- for t in range(args.tasks_per_metaupdate): # reset private network weights model.reset_context_params() # get data for current task x = task_family_train.sample_inputs( args.k_meta_train, args.use_ordered_pixels).to(args.device) y = target_functions[t](x) train_inputs = torch.cat([x, y], dim=1) a = encoder(train_inputs) #embedding,_ = torch.max(a,dim=0) embedding = torch.mean(a, dim=0) logits = p_encoder(embedding) logits = logits.reshape([latent_dim, categorical_dim]) y = gumbel_softmax(logits, temp, hard=True) y = y[:, 1] #print(temp) #model.set_context_params(embedding) #print(model.context_params) for _ in range(args.num_inner_updates): # forward through model train_outputs = model(x) # get targets train_targets = target_functions[t](x) # ------------ update on current task ------------ # compute loss for current task task_loss = F.mse_loss(train_outputs, train_targets) # compute gradient wrt context params task_gradients = \ torch.autograd.grad(task_loss, model.context_params, create_graph=not args.first_order)[0] # update context params (this will set up the computation graph correctly) model.context_params = model.context_params - args.lr_inner * task_gradients * y #print(model.context_params) # ------------ compute meta-gradient on test loss of current task ------------ # get test data test_inputs = task_family_train.sample_inputs( args.k_meta_test, args.use_ordered_pixels).to(args.device) # get outputs after update test_outputs = model(test_inputs) # get the correct targets test_targets = target_functions[t](test_inputs) # compute loss after updating context (will backprop through inner loop) loss_meta = F.mse_loss(test_outputs, test_targets) #print(torch.norm(y,1)/1000) #loss_meta += torch.norm(y,1)/700 qy = F.softmax(logits, dim=-1) log_ratio = torch.log(qy * categorical_dim + 1e-20) KLD = torch.sum(qy * log_ratio, dim=-1).mean() / 5 # print(KLD) loss_meta += KLD # compute gradient + save for current task task_grad = torch.autograd.grad(loss_meta, model.parameters(), retain_graph=True) for i in range(len(task_grad)): # clip the gradient meta_gradient[i] += task_grad[i].detach().clamp_(-10, 10) task_grad_place = torch.autograd.grad(loss_meta, p_encoder.parameters(), retain_graph=True) for i in range(len(task_grad_place)): # clip the gradient place_gradient[i] += task_grad_place[i].detach().clamp_( -10, 10) task_grad_encoder = torch.autograd.grad(loss_meta, encoder.parameters()) for i in range(len(task_grad_encoder)): # clip the gradient encoder_gradient[i] += task_grad_encoder[i].detach().clamp_( -10, 10) # ------------ meta update ------------ # assign meta-gradient for i, param in enumerate(model.parameters()): param.grad = meta_gradient[i] / args.tasks_per_metaupdate meta_optimiser.step() # do update step on shared model for i, param in enumerate(p_encoder.parameters()): param.grad = place_gradient[i] / args.tasks_per_metaupdate p_optimiser.step() for i, param in enumerate(encoder.parameters()): param.grad = encoder_gradient[i] / args.tasks_per_metaupdate encoder_optimiser.step() # reset context params model.reset_context_params() if i_iter % 350 == 1: temp = np.maximum(temp * np.exp(-ANNEAL_RATE * i_iter), 0.5) print(temp) # ------------ logging ------------ if i_iter % log_interval == 0: # evaluate on training set loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_train, num_updates=args.num_inner_updates, encoder=encoder, p_encoder=p_encoder) logger.train_loss.append(loss_mean) logger.train_conf.append(loss_conf) # evaluate on test set loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_valid, num_updates=args.num_inner_updates, encoder=encoder, p_encoder=p_encoder) logger.valid_loss.append(loss_mean) logger.valid_conf.append(loss_conf) # evaluate on validation set if i_iter % log_interval == 0: loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_test, num_updates=args.num_inner_updates, encoder=encoder, p_encoder=p_encoder) logger.test_loss.append(loss_mean) logger.test_conf.append(loss_conf) # save