def run_PCA(self, num_dim, percent_variability):
# get covariance matrix (uses compact trick)
N = self.data_matrix.shape[0]
data_matrix_2d = self.data_matrix.reshape(
self.data_matrix.shape[0],
-1).T # flatten data instances and transpose
mean = np.mean(data_matrix_2d, axis=1)
centered_data_matrix_2d = (data_matrix_2d.T - mean).T
trick_cov_matrix = np.dot(
centered_data_matrix_2d.T,
centered_data_matrix_2d) * 1.0 / np.sqrt(N - 1)
# get eignevectors and eigenvalues
eigen_values, eigen_vectors = np.linalg.eigh(trick_cov_matrix)
eigen_vectors = np.dot(centered_data_matrix_2d, eigen_vectors)
for i in range(N):
eigen_vectors[:, i] = eigen_vectors[:, i] / np.linalg.norm(
eigen_vectors[:, i])
eigen_values = np.flip(eigen_values)
eigen_vectors = np.flip(eigen_vectors, 1)
# get num PCA components
cumDst = np.cumsum(eigen_values) / np.sum(eigen_values)
if num_dim == 0:
cumDst = np.cumsum(eigen_values) / np.sum(eigen_values)
num_dim = np.where(cumDst > float(percent_variability))[0][0] + 1
W = eigen_vectors[:, :num_dim]
PCA_scores = np.matmul(centered_data_matrix_2d.T, W)
sw_message(f"The PCA modes of particles being retained : {num_dim}")
sw_message(f"Variablity preserved: {str(float(cumDst[num_dim-1]))}")
self.num_dim = num_dim
self.PCA_scores = PCA_scores
self.eigen_vectors = eigen_vectors
self.eigen_values = eigen_values
return num_dim
def visualizeAugmentation(data_csv, viz_type='splom', show=True):
if viz_type == 'splom':
Visualize.splom(data_csv)
elif viz_type == 'violin':
Visualize.violin(data_csv, show)
else:
sw_message("Error visualization type unrecognized.")
def fit(self, embedded_matrix, mixture_num):
sw_message("Fitting Gaussian mixture model...")
self.embedded_matrix = embedded_matrix
if mixture_num == 0:
mixture_num = self.selectClusterNum()
self.GMM = GaussianMixture(mixture_num, covariance_type='full', random_state=0)
self.GMM.fit(self.embedded_matrix)
sw_message("Gaussian mixture model converged: " + str(self.GMM.converged_))
def selectClusterNum(self):
n_components = np.arange(1, self.embedded_matrix.shape[1])
models = [GaussianMixture(n, covariance_type='full', random_state=0).fit(self.embedded_matrix) for n in n_components]
bic_min_index = np.argmin(np.array([m.bic(self.embedded_matrix) for m in models]))
aic_min_index = np.argmin(np.array([m.aic(self.embedded_matrix) for m in models]))
avg_index = int((bic_min_index + aic_min_index) / 2)
mixture_num = n_components[avg_index]
sw_message("Using " + str(mixture_num) + " components.")
return mixture_num
def get_test_loader(loader_dir, test_img_list, down_factor=1, down_dir=None):
sw_message("Creating test torch loader:")
# get data
image_paths = []
scores = []
models = []
test_names = []
for index in range(len(test_img_list)):
image_path = test_img_list[index]
# add name
prefix = get_prefix(image_path)
test_names.append(prefix)
image_paths.append(image_path)
# add label placeholders
scores.append([1])
models.append([1])
images = get_images(loader_dir, image_paths, down_factor, down_dir)
test_data = DeepSSMdataset(images, scores, models)
# Write test names to file so they are saved somewhere
name_file = open(loader_dir + 'test_names.txt', 'w+')
name_file.write(str(test_names))
name_file.close()
sw_message("Test names saved to: " + loader_dir + "test_names.txt")
# Make loader
sw_message("Creating and saving test dataloader...")
testloader = DataLoader(test_data,
batch_size=1,
shuffle=False,
num_workers=8,
pin_memory=torch.cuda.is_available())
test_path = loader_dir + 'test'
torch.save(testloader, test_path)
sw_message("Test loader complete.\n")
return test_path, test_names
def fit(self, embedded_matrix):
sw_message("Fitting KDE...")
