def forward(self, img, z, mcr):
'''
Steps:
- Modify image if using mcr
- D1 + imgs -> d1_real
- G + noise -> gen_imgs
- D1 + gen_imgs -> d1_fake
- Adv (D2) + gen_imgs
Return D outputs and gen_imgs
'''
print('MCR in forward', mcr)
if mcr: ### Multi-channel rescaling. Add extra channel for real images. Generated images are rescaled inside generator
linear_scale = 1 / self.linear_scaler
ch2 = lbann.Tanh(
lbann.WeightedSum(self.inv_transform(lbann.Identity(img)),
scaling_factors=str(linear_scale)))
y = lbann.Concatenation(lbann.Identity(img), ch2, axis=0)
img = lbann.Reshape(y, dims='2 128 128')
else:
img = lbann.Reshape(img, dims='1 128 128')
d1_real = self.forward_discriminator1(img) #instance1
gen_img = self.forward_generator(z, mcr=mcr)
d1_fake = self.forward_discriminator1(
lbann.StopGradient(gen_img)) #instance2
d_adv = self.forward_discriminator2(
gen_img) #instance 3 //need to freeze
#d1s share weights, d1_w is copied to d_adv (through replace weight callback) and freeze
return d1_real, d1_fake, d_adv, gen_img, img
def forward(self, img, z):
d1_real = self.forward_discriminator1(img) #instance1
gen_img = self.forward_generator(z, self.g_ps)
d1_fake = self.forward_discriminator1(
lbann.StopGradient(gen_img, name='stop_gradient')) #instance2
d_adv = self.forward_discriminator2(
gen_img) #instance 3 //need to freeze
return d1_real, d1_fake, d_adv, gen_img
def forward(self, img, z):
#description
d1_real = self.forward_discriminator1(img) #instance1
gen_img = self.forward_generator(z)
d1_fake = self.forward_discriminator1(lbann.StopGradient(gen_img)) #instance2
d_adv = self.forward_discriminator2(gen_img) #instance 3 //need to freeze
#d1s share weights, d1_w is copied to d_adv (through replace weight callback) and freeze
return d1_real, d1_fake, d_adv,gen_img
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 forward(
self,
motif_indices,
motif_size,
walk_indices,
walk_length,
):
# Apply generator
fake_motif_indices, gen_prob, gen_log_prob = self.generator(
walk_length,
walk_indices,
motif_size,
)
# Get discriminator embeddings in log-space
all_motif_indices = lbann.Concatenation(motif_indices,
fake_motif_indices)
all_motif_log_embeddings = self.discriminator.get_log_embeddings(
all_motif_indices)
all_motif_log_embeddings = lbann.Slice(
all_motif_log_embeddings,
slice_points=str_list([0, motif_size, 2 * motif_size]),
)
real_motif_log_embeddings = lbann.Identity(all_motif_log_embeddings)
fake_motif_log_embeddings = lbann.Identity(all_motif_log_embeddings)
# Apply discriminator
real_disc_prob, real_disc_log_not_prob \
= self.discriminator(motif_size, real_motif_log_embeddings)
fake_disc_prob, fake_disc_log_not_prob \
= self.discriminator(motif_size, fake_motif_log_embeddings)
# Loss function
# L_disc = - log(D(real)) - log(1-D(fake))
# L_gen = - log(G) * stop_gradient(log(1-D(fake)))
real_disc_log_prob \
= lbann.Log(lbann.Clamp(real_disc_prob, min=1e-37, max=1))
disc_loss = lbann.WeightedSum(
real_disc_log_prob,
fake_disc_log_not_prob,
scaling_factors=str_list([-1, -1]),
)
gen_loss = lbann.Multiply(
gen_log_prob,
lbann.StopGradient(fake_disc_log_not_prob),
)
loss = lbann.Add(disc_loss, gen_loss)
return loss, real_disc_prob, fake_disc_prob, gen_prob
def forward(self, z, y):
z_sample = self.encoder(y)
y_recon = self.decoder(z_sample)
#d real/fake share weights, shared weights is copied to d_adv
#(through replace weight callback) and freeze
d_real = self.discriminator0(lbann.Concatenation([y, z], axis=0))
y_z_sample = lbann.Concatenation([y, z_sample], axis=0)
d_fake = self.discriminator0(lbann.StopGradient(y_z_sample))
d_adv = self.discriminator1(y_z_sample) #freeze
return d_real, d_fake, d_adv, y_recon
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)