logging results utils.save_obj(logger, path) # save best model if logger.valid_loss[-1] == np.min(logger.valid_loss): print('saving best model at iter', i_iter) logger.best_valid_model = copy.deepcopy(model) logger.best_encoder_valid_model = copy.deepcopy(encoder) logger.best_place_valid_model = copy.deepcopy(p_encoder) if i_iter % (4 * log_interval) == 0: print('saving model at iter', i_iter) logger.valid_model.append(copy.deepcopy(model)) logger.encoder_valid_model.append(copy.deepcopy(encoder)) logger.place_valid_model.append(copy.deepcopy(p_encoder)) # visualise results if args.task == 'celeba': task_family_train.visualise( task_family_train, task_family_test, copy.deepcopy(logger.best_valid_model), args, i_iter) # print current results logger.print_info(i_iter, start_time) start_time = time.time() return logger
def run(args, log_interval=5000, rerun=False): assert args.maml # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) if not os.path.isdir('{}/{}_result_files/'.format(code_root, args.task)): os.mkdir('{}/{}_result_files/'.format(code_root, args.task)) path = '{}/{}_result_files/'.format(code_root, args.task) + utils.get_path_from_args(args) if os.path.exists(path + '.pkl') and not rerun: return utils.load_obj(path) start_time = time.time() # correctly seed everything utils.set_seed(args.seed) # --- initialise everything --- # get the task family if args.task == 'sine': task_family_train = tasks_sine.RegressionTasksSinusoidal() task_family_valid = tasks_sine.RegressionTasksSinusoidal() task_family_test = tasks_sine.RegressionTasksSinusoidal() elif args.task == 'celeba': task_family_train = tasks_celebA.CelebADataset('train', args.device) task_family_valid = tasks_celebA.CelebADataset('valid', args.device) task_family_test = tasks_celebA.CelebADataset('test', args.device) else: raise NotImplementedError #initialize transformer transformer = FCNet(task_family_train.num_inputs, 3, 128, 128).to(args.device) # initialise network model_inner = MamlModel(128, task_family_train.num_outputs, n_weights=args.num_hidden_layers, num_context_params=args.num_context_params, device=args.device ).to(args.device) model_outer = copy.deepcopy(model_inner) print("MAML: ", model_outer) print("Transformer: ", transformer) # intitialise meta-optimiser meta_optimiser = optim.Adam(model_outer.weights + model_outer.biases + [model_outer.task_context], args.lr_meta) opt_transformer = torch.optim.Adam(transformer.parameters(), 0.01) # initialise loggers logger = Logger() logger.best_valid_model = copy.deepcopy(model_outer) for i_iter in range(args.n_iter): #meta_train_error = 0.0 # copy weights of network copy_weights = [w.clone() for w in model_outer.weights] copy_biases = [b.clone() for b in model_outer.biases] copy_context = model_outer.task_context.clone() # get all shared parameters and initialise cumulative gradient meta_gradient = [0 for _ in range(len(copy_weights + copy_biases) + 1)] # sample tasks target_functions = task_family_train.sample_tasks(args.tasks_per_metaupdate) for t in range(args.tasks_per_metaupdate): #gradient initialization for transformer acc_grads = fsn.phi_gradients(transformer, args.device) # reset network weights model_inner.weights = [w.clone() for w in copy_weights] model_inner.biases = [b.clone() for b in copy_biases] model_inner.task_context = copy_context.clone() # get data for current task train_inputs = task_family_train.sample_inputs(args.k_meta_train, args.use_ordered_pixels).to(args.device) # get test data test_inputs = task_family_train.sample_inputs(args.k_meta_test, args.use_ordered_pixels).to(args.device) transformed_train_inputs = transformer(train_inputs)#.to(args.device) transformed_test_inputs = transformer(test_inputs)#.to(args.device) # transformer task loss # with torch.no_grad(): targets0 = target_functions[t](train_inputs) L0 = F.mse_loss(model_inner(transformed_train_inputs), targets0) targets1 = target_functions[t](test_inputs) L1 = F.mse_loss(model_inner(transformed_test_inputs), targets1) trans_loss = fsn.cosine_loss(L0, L1, model_inner, args.device) #trans_loss = evaluation_error + trans_loss for step in range(args.num_inner_updates): # print("iteration:" , i_iter, "innerstep: ", step) outputs = model_inner(transformed_train_inputs) # make prediction using the current model #outputs = model_inner(train_inputs) # get targets targets = target_functions[t](train_inputs) # ------------ update on current task ------------ # compute loss for current task loss_task = F.mse_loss(outputs, targets) # compute the gradient wrt current model params = [w for w in model_inner.weights] + [b for b in model_inner.biases] + [model_inner.task_context] grads = torch.autograd.grad(loss_task, params, create_graph=True, retain_graph=True) # make an update on the inner model using the current model (to build up computation graph) for i in range(len(model_inner.weights)): if not args.first_order: model_inner.weights[i] = model_inner.weights[i] - args.lr_inner * grads[i].clamp_(-10, 10) else: model_inner.weights[i] = model_inner.weights[i] - args.lr_inner * grads[i].detach().clamp_(-10, 10) for j in range(len(model_inner.biases)): if not args.first_order: model_inner.biases[j] = model_inner.biases[j] - args.lr_inner * grads[i + j + 1].clamp_(-10, 10) else: model_inner.biases[j] = model_inner.biases[j] - args.lr_inner * grads[i + j + 1].detach().clamp_(-10, 10) if not args.first_order: model_inner.task_context = model_inner.task_context - args.lr_inner * grads[i + j + 2].clamp_(-10, 10) else: model_inner.task_context = model_inner.task_context - args.lr_inner * grads[i + j + 2].detach().clamp_(-10, 10) # ------------ compute meta-gradient on test loss of current task ------------ # get outputs after update test_outputs = model_inner(transformed_test_inputs) # get the correct targets test_targets = target_functions[t](test_inputs) # compute loss (will backprop through inner loop) loss_meta = F.mse_loss(test_outputs, test_targets) #meta_train_error += loss_meta.item() # transformer gradients trans_loss = loss_meta grads_phi = list(torch.autograd.grad(trans_loss, transformer.parameters(), retain_graph=True, create_graph=True)) for p, l in zip(acc_grads, grads_phi): l = l p.data = torch.add(p, (1 / args.tasks_per_metaupdate), l.detach().clamp_(-10,10)) # compute gradient w.r.t. *outer model* task_grads = torch.autograd.grad(loss_meta, model_outer.weights + model_outer.biases + [model_outer.task_context]) for i in range(len(model_inner.weights + model_inner.biases) + 1): meta_gradient[i] += task_grads[i].detach().clamp_(-10, 10) # ------------ meta update ------------ opt_transformer.zero_grad() meta_optimiser.zero_grad() # parameter gradient attributes of transformer updated for k, p in zip(transformer.parameters(), acc_grads): k.grad = p # print(meta_gradient) # assign meta-gradient for i in range(len(model_outer.weights)): model_outer.weights[i].grad = meta_gradient[i] / args.tasks_per_metaupdate meta_gradient[i] = 0 for j in range(len(model_outer.biases)): model_outer.biases[j].grad = meta_gradient[i + j + 1] / args.tasks_per_metaupdate meta_gradient[i + j + 1] = 0 model_outer.task_context.grad = meta_gradient[i + j + 2] / args.tasks_per_metaupdate meta_gradient[i + j + 2] = 0 # do update step on outer model meta_optimiser.step() opt_transformer.step() # ------------ logging ------------ if i_iter % log_interval == 0:# and i_iter > 0: # evaluate on training set loss_mean, loss_conf = eval(args, copy.copy(model_outer), task_family=task_family_train, num_updates=args.num_inner_updates, transformer=transformer) logger.train_loss.append(loss_mean) logger.train_conf.append(loss_conf) # evaluate on test set loss_mean, loss_conf = eval(args, copy.copy(model_outer), task_family=task_family_valid, num_updates=args.num_inner_updates, transformer=transformer) logger.valid_loss.append(loss_mean) logger.valid_conf.append(loss_conf) # evaluate on validation set loss_mean, loss_conf = eval(args, copy.copy(model_outer), task_family=task_family_test, num_updates=args.num_inner_updates, transformer=transformer) logger.test_loss.append(loss_mean) logger.test_conf.append(loss_conf) # save logging results utils.save_obj(logger, path) # save best model if logger.valid_loss[-1] == np.min(logger.valid_loss): print('saving best model at iter', i_iter) logger.best_valid_model = copy.copy(model_outer) # visualise results if args.task == 'celeba': task_family_train.visualise(task_family_train, task_family_test, copy.copy(logger.best_valid_model), args, i_iter, transformer) # print current results logger.print_info(i_iter, start_time) start_time = time.time() return logger
def run(args, log_interval=5000, rerun=False): assert not args.maml # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) if not os.path.isdir('{}/{}_result_files/'.format(code_root, args.task)): os.mkdir('{}/{}_result_files/'.format(code_root, args.task)) path = '{}/{}_result_files/'.format( code_root, args.task) + utils.get_path_from_args(args) if os.path.exists(path + '.pkl') and not rerun: return utils.load_obj(path) start_time = time.time() utils.set_seed(args.seed) # --- initialise everything --- # get the task family if args.task == 'sine': task_family_train = tasks_sine.RegressionTasksSinusoidal() task_family_valid = tasks_sine.RegressionTasksSinusoidal() task_family_test = tasks_sine.RegressionTasksSinusoidal() elif args.task == 'celeba': task_family_train = tasks_celebA.CelebADataset('train', device=args.device) task_family_valid = tasks_celebA.CelebADataset('valid', device=args.device) task_family_test = tasks_celebA.CelebADataset('test', device=args.device) else: raise NotImplementedError # initialise network model = CaviaModel(n_in=task_family_train.num_inputs, n_out=task_family_train.num_outputs, num_context_params=args.num_context_params, n_hidden=args.num_hidden_layers, device=args.device).to(args.device) # intitialise meta-optimiser # (only on shared params - context parameters are *not* registered parameters of the model) meta_optimiser = optim.Adam(model.parameters(), args.lr_meta) # initialise loggers logger = Logger() logger.best_valid_model = copy.deepcopy(model) # --- main training loop --- for i_iter in range(args.n_iter): # initialise meta-gradient meta_gradient = [0 for _ in range(len(model.state_dict()))] # sample tasks target_functions = task_family_train.sample_tasks( args.tasks_per_metaupdate) # --- inner loop --- for t in range(args.tasks_per_metaupdate): # reset private network weights model.reset_context_params() # get data for current task train_inputs = task_family_train.sample_inputs( args.k_meta_train, args.use_ordered_pixels).to(args.device) for _ in range(args.num_inner_updates): # forward through model train_outputs = model(train_inputs) # get targets train_targets = target_functions[t](train_inputs) # ------------ update on current task ------------ # compute loss for current task task_loss = F.mse_loss(train_outputs, train_targets) # compute gradient wrt context params task_gradients = \ torch.autograd.grad(task_loss, model.context_params, create_graph=not args.first_order)[0] # update context params (this will set up the computation graph correctly) model.context_params = model.context_params - args.lr_inner * task_gradients # ------------ compute meta-gradient on test loss of current task ------------ # get test data test_inputs = task_family_train.sample_inputs( args.k_meta_test, args.use_ordered_pixels).to(args.device) # get outputs after update test_outputs = model(test_inputs) # get the correct targets test_targets = target_functions[t](test_inputs) # compute loss after updating context (will