self.embedded_matrix = embedded_matrix
# get sigma squared
nearest_neighbor_dists = []
cov = np.cov(embedded_matrix.T)
for i in embedded_matrix:
smallest = np.Inf
for j in embedded_matrix:
dist = Mdist(i,j,cov)
if dist < smallest and dist != 0:
smallest = dist
nearest_neighbor_dists.append(smallest)
self.sigma_squared = np.mean(np.array(nearest_neighbor_dists))/embedded_matrix.shape[1]
def compute_pvalues_for_group_difference_data(group_0_data,
group_1_data,
permutations=100):
number_of_particles = group_0_data.shape[0]
group_0_size = group_0_data.shape[-1]
group_1_size = group_1_data.shape[-1]
subset_size = min(group_0_size, group_1_size)
pvalues_matrix = np.zeros((number_of_particles, permutations))
idx = 0
for p in range(permutations):
sw_progress(float(idx) / permutations)
idx = idx + 1
if sw_check_abort():
sw_message("Aborted")
return
group_0_index = np.random.choice(group_0_size,
subset_size,
replace=False)
group_1_index = np.random.choice(group_1_size,
subset_size,
replace=False)
group_0_subset = group_0_data[:, :, group_0_index]
group_1_subset = group_1_data[:, :, group_1_index]
for particle_id in range(number_of_particles):
x = group_0_subset[particle_id, :, :].T
y = group_1_subset[particle_id, :, :].T
stats, f, pv, s = hotelling_t2(x, y)
pvalues_matrix[particle_id, p] = pv
corrected_pvalue_matrix = np.zeros((number_of_particles, 1))
for particle_id in range(number_of_particles):
r, pval = multi.fdrcorrection(pvalues_matrix[particle_id, :],
alpha=0.05)
corrected_pvalue_matrix[particle_id, 0] = np.mean(pval)
return corrected_pvalue_matrix
def get_all_train_data(loader_dir, data_csv, down_factor, down_dir):
# get all data and targets
sw_message("Reading all data...")
image_paths = []
scores = []
models = []
prefixes = []
with open(data_csv, newline='') as csvfile:
datareader = csv.reader(csvfile)
index = 0
for row in datareader:
image_path = row[0]
model_path = row[1]
pca_scores = row[2:]
# add name
prefix = get_prefix(image_path)
# data error check
if prefix not in get_prefix(model_path):
print("Error: Images and models mismatched in csv.")
print(index)
print(prefix)
print(get_prefix(model_path))
exit()
prefixes.append(prefix)
# add image path
image_paths.append(image_path)
# add score (un-normalized)
pca_scores = [float(i) for i in pca_scores]
scores.append(pca_scores)
# add model
mdl = get_particles(model_path)
models.append(mdl)
index += 1
images = get_images(loader_dir, image_paths, down_factor, down_dir)
scores = whiten_PCA_scores(scores, loader_dir)
return images, scores, models, prefixes
def get_train_val_loaders(loader_dir,
data_csv,
batch_size=1,
down_factor=1,
down_dir=None,
train_split=0.80):
sw_message("Creating training and validation torch loaders:")
if not os.path.exists(loader_dir):
os.makedirs(loader_dir)
images, scores, models, prefixes = get_all_train_data(
loader_dir, data_csv, down_factor, down_dir)
images, scores, models, prefixes = shuffle_data(images, scores, models,
prefixes)
# split into train and validation (e.g. 80% vs 20%)
cut = int(len(images) * train_split)
sw_message("Turning to tensors...")
train_data = DeepSSMdataset(images[:cut], scores[:cut], models[:cut])
sw_message(str(len(train_data)) + ' in training set')
val_data = DeepSSMdataset(images[cut:], scores[cut:], models[cut:])
sw_message(str(len(val_data)) + ' in validation set')
sw_message("Saving data loaders...")
trainloader = DataLoader(train_data,
batch_size=batch_size,
shuffle=True,
num_workers=8,
pin_memory=torch.cuda.is_available())
train_path = loader_dir + 'train'
torch.save(trainloader, train_path)
validationloader = DataLoader(val_data,
batch_size=1,
shuffle=True,
num_workers=8,
pin_memory=torch.cuda.is_available())
val_path = loader_dir + 'validation'
torch.save(validationloader, val_path)
sw_message("Training and validation loaders complete.\n")
return train_path, val_path
def shuffle_data(images, scores, models, prefixes):
sw_message("Shuffling.")
