def forward_encoder(self, x_emb): """Encoder step, emulating z ~ E(x) = q_E(z|x) :param x_emb: (n_batch, len(x), d_z) of floats, embeddings for input sentence x :return: (n_batch, d_z) of floats, sample of latent vector z :return: float, kl term component of loss """ # _, h = self.encoder_rnn(x, None) h = self.encoder_rnn(x_emb, None) h = lbann.Slice( h, slice_points=str_list( [self.input_feature_dims - 1, self.input_feature_dims]), axis=0, ) h = lbann.Identity(h) mu, logvar = self.q_mu(h), self.q_logvar(h) # Set datatype of previous layers # Note: Depth-first search from mu and logvar to x_emb stack = [mu, logvar] in_stack = {l: True for l in stack} while stack: l = stack.pop() if type(l) not in (lbann.Slice, lbann.Reshape, lbann.Tessellate): l.datatype = self.datatype for parent in l.parents: if parent not in in_stack and parent is not x_emb: stack.append(parent) in_stack[parent] = True # eps = torch.randn_like(mu) eps = lbann.Gaussian(mean=0, stdev=1, hint_layer=mu) # z = mu + (logvar / 2).exp() * eps z = lbann.Add([ mu, (lbann.Multiply([ lbann.Exp(lbann.WeightedSum(logvar, scaling_factors='0.5')), eps ])) ]) # kl_loss = 0.5 * (logvar.exp() + mu ** 2 - 1 - logvar).sum(1).mean() kl_loss = lbann.Reduction( lbann.WeightedSum( lbann.Exp(logvar), lbann.Square(mu), self.constant(1, hint_layer=mu), logvar, scaling_factors='0.5 0.5 -0.5 -0.5', ), mode='sum', ) return z, kl_loss
def forward(self, x, z): """Do the WAE forward step :param x: list of tensors of longs, embed representation of input :return: float, kl term component of loss :return: float, recon component of loss """ x = lbann.Slice(x, slice_points=str_list([0, self.input_feature_dims])) x = lbann.Identity(x) x_emb = lbann.Embedding(x, num_embeddings=self.dictionary_size, embedding_dim=self.embedding_size, name='emb', weights=self.emb_weights) # Encoder: x -> z, kl_loss z_sample = self.forward_encoder(x_emb) eps = lbann.Gaussian(mean=self.gmean, stdev=self.gstd, hint_layer=z_sample) z_sample = lbann.Add([z_sample, eps]) # Decoder: x, z -> recon_loss #pred = self.forward_decoder(x_emb, z_sample) pred, arg_max = self.forward_decoder(x_emb, z_sample) recon_loss = self.compute_loss(x, pred) # Hack to remove blocking GPU allreduce in evaluation layer #kl_loss = lbann.Identity(kl_loss, device='CPU') recon_loss = lbann.Identity(recon_loss, device='CPU') z_prior = lbann.Tessellate( lbann.Reshape(z, dims=str_list([1, self.zdim])), dims=str_list([self.input_feature_dims, self.zdim]), ) d_real = self.discriminator0( lbann.Concatenation([x_emb, z_prior], axis=1)) z_sample0 = lbann.Tessellate( lbann.Reshape(z_sample, dims=str_list([1, self.zdim])), dims=str_list([self.input_feature_dims, self.zdim]), ) y_z_sample = lbann.Concatenation([x_emb, z_sample0], axis=1) d_fake = self.discriminator0(lbann.StopGradient(y_z_sample)) d_adv = self.discriminator1(y_z_sample) #freeze return recon_loss, d_real, d_fake, d_adv, arg_max
def construct_model(num_epochs,mcr,spectral_loss,save_batch_interval): """Construct LBANN model. """ import lbann # Layer graph input = lbann.Input(target_mode='N/A',name='inp_img') ### Create expected labels for real and fake data (with label flipping = 0.01) prob_flip=0.01 label_flip_rand = lbann.Uniform(min=0,max=1, neuron_dims='1') label_flip_prob = lbann.Constant(value=prob_flip, num_neurons='1') ones = lbann.GreaterEqual(label_flip_rand,label_flip_prob, name='is_real') zeros = lbann.LogicalNot(ones,name='is_fake') gen_ones=lbann.Constant(value=1.0,num_neurons='1')## All ones: no flip. Input for training Generator. #============================================== ### Implement GAN ##Create the noise vector z = lbann.Reshape(lbann.Gaussian(mean=0.0,stdev=1.0, neuron_dims="64", name='noise_vec'),dims='1 64') ## Creating the GAN object and implementing forward pass for both networks ### d1_real, d1_fake, d_adv, gen_img, img = ExaGAN.CosmoGAN(mcr)(input,z,mcr) #============================================== ### Compute quantities for adding to Loss and Metrics d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real,ones],name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake,zeros],name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv,gen_ones],name='d_adv_bce') #img_loss = lbann.MeanSquaredError([gen_img,img]) #l1_loss = lbann.L1Norm(lbann.WeightedSum([gen_img,img], scaling_factors="1 -1")) #============================================== ### Set up source and destination layers layers = list(lbann.traverse_layer_graph(input)) weights = set() src_layers,dst_layers = [],[] for l in layers: if(l.weights and "disc1" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2, analogous to discrim.trainable=False in Keras if(l.weights and "disc2" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) #l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) #============================================== ### Define Loss and Metrics #Define loss (Objective function) loss_list=[d1_real_bce,d1_fake_bce,d_adv_bce] ## Usual GAN loss function # loss_list=[d1_real_bce,d1_fake_bce] ## skipping adversarial loss for G for testing spectral loss if spectral_loss: dft_gen_img = lbann.DFTAbs(gen_img) dft_img = lbann.StopGradient(lbann.DFTAbs(img)) spec_loss = lbann.Log(lbann.MeanSquaredError(dft_gen_img, dft_img)) loss_list.append(lbann.LayerTerm(spec_loss, scale=8.0)) loss = lbann.ObjectiveFunction(loss_list) #Define metrics metrics = [lbann.Metric(d1_real_bce,name='d_real'),lbann.Metric(d1_fake_bce, name='d_fake'), lbann.Metric(d_adv_bce,name='gen_adv')] if spectral_loss: metrics.append(lbann.Metric(spec_loss,name='spec_loss')) #============================================== ### Define callbacks list callbacks_list=[] dump_outputs=True save_model=False print_model=False callbacks_list.append(lbann.CallbackPrint()) callbacks_list.append(lbann.CallbackTimer()) callbacks_list.append(lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers),batch_interval=1)) if dump_outputs: #callbacks_list.append(lbann.CallbackDumpOutputs(layers='inp_img gen_img_instance1_activation', execution_modes='train validation', directory='dump_outs',batch_interval=save_batch_interval,format='npy')) callbacks_list.append(lbann.CallbackDumpOutputs(layers='gen_img_instance1_activation', execution_modes='train validation', directory='dump_outs',batch_interval=save_batch_interval,format='npy')) if save_model : callbacks_list.append(lbann.CallbackSaveModel(dir='models')) if print_model: callbacks_list.append(lbann.CallbackPrintModelDescription()) ### Construct model return lbann.Model(num_epochs, weights=weights, layers=layers, metrics=metrics, objective_function=loss, callbacks=callbacks_list)
def construct_model(run_args): """Construct LBANN model. Initial model for ATOM molecular VAE """ import lbann print("Dump model dir ", run_args.dump_model_dir) assert run_args.dump_model_dir, "evaluate script asssumes a pretrained WAE model" pad_index = run_args.pad_index assert pad_index is not None sequence_length = run_args.sequence_length assert sequence_length is not None print("sequence length is {}".format(sequence_length)) data_layout = "data_parallel" # Layer graph input_ = lbann.Identity(lbann.Input(name='inp', data_field='samples'), name='inp1') wae_loss = [] input_feature_dims = sequence_length embedding_size = run_args.embedding_dim dictionary_size = run_args.num_embeddings assert embedding_size is not None assert dictionary_size is not None save_output = True if run_args.dump_outputs_dir else False print("save output? ", save_output, "out dir ", run_args.dump_outputs_dir) z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims=run_args.z_dim) waemodel = molwae.MolWAE(input_feature_dims, dictionary_size, embedding_size, pad_index, run_args.z_dim, save_output) recon, d1_real, d1_fake, d_adv, arg_max = waemodel(input_, z) zero = lbann.Constant(value=0.0, num_neurons='1', name='zero') one = lbann.Constant(value=1.0, num_neurons='1', name='one') d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real, one], name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake, zero], name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv, one], name='d_adv_bce') wae_loss.append(recon) layers = list(lbann.traverse_layer_graph(input_)) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if (l.weights and "disc0" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2 if (l.weights and "disc1" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) l2_weights = [ w for w in weights if not isinstance(w.optimizer, lbann.NoOptimizer) ] l2_reg = lbann.L2WeightRegularization(weights=l2_weights, scale=1e-4) wae_loss.append(d1_real_bce) wae_loss.append(d_adv_bce) wae_loss.append(d1_fake_bce) wae_loss.append(l2_reg) print("LEN wae loss ", len(wae_loss)) obj = lbann.ObjectiveFunction(wae_loss) # Initialize check metric callback metrics = [ lbann.Metric(d_adv_bce, name='adv_loss'), lbann.Metric(recon, name='recon') ] callbacks = [ lbann.CallbackPrint(), #lbann.CallbackStepLearningRate(step=10, amt=0.5), lbann.CallbackTimer() ] callbacks.append( lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers), batch_interval=2)) #Dump output (activation) for post processing if (run_args.dump_outputs_dir): pred_tensor = lbann.Concatenation(arg_max, name='pred_tensor') callbacks.append( lbann.CallbackDumpOutputs( batch_interval=run_args.dump_outputs_interval, execution_modes='test', directory=run_args.dump_outputs_dir, layers=f'inp pred_tensor {waemodel.q_mu.name}')) # Construct model return lbann.Model(run_args.num_epochs, weights=weights, layers=layers, objective_function=obj, metrics=metrics, callbacks=callbacks)
def construct_model(): """Construct LBANN model. ExaGAN model """ import lbann # Layer graph input = lbann.Input(target_mode='N/A',name='inp_img') #label flipping label_flip_rand = lbann.Uniform(min=0,max=1, neuron_dims='1') label_flip_prob = lbann.Constant(value=0.01, num_neurons='1') one = lbann.GreaterEqual(label_flip_rand,label_flip_prob, name='is_real') zero = lbann.LogicalNot(one,name='is_fake') z = lbann.Reshape(lbann.Gaussian(mean=0.0,stdev=1.0, neuron_dims="64", name='noise_vec'),dims='1 64') d1_real, d1_fake, d_adv, gen_img = ExaGAN.CosmoGAN()(input,z) d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real,one],name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake,zero],name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv,one],name='d_adv_bce') layers = list(lbann.traverse_layer_graph(input)) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if(l.weights and "disc1" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2, analogous to discrim.trainable=False in Keras if(l.weights and "disc2" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) #l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) obj = lbann.ObjectiveFunction([d1_real_bce,d1_fake_bce,d_adv_bce]) # Initialize check metric callback metrics = [lbann.Metric(d1_real_bce,name='d_real'), lbann.Metric(d1_fake_bce, name='d_fake'), lbann.Metric(d_adv_bce,name='gen')] callbacks = [lbann.CallbackPrint(), lbann.CallbackTimer(), #Uncomment to dump output for plotting and further statistical analysis #lbann.CallbackDumpOutputs(layers='inp_img gen_img_instance1_activation', # execution_modes='train validation', # directory='dump_outs', # batch_interval=100, # format='npy'), lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers), batch_interval=2)] # Construct model num_epochs = 20 return lbann.Model(num_epochs, weights=weights, layers=layers, metrics=metrics, objective_function=obj, callbacks=callbacks)
def construct_model(run_args): """Construct LBANN model. Initial model for ATOM molecular VAE """ import lbann print("Dump model dir ", run_args.dump_model_dir) assert run_args.dump_model_dir, "evaluate script asssumes a pretrained WAE model" pad_index = run_args.pad_index assert pad_index is not None sequence_length = run_args.sequence_length assert sequence_length is not None print("sequence length is {}".format(sequence_length)) data_layout = "data_parallel" # Layer graph input_ = lbann.Identity(lbann.Input(name='inp', target_mode="N/A"), name='inp1') wae_loss = [] input_feature_dims = sequence_length embedding_size = run_args.embedding_dim dictionary_size = run_args.num_embeddings assert embedding_size is not None assert dictionary_size is not None save_output = False print("save output? ", save_output, "out dir ", run_args.dump_outputs_dir) z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="128") x = lbann.Slice(input_, slice_points=str_list([0, input_feature_dims])) x = lbann.Identity(x) waemodel = molwae.MolWAE(input_feature_dims, dictionary_size, embedding_size, pad_index, save_output) x_emb = lbann.Embedding(x, num_embeddings=waemodel.dictionary_size, embedding_dim=waemodel.embedding_size, name='emb', weights=waemodel.emb_weights) latentz = waemodel.forward_encoder(x_emb) fake_loss = lbann.MeanAbsoluteError(latentz, z) layers = list(lbann.traverse_layer_graph(input_)) # Setup objective function weights = set() for l in layers: weights.update(l.weights) obj = lbann.ObjectiveFunction(fake_loss) callbacks = [lbann.CallbackPrint(), lbann.CallbackTimer()] #Dump output (activation) for post processing conc_out = lbann.Concatenation([input_, latentz], name='conc_out') callbacks.append( lbann.CallbackDumpOutputs( batch_interval=run_args.dump_outputs_interval, execution_modes='test', directory=run_args.dump_outputs_dir, layers=f'{conc_out.name}')) # Construct model return lbann.Model(run_args.num_epochs, weights=weights, layers=layers, objective_function=obj, callbacks=callbacks)
sd = lbann.Exp(logsd, name="sd") var = lbann.Square(sd, name="var") meansq = lbann.Square(mu, name="meansq") kldiv_plus_half = lbann.WeightedSum([meansq, var, logsd], name="kldiv_plus_half", scaling_factors='0.5 0.5 -1') kldiv_full = lbann.Rsqrt(kldiv_plus_half, name="kldiv_full") kldiv = lbann.Reduction(kldiv_full, name="kldiv", mode="sum") # Generate sample noise = lbann.Gaussian(name="noise", mean=0, stdev=1, hint_layer=mu) sdnoise = lbann.Hadamard([noise, sd], name="sdnoise") sample = lbann.Add([mu, sdnoise], name="sample") # Decoder decode4 = lbann.FullyConnected(sample, name="decode4", has_bias=True, hint_layer=encode3) decode4neuron = lbann.Relu(decode4, name="decode4neuron") decode3 = lbann.FullyConnected(decode4neuron, name="decode3",
def construct_macc_surrogate_model(xdim, ydim, zdim, wae_mcf, surrogate_mcf, lambda_cyc, useCNN, dump_models, pretrained_dir, ltfb_batch_interval, num_epochs): """Construct MACC surrogate model. See https://arxiv.org/pdf/1912.08113.pdf model architecture and other details """ # Layer graph input = lbann.Input(data_field='samples', name='inp_data') # data is 64*64*4 images + 15 scalar + 5 param inp_slice = lbann.Slice(input, axis=0, slice_points=str_list([0, ydim, ydim + xdim]), name='inp_slice') gt_y = lbann.Identity(inp_slice, name='gt_y') gt_x = lbann.Identity(inp_slice, name='gt_x') #param not used zero = lbann.Constant(value=0.0, num_neurons='1', name='zero') one = lbann.Constant(value=1.0, num_neurons='1', name='one') z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="20") wae = macc_network_architectures.MACCWAE( zdim, ydim, cf=wae_mcf, use_CNN=useCNN) #pretrained, freeze inv = macc_network_architectures.MACCInverse(xdim, cf=surrogate_mcf) fwd = macc_network_architectures.MACCForward(zdim, cf=surrogate_mcf) y_pred_fwd = wae.encoder(gt_y) param_pred_ = wae.encoder(gt_y) input_fake = inv(param_pred_) output_cyc = fwd(input_fake) y_image_re2 = wae.decoder(output_cyc) '''**** Train cycleGAN input params <--> latent space of (images, scalars) ****''' output_fake = fwd(gt_x) y_image_re = wae.decoder(output_fake) param_pred2_ = wae.encoder(y_image_re) input_cyc = inv(param_pred2_) L_l2_x = lbann.MeanSquaredError(input_fake, gt_x) L_cyc_x = lbann.MeanSquaredError(input_cyc, gt_x) L_l2_y = lbann.MeanSquaredError(output_fake, y_pred_fwd) L_cyc_y = lbann.MeanSquaredError(output_cyc, y_pred_fwd) #@todo slice here to separate scalar from image img_sca_loss = lbann.MeanSquaredError(y_image_re, gt_y) #L_cyc = L_cyc_y + L_cyc_x L_cyc = lbann.Add(L_cyc_y, L_cyc_x) #loss_gen0 = L_l2_y + lamda_cyc*L_cyc loss_gen0 = lbann.WeightedSum([L_l2_y, L_cyc], scaling_factors=f'1 {lambda_cyc}') loss_gen1 = lbann.WeightedSum([L_l2_x, L_cyc_y], scaling_factors=f'1 {lambda_cyc}') #loss_gen1 = L_l2_x + lamda_cyc*L_cyc_y layers = list(lbann.traverse_layer_graph(input)) weights = set() #Freeze appropriate (pretrained) weights pretrained_models = ["wae"] #add macc? for l in layers: for idx in range(len(pretrained_models)): if (l.weights and pretrained_models[idx] in l.name): for w in range(len(l.weights)): l.weights[w].optimizer = lbann.NoOptimizer() weights.update(l.weights) l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) #d_adv_bce = lbann.LayerTerm(d_adv_bce,scale=0.01) # Setup objective function obj = lbann.ObjectiveFunction([loss_gen0, loss_gen1, l2_reg]) # Initialize check metric callback metrics = [ lbann.Metric(img_sca_loss, name='fw_loss'), lbann.Metric(L_l2_x, name='inverse loss'), lbann.Metric(L_cyc_y, name='output cycle loss'), lbann.Metric(L_cyc_x, name='param cycle loss') ] callbacks = [ lbann.CallbackPrint(), lbann.CallbackSaveModel(dir=dump_models), lbann.CallbackLoadModel(dirs=str(pretrained_dir)), lbann.CallbackTimer() ] if (ltfb_batch_interval > 0): callbacks.append( lbann.CallbackLTFB(batch_interval=ltfb_batch_interval, metric='fw_loss', low_score_wins=True, exchange_hyperparameters=True)) # Construct model return lbann.Model(num_epochs, weights=weights, layers=layers, metrics=metrics, objective_function=obj, callbacks=callbacks)
def construct_jag_wae_model(ydim, zdim, mcf, useCNN, dump_models, ltfb_batch_interval, num_epochs): """Construct LBANN model. JAG Wasserstein autoencoder model """ # Layer graph input = lbann.Input(data_field='samples', name='inp_data') # data is 64*64*4 images + 15 scalar + 5 param #inp_slice = lbann.Slice(input, axis=0, slice_points="0 16399 16404",name='inp_slice') inp_slice = lbann.Slice(input, axis=0, slice_points=str_list([0, ydim, ydim + 5]), name='inp_slice') gt_y = lbann.Identity(inp_slice, name='gt_y') gt_x = lbann.Identity(inp_slice, name='gt_x') #param not used zero = lbann.Constant(value=0.0, num_neurons='1', name='zero') one = lbann.Constant(value=1.0, num_neurons='1', name='one') z_dim = 20 #Latent space dim z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="20") model = macc_network_architectures.MACCWAE(zdim, ydim, cf=mcf, use_CNN=useCNN) d1_real, d1_fake, d_adv, pred_y = model(z, gt_y) d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real, one], name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake, zero], name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv, one], name='d_adv_bce') img_loss = lbann.MeanSquaredError([pred_y, gt_y]) rec_error = lbann.L2Norm2( lbann.WeightedSum([pred_y, gt_y], scaling_factors="1 -1")) layers = list(lbann.traverse_layer_graph(input)) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if (l.weights and "disc0" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2 if (l.weights and "disc1" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) d_adv_bce = lbann.LayerTerm(d_adv_bce, scale=0.01) obj = lbann.ObjectiveFunction( [d1_real_bce, d1_fake_bce, d_adv_bce, img_loss, rec_error, l2_reg]) # Initialize check metric callback metrics = [lbann.Metric(img_loss, name='recon_error')] #pred_y = macc_models.MACCWAE.pred_y_name callbacks = [ lbann.CallbackPrint(), lbann.CallbackTimer(), lbann.CallbackPrintModelDescription(), lbann.CallbackSaveModel(dir=dump_models), lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers), batch_interval=2) ] if (ltfb_batch_interval > 0): callbacks.append( lbann.CallbackLTFB(batch_interval=ltfb_batch_interval, metric='recon_error', low_score_wins=True, exchange_hyperparameters=True)) # Construct model return lbann.Model(num_epochs, weights=weights, layers=layers, metrics=metrics, objective_function=obj, callbacks=callbacks)
def construct_model(): """Construct LBANN model. JAG Wasserstein autoencoder model """ import lbann # Layer graph input = lbann.Input(target_mode='N/A',name='inp_data') # data is 64*64*4 images + 15 scalar + 5 param inp_slice = lbann.Slice(input, axis=0, slice_points="0 16399 16404",name='inp_slice') gt_y = lbann.Identity(inp_slice,name='gt_y') gt_x = lbann.Identity(inp_slice, name='gt_x') #param not used zero = lbann.Constant(value=0.0,num_neurons='1',name='zero') one = lbann.Constant(value=1.0,num_neurons='1',name='one') y_dim = 16399 #image+scalar shape z_dim = 20 #Latent space dim z = lbann.Gaussian(mean=0.0,stdev=1.0, neuron_dims="20") d1_real, d1_fake, d_adv, pred_y = jag_models.WAE(z_dim,y_dim)(z,gt_y) d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real,one],name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake,zero],name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv,one],name='d_adv_bce') img_loss = lbann.MeanSquaredError([pred_y,gt_y]) rec_error = lbann.L2Norm2(lbann.WeightedSum([pred_y,gt_y], scaling_factors="1 -1")) layers = list(lbann.traverse_layer_graph(input)) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if(l.weights and "disc0" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2 if(l.weights and "disc1" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) d_adv_bce = lbann.LayerTerm(d_adv_bce,scale=0.01) obj = lbann.ObjectiveFunction([d1_real_bce,d1_fake_bce,d_adv_bce,img_loss,rec_error,l2_reg]) # Initialize check metric callback metrics = [lbann.Metric(img_loss, name='recon_error')] callbacks = [lbann.CallbackPrint(), lbann.CallbackTimer(), lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers), batch_interval=2)] # Construct model num_epochs = 100 return lbann.Model(num_epochs, weights=weights, layers=layers, metrics=metrics, objective_function=obj, callbacks=callbacks)
def construct_model(args): """Construct LBANN for CosmoGAN 3D model. """ obj = [] metrics = [] callbacks = [] w = [args.input_width]*3 w.insert(0,args.input_channel) _sample_dims = w ps = None #have model and input ps if(args.use_distconv): ps = get_parallel_strategy_args( sample_groups=args.mini_batch_size, depth_groups=args.depth_groups, height_groups=args.height_groups, ) g_device = 'GPU' input_ = lbann.Input(name='input', data_field='samples') input_ = lbann.Reshape(input_, dims=list2str(_sample_dims),name='in_reshape', device=g_device), x1 = lbann.Identity(input_, parallel_strategy=None, name='x1') x2 = lbann.Identity(input_, name='x2') if args.compute_mse else None zero = lbann.Constant(value=0.0,num_neurons='1',name='zero',device=g_device) one = lbann.Constant(value=1.0,num_neurons='1',name='one', device=g_device) z = lbann.Reshape(lbann.Gaussian(mean=0.0,stdev=1.0, neuron_dims="64", name='noise_vec', device=g_device), dims='1 64', name='noise_vec_reshape',device=g_device) print("RUN ARGS ", args) d1_real,d1_fake,d_adv, gen_img = model.Exa3DGAN(args.input_width,args.input_channel, g_device,ps,use_bn=args.use_bn)(x1,z) layers=list(lbann.traverse_layer_graph([d1_real, d1_fake])) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if(l.weights and "disc1" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2, analogous to discrim.trainable=False in Keras if(l.weights and "disc2" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real,one],name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake,zero],name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv,one],name='d_adv_bce') mse = lbann.MeanSquaredError([gen_img, x2], name='MSE') if args.compute_mse else None obj.append(d1_real_bce) obj.append(d1_fake_bce) obj.append(d_adv_bce) metrics.append(lbann.Metric(d_adv_bce, name='d_adv_bce')) metrics.append(lbann.Metric(d1_real_bce, name='d1_real_bce')) metrics.append(lbann.Metric(d1_fake_bce, name='d1_fake_bce')) if (mse is not None): obj.append(mse) metrics.append(lbann.Metric(mse, name='MSE')) callbacks.append(lbann.CallbackPrint()) callbacks.append(lbann.CallbackTimer()) callbacks.append(lbann.CallbackGPUMemoryUsage()) # ------------------------------------------ # Construct model # ------------------------------------------ return lbann.Model(args.num_epochs, weights=weights, layers=layers, objective_function=obj, metrics=metrics, callbacks=callbacks)
def construct_model(): """Construct MACC surrogate model. See https://arxiv.org/pdf/1912.08113.pdf model architecture and other details """ import lbann # Layer graph input = lbann.Input(data_field='samples', name='inp_data') # data is 64*64*4 images + 15 scalar + 5 param inp_slice = lbann.Slice(input, axis=0, slice_points=str_list( [0, args.ydim, args.ydim + args.xdim]), name='inp_slice') gt_y = lbann.Identity(inp_slice, name='gt_y') gt_x = lbann.Identity(inp_slice, name='gt_x') #param not used zero = lbann.Constant(value=0.0, num_neurons='1', name='zero') one = lbann.Constant(value=1.0, num_neurons='1', name='one') z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="20") wae = macc_models.MACCWAE(args.zdim, args.ydim, cf=args.wae_mcf, use_CNN=args.useCNN) #pretrained, freeze inv = macc_models.MACCInverse(args.xdim, cf=args.surrogate_mcf) fwd = macc_models.MACCForward(args.zdim, cf=args.surrogate_mcf) y_pred_fwd = wae.encoder(gt_y) param_pred_ = wae.encoder(gt_y) input_fake = inv(param_pred_) output_cyc = fwd(input_fake) y_image_re2 = wae.decoder(output_cyc) '''**** Train cycleGAN input params <--> latent space of (images, scalars) ****''' output_fake = fwd(gt_x) y_image_re = wae.decoder(output_fake) y_out = wae.decoder(y_pred_fwd) param_pred2_ = wae.encoder(y_image_re) input_cyc = inv(param_pred2_) L_l2_x = lbann.MeanSquaredError( input_fake, gt_x) #(x,inv(enc(y)), (encoder+)inverse loss L_cyc_x = lbann.MeanSquaredError( input_cyc, gt_x) #param, x cycle loss, from latent space L_l2_y = lbann.MeanSquaredError( output_fake, y_pred_fwd) #pred error into latent space (enc(y),fw(x)) L_cyc_y = lbann.MeanSquaredError( output_cyc, y_pred_fwd) # pred error into latent space (enc(y), fw(inv(enc(y)))) #@todo slice here to separate scalar from image img_sca_loss = lbann.MeanSquaredError( y_image_re, gt_y) # (y,dec(fw(x))) #forward model to decoder, no latent space dec_fw_inv_enc_y = lbann.MeanSquaredError( y_image_re2, gt_y) #(y, dec(fw(inv(enc(y))))) y->enc_z->x'->fw_z->y' wae_loss = lbann.MeanSquaredError(y_out, gt_y) #(y, dec(enc(y)) ' #L_cyc = L_cyc_y + L_cyc_x L_cyc = lbann.Add(L_cyc_y, L_cyc_x) #loss_gen0 = L_l2_y + lamda_cyc*L_cyc loss_gen0 = lbann.WeightedSum([L_l2_y, L_cyc], scaling_factors=f'1 {args.lamda_cyc}') loss_gen1 = lbann.WeightedSum([L_l2_x, L_cyc_y], scaling_factors=f'1 {args.lamda_cyc}') #loss_gen1 = L_l2_x + lamda_cyc*L_cyc_y conc_out = lbann.Concatenation( [gt_x, wae_loss, img_sca_loss, dec_fw_inv_enc_y, L_l2_x], name='x_errors') layers = list(lbann.traverse_layer_graph(input)) weights = set() for l in layers: weights.update(l.weights) # Setup objective function obj = lbann.ObjectiveFunction([loss_gen0, loss_gen1]) # Initialize check metric callback metrics = [ lbann.Metric(img_sca_loss, name='img_re1'), lbann.Metric(dec_fw_inv_enc_y, name='img_re2'), lbann.Metric(wae_loss, name='wae_loss'), lbann.Metric(L_l2_x, name='inverse loss'), lbann.Metric(L_cyc_y, name='output cycle loss'), lbann.Metric(L_cyc_x, name='param cycle loss') ] callbacks = [ lbann.CallbackPrint(), lbann.CallbackDumpOutputs(layers=f'{conc_out.name}', execution_modes='test', directory=args.dump_outputs, batch_interval=1, format='npy'), lbann.CallbackTimer() ] # Construct model num_epochs = 1 return lbann.Model(num_epochs, weights=weights, layers=layers, serialize_io=True, metrics=metrics, objective_function=obj, callbacks=callbacks)
def construct_model(num_epochs, mcr, save_batch_interval=82): """Construct LBANN model. """ import lbann # Layer graph input = lbann.Input(target_mode='N/A', name='inp_img') #============================================== ##Create the noise vector z = lbann.Reshape(lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="64", name='noise_vec'), dims='1 64') ## Creating the GAN object and implementing forward pass for both networks ### gen_img = model_GAN.CosmoGAN(mcr)(input, z, mcr) # #============================================== # ### Set up source and destination layers # layers = list(lbann.traverse_layer_graph(input)) # weights = set() # src_layers,dst_layers = [],[] # for l in layers: # if(l.weights and "disc1" in l.name and "instance1" in l.name): # src_layers.append(l.name) # #freeze weights in disc2, analogous to discrim.trainable=False in Keras # if(l.weights and "disc2" in l.name): # dst_layers.append(l.name) # for idx in range(len(l.weights)): # l.weights[idx].optimizer = lbann.NoOptimizer() # weights.update(l.weights) # #============================================== # ### Define Loss and Metrics # #Define loss (Objective function) # loss_list=[d1_real_bce,d1_fake_bce,d_adv_bce] ## Usual GAN loss function # # loss_list.append(l2_reg) # loss = lbann.ObjectiveFunction(loss_list) # #Define metrics # metrics = [lbann.Metric(d1_real_bce,name='d_real'),lbann.Metric(d1_fake_bce, name='d_fake'), lbann.Metric(d_adv_bce,name='gen'), # #lbann.Metric(img_loss, name='msq_error') ,lbann.Metric(l1_loss, name='l1norm_error') # # ,lbann.Metric(l2_reg) # ] #============================================== ### Define callbacks list callbacks_list = [] print_model = False fname = '' callbacks_list.append(lbann.CallbackPrint()) callbacks_list.append(lbann.CallbackTimer()) callbacks_list.append(lbann.CallbackLoadModel(dirs=str(fname))) if print_model: callbacks_list.append(lbann.CallbackPrintModelDescription()) ### Construct model return lbann.Model( num_epochs, # weights=weights, # layers=layers, # metrics=metrics, # objective_function=loss, callbacks=callbacks_list)