backprop through inner loop) loss_meta = F.mse_loss(test_outputs, test_targets) # compute gradient + save for current task task_grad = torch.autograd.grad(loss_meta, model.parameters()) for i in range(len(task_grad)): # clip the gradient meta_gradient[i] += task_grad[i].detach().clamp_(-10, 10) # ------------ meta update ------------ # assign meta-gradient for i, param in enumerate(model.parameters()): param.grad = meta_gradient[i] / args.tasks_per_metaupdate # do update step on shared model meta_optimiser.step() # reset context params model.reset_context_params() # ------------ logging ------------ if i_iter % log_interval == 0: # evaluate on training set loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_train, num_updates=args.num_inner_updates) logger.train_loss.append(loss_mean) logger.train_conf.append(loss_conf) # evaluate on test set loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_valid, num_updates=args.num_inner_updates) logger.valid_loss.append(loss_mean) logger.valid_conf.append(loss_conf) # evaluate on validation set loss_mean, loss_conf = eval_cavia( args, copy.deepcopy(model), task_family=task_family_test, num_updates=args.num_inner_updates) logger.test_loss.append(loss_mean) logger.test_conf.append(loss_conf) # save logging results utils.save_obj(logger, path) # save best model if logger.valid_loss[-1] == np.min(logger.valid_loss): print('saving best model at iter', i_iter) logger.best_valid_model = copy.deepcopy(model) # visualise results if args.task == 'celeba': task_family_train.visualise( task_family_train, task_family_test, copy.deepcopy(logger.best_valid_model), args, i_iter) # print current results logger.print_info(i_iter, start_time) start_time = time.time() return logger
def run(args, log_interval=5000, rerun=False): assert args.maml # see if we already ran this experiment code_root = os.path.dirname(os.path.realpath(__file__)) if not os.path.isdir('{}/{}_result_files/'.format(code_root, args.task)): os.mkdir('{}/{}_result_files/'.format(code_root, args.task)) path = '{}/{}_result_files/'.format( code_root, args.task) + utils.get_path_from_args(args) if os.path.exists(path + '.pkl') and not rerun: return utils.load_obj(path) start_time = time.time() # correctly seed everything utils.set_seed(args.seed) # --- initialise everything --- # get the task family if args.task == 'sine': task_family_train = tasks_sine.RegressionTasksSinusoidal() task_family_valid = tasks_sine.RegressionTasksSinusoidal() task_family_test = tasks_sine.RegressionTasksSinusoidal() elif args.task == 'celeba': task_family_train = tasks_celebA.CelebADataset('train') task_family_valid = tasks_celebA.CelebADataset('valid') task_family_test = tasks_celebA.CelebADataset('test') else: raise NotImplementedError # initialise network model_inner = MamlModel(task_family_train.num_inputs, task_family_train.num_outputs, n_weights=args.num_hidden_layers, num_context_params=args.num_context_params, device=args.device).to(args.device) model_outer = copy.deepcopy(model_inner) # intitialise meta-optimiser meta_optimiser = optim.Adam( model_outer.weights + model_outer.biases + [model_outer.task_context], args.lr_meta) # initialise loggers logger = Logger() logger.best_valid_model = copy.deepcopy(model_outer) for i_iter in range(args.n_iter): # copy weights of network copy_weights = [w.clone() for w in model_outer.weights] copy_biases = [b.clone() for b in model_outer.biases] copy_context = model_outer.task_context.clone() # get all shared parameters and initialise cumulative gradient meta_gradient = [0 for _ in range(len(copy_weights + copy_biases) + 1)] # sample tasks target_functions = task_family_train.sample_tasks( args.tasks_per_metaupdate) for t in range(args.tasks_per_metaupdate): # reset network weights model_inner.weights = [w.clone() for w in copy_weights] model_inner.biases = [b.clone() for b in copy_biases] model_inner.task_context = copy_context.clone() # get data for current task train_inputs = task_family_train.sample_inputs( args.k_meta_train, args.use_ordered_pixels).to(args.device) for _ in range(args.num_inner_updates): # forward through network outputs = model_outer(train_inputs) # get targets targets = target_functions[t](train_inputs) # ------------ update on current task ------------ # compute loss for current task loss_task = F.mse_loss(outputs, targets) # update private parts of network and keep correct computation graph params = [w for w in model_outer.weights] + [ b for b in model_outer.biases ] + [model_outer.task_context] grads = torch.autograd.grad(loss_task, params, create_graph=True, retain_graph=True) for i in range(len(model_inner.weights)): if not args.first_order: model_inner.weights[i] = model_outer.weights[ i] - args.lr_inner * grads[i] else: model_inner.weights[i] = model_outer.weights[ i] - args.lr_inner * grads[i].detach() for j in range(len(model_inner.biases)): if not args.first_order: model_inner.biases[j] = model_outer.biases[ j] - args.lr_inner * grads[i + j + 1] else: model_inner.biases[j] = model_outer.biases[ j] - args.lr_inner * grads[i + j + 1].detach() if not args.first_order: model_inner.task_context = model_outer.task_context - args.lr_inner * grads[ i + j + 2] else: model_inner.task_context = model_outer.task_context - args.lr_inner * grads[ i + j + 2].detach() # ------------ compute meta-gradient on test loss of current task ------------ # get test data test_inputs = task_family_train.sample_inputs( args.k_meta_test, args.use_ordered_pixels).to(args.device) # get outputs after update test_outputs = model_inner(test_inputs) # get the correct targets test_targets = target_functions[t](test_inputs) # compute loss (will backprop through inner loop) loss_meta = F.mse_loss(test_outputs, test_targets) # compute gradient w.r.t. *outer model* task_grads = torch.autograd.grad( loss_meta, model_outer.weights + model_outer.biases + [model_outer.task_context]) for i in range(len(model_inner.weights + model_inner.biases) + 1): meta_gradient[i] += task_grads[i].detach() # ------------ meta update ------------ meta_optimiser.zero_grad() # print(meta_gradient) # assign meta-gradient for i in range(len(model_outer.weights)): model_outer.weights[ i].grad = meta_gradient[i] / args.tasks_per_metaupdate meta_gradient[i] = 0 for j in range(len(model_outer.biases)): model_outer.biases[j].grad = meta_gradient[ i + j + 1] / args.tasks_per_metaupdate meta_gradient[i + j + 1] = 0 model_outer.task_context.grad = meta_gradient[ i + j + 2] / args.tasks_per_metaupdate meta_gradient[i + j + 2] = 0 # do update step on outer model meta_optimiser.step() # ------------ logging ------------ if i_iter % log_interval == 0: # evaluate on training set loss_mean, loss_conf = eval(args, copy.deepcopy(model_outer), task_family=task_family_train, num_updates=args.num_inner_updates) logger.train_loss.append(loss_mean) logger.train_conf.append(loss_conf) # evaluate on test set loss_mean, loss_conf = eval(args, copy.deepcopy(model_outer), task_family=task_family_valid, num_updates=args.num_inner_updates) logger.valid_loss.append(loss_mean) logger.valid_conf.append(loss_conf) # evaluate on validation set loss_mean, loss_conf = eval(args, copy.deepcopy(model_outer), task_family=task_family_test, num_updates=args.num_inner_updates) logger.test_loss.append(loss_mean) logger.test_conf.append(loss_conf) # save logging results utils.save_obj(logger, path) # save best model if logger.valid_loss[-1] == np.min(logger.valid_loss): print('saving best model at iter', i_iter) logger.best_valid_model = copy.deepcopy(model_outer) # visualise results if args.task == 'celeba': tasks_celebA.visualise(task_family_train, task_family_test, copy.deepcopy(logger.best_valid_model), args, i_iter) # print current results logger.print_info(i_iter, start_time) start_time = time.time() return logger