c = list(zip(images, scores, models, prefixes))
random.shuffle(c)
images, scores, models, prefixes = zip(*c)
return images, scores, models, prefixes
def test(config_file):
with open(config_file) as json_file:
parameters = json.load(json_file)
model_dir = parameters["paths"]["out_dir"] + parameters["model_name"] + '/'
pred_dir = model_dir + 'predictions/'
if not os.path.exists(pred_dir):
os.makedirs(pred_dir)
if parameters["use_best_model"]:
model_path = model_dir + 'best_model.torch'
else:
model_path = model_dir + 'final_model.torch'
if parameters["fine_tune"]["enabled"]:
model_path_ft = model_path.replace(".torch", "_ft.torch")
else:
model_path_ft = model_path
loader_dir = parameters["paths"]["loader_dir"]
# load the loaders
sw_message("Loading test data loader...")
test_loader = torch.load(loader_dir + "test")
print("Done.\n")
# initalizations
sw_message("Loading trained model...")
model_pca = model.DeepSSMNet(config_file)
model_pca.load_state_dict(torch.load(model_path))
device = model_pca.device
model_pca.to(device)
model_pca.eval()
model_ft = model.DeepSSMNet(config_file)
model_ft.load_state_dict(torch.load(model_path_ft))
model_ft.to(device)
model_ft.eval()
# Get test names
test_names_file = loader_dir + 'test_names.txt'
f = open(test_names_file, 'r')
test_names_string = f.read()
f.close()
test_names_string = test_names_string.replace("[", "").replace(
"]", "").replace("'", "").replace(" ", "")
test_names = test_names_string.split(",")
print("Done.\n")
sw_message("Predicting for test images...")
index = 0
pred_scores = []
predPath_ft = pred_dir + '/FT_Predictions'
if not os.path.exists(predPath_ft):
os.makedirs(predPath_ft)
predPath_pca = pred_dir + '/PCA_Predictions'
if not os.path.exists(predPath_pca):
os.makedirs(predPath_pca)
for img, pca, mdl in test_loader:
if sw_check_abort():
sw_message("Aborted")
return
sw_message(f"Predicting {index+1}/{len(test_loader)}")
sw_progress((index + 1) / len(test_loader))
img = img.to(device)
[pred, pred_mdl_pca] = model_pca(img)
[pred, pred_mdl_ft] = model_ft(img)
pred_scores.append(pred.cpu().data.numpy()[0])
nmpred = predPath_pca + '/predicted_pca_' + test_names[
index] + '.particles'
np.savetxt(nmpred, pred_mdl_pca.squeeze().detach().cpu().numpy())
nmpred = predPath_ft + '/predicted_ft_' + test_names[
index] + '.particles'
np.savetxt(nmpred, pred_mdl_ft.squeeze().detach().cpu().numpy())
index += 1
print("Predicted test " + str(index) + ".")
print("Done.\n")
return pred_dir
def supervised_train(config_file):
with open(config_file) as json_file:
parameters = json.load(json_file)
model_dir = parameters['paths']['out_dir'] + parameters['model_name'] + '/'
if not os.path.exists(model_dir):
os.makedirs(model_dir)
loader_dir = parameters['paths']['loader_dir']
aug_dir = parameters['paths']['aug_dir']
num_epochs = parameters['trainer']['epochs']
learning_rate = parameters['trainer']['learning_rate']
eval_freq = parameters['trainer']['val_freq']
decay_lr = parameters['trainer']['decay_lr']
fine_tune = parameters['fine_tune']['enabled']
loss_func = method_to_call = getattr(losses, parameters["loss"]["function"])
# load the loaders
train_loader_path = loader_dir + "train"
validation_loader_path = loader_dir + "validation"
print("Loading data loaders...")
train_loader = torch.load(train_loader_path)
val_loader = torch.load(validation_loader_path)
print("Done.")