def construct_model(run_args): """Construct LBANN model. Initial model for ATOM molecular VAE """ import lbann pad_index = run_args.pad_index assert pad_index is not None sequence_length = run_args.sequence_length assert sequence_length is not None print("sequence length is {}".format(sequence_length)) data_layout = "data_parallel" # Layer graph input_ = lbann.Identity(lbann.Input(name='inp', target_mode="N/A"), name='inp1') vae_loss = [] input_feature_dims = sequence_length embedding_size = run_args.embedding_dim dictionary_size = run_args.num_embeddings assert embedding_size is not None assert dictionary_size is not None save_output = True if run_args.dump_outputs_dir else False print("save output? ", save_output, "out dir ", run_args.dump_outputs_dir) z = lbann.Gaussian(mean=0.0, stdev=1.0, neuron_dims="128") recon, d1_real, d1_fake, d_adv, arg_max = molwae.MolWAE( input_feature_dims, dictionary_size, embedding_size, pad_index, save_output)(input_, z) zero = lbann.Constant(value=0.0, num_neurons='1', name='zero') one = lbann.Constant(value=1.0, num_neurons='1', name='one') d1_real_bce = lbann.SigmoidBinaryCrossEntropy([d1_real, one], name='d1_real_bce') d1_fake_bce = lbann.SigmoidBinaryCrossEntropy([d1_fake, zero], name='d1_fake_bce') d_adv_bce = lbann.SigmoidBinaryCrossEntropy([d_adv, one], name='d_adv_bce') vae_loss.append(recon) layers = list(lbann.traverse_layer_graph(input_)) # Setup objective function weights = set() src_layers = [] dst_layers = [] for l in layers: if (l.weights and "disc0" in l.name and "instance1" in l.name): src_layers.append(l.name) #freeze weights in disc2 if (l.weights and "disc1" in l.name): dst_layers.append(l.name) for idx in range(len(l.weights)): l.weights[idx].optimizer = lbann.NoOptimizer() weights.update(l.weights) l2_reg = lbann.L2WeightRegularization(weights=weights, scale=1e-4) vae_loss.append(d1_real_bce) vae_loss.append(d_adv_bce) vae_loss.append(d1_fake_bce) vae_loss.append(l2_reg) print("LEN vae loss ", len(vae_loss)) obj = lbann.ObjectiveFunction(vae_loss) # Initialize check metric callback metrics = [ lbann.Metric(d_adv_bce, name='adv_loss'), lbann.Metric(recon, name='recon') ] callbacks = [ lbann.CallbackPrint(), #lbann.CallbackStepLearningRate(step=10, amt=0.5), lbann.CallbackTimer() ] if (run_args.dump_weights_interval > 0): callbacks.append( lbann.CallbackDumpWeights( directory=run_args.dump_weights_dir, epoch_interval=run_args.dump_weights_interval)) if (run_args.ltfb): send_name = ('' if run_args.weights_to_send == 'All' else run_args.weights_to_send) #hack for Merlin empty string weights_to_ex = [w.name for w in weights if send_name in w.name] print("LTFB Weights to exchange ", weights_to_ex) callbacks.append( lbann.CallbackLTFB(batch_interval=run_args.ltfb_batch_interval, metric='recon', weights=list2str(weights_to_ex), low_score_wins=True, exchange_hyperparameters=True)) callbacks.append( lbann.CallbackReplaceWeights(source_layers=list2str(src_layers), destination_layers=list2str(dst_layers), batch_interval=2)) #Dump final weight for inference if (run_args.dump_model_dir): callbacks.append(lbann.CallbackSaveModel(dir=run_args.dump_model_dir)) #Dump output (activation) for post processing if (run_args.dump_outputs_dir): pred_tensor = lbann.Concatenation(arg_max, name='pred_tensor') callbacks.append( lbann.CallbackDumpOutputs( batch_interval=run_args.dump_outputs_interval, execution_modes='test', directory=run_args.dump_outputs_dir, layers='inp pred_tensor')) if (run_args.warmup): callbacks.append( lbann.CallbackLinearGrowthLearningRate(target=run_args.lr / 512 * run_args.batch_size, num_epochs=5)) # Construct model return lbann.Model(run_args.num_epochs, weights=weights, layers=layers, objective_function=obj, metrics=metrics, callbacks=callbacks)