# initializations
num_pca = train_loader.dataset.pca_target[0].shape[0]
num_corr = train_loader.dataset.mdl_target[0].shape[0]
print("Defining net...")
net = model.DeepSSMNet(config_file)
device = net.device
net.to(device)
# initialize model weights
net.apply(weight_init(module=nn.Conv2d, initf=nn.init.xavier_normal_))
net.apply(weight_init(module=nn.Linear, initf=nn.init.xavier_normal_))
# these lines are for the fine tuning layer initialization
whiten_mean = np.load(loader_dir + '/mean_PCA.npy')
whiten_std = np.load(loader_dir + '/std_PCA.npy')
orig_mean = np.loadtxt(aug_dir + '/PCA_Particle_Info/mean.particles')
orig_pc = np.zeros([num_pca, num_corr*3])
for i in range(num_pca):
temp = np.loadtxt(aug_dir + '/PCA_Particle_Info/pcamode' + str(i) + '.particles')
orig_pc[i, :] = temp.flatten()
bias = torch.from_numpy(orig_mean.flatten()).to(device) # load the mean here
weight = torch.from_numpy(orig_pc.T).to(device) # load the PCA vectors here
net.decoder.fc_fine.bias.data.copy_(bias)
net.decoder.fc_fine.weight.data.copy_(weight)
# define optimizer
# for the initial steps set the gradient of the final layer to be zero
for param in net.decoder.fc_fine.parameters():
param.requires_grad = False
train_params = net.parameters()
opt = torch.optim.Adam(train_params, learning_rate)
opt.zero_grad()
scheduler = StepLR(opt, step_size=1, gamma=0.99)
print("Done.")
# train
print("Beginning training on device = " + device + '\n')
# Initialize logger
logger = open(model_dir + "train_log.csv", "w+", buffering=1)
log_print(logger, ["Epoch", "LR", "Train_Err", "Train_Rel_Err", "Val_Err", "Val_Rel_Err", "Sec"])
# Initialize training plot
train_plot = plt.figure()
axe = train_plot.add_subplot(111)
axe.set_title('DeepSSM Training')
sp_train, = axe.plot([],[],label='Training',ms=10,color='b',marker='o',ls='')
sp_val, = axe.plot([],[],label='Validation',ms=10,color='r',marker='o',ls='')
axe.set_xlabel('Epochs')
axe.set_xlim(0,num_epochs+1)
axe.set_ylabel('PCA MSE')
axe.legend()
train_plot.savefig(model_dir + "training_plot.png", dpi=300)
# initialize
epochs = []
plot_train_losses = []
plot_val_losses = []
t0 = time.time()
best_val_rel_error = np.Inf
for e in range(1, num_epochs + 1):
if sw_check_abort():
sw_message("Aborted")
return
sw_message(f"Epoch {e}/{num_epochs}")
sw_progress(e / (num_epochs+1))
torch.cuda.empty_cache()
# train
net.train()
train_losses = []
train_rel_losses = []
pred_particles = []
true_particles = []
for img, pca, mdl in train_loader:
opt.zero_grad()
img = img.to(device)
pca = pca.to(device)
[pred_pca, pred_mdl] = net(img)
loss = loss_func(pred_pca, pca)
loss.backward()
opt.step()
train_losses.append(loss.item())
train_rel_loss = loss_func(pred_pca, pca) / loss_func(pred_pca*0, pca)
train_rel_losses.append(train_rel_loss.item())
pred_particles.extend(pred_mdl.detach().cpu().numpy())
true_particles.extend(mdl.detach().cpu().numpy())
train_viz.write_examples(np.array(pred_particles), np.array(true_particles), model_dir + "examples/train_")
# test validation
pred_particles = []
true_particles = []
if ((e % eval_freq) == 0 or e == 1):
net.eval()
val_losses = []
val_rel_losses = []
for img, pca, mdl in val_loader:
opt.zero_grad()
img = img.to(device)
pca = pca.to(device)
[pred_pca, pred_mdl] = net(img)
v_loss = loss_func(pred_pca, pca)
val_losses.append(v_loss.item())
val_rel_loss = loss_func(pred_pca, pca) / loss_func(pred_pca*0, pca)
val_rel_losses.append(val_rel_loss.item())
pred_particles.extend(pred_mdl.detach().cpu().numpy())
true_particles.extend(mdl.detach().cpu().numpy())
train_viz.write_examples(np.array(pred_particles), np.array(true_particles), model_dir + "examples/validation_")
# log
train_mr_MSE = np.mean(np.sqrt(train_losses))
val_mr_MSE = np.mean(np.sqrt(val_losses))
train_rel_err = np.mean(train_rel_losses)
val_rel_err = np.mean(val_rel_losses)
log_print(logger, [e, scheduler.get_lr()[0], train_mr_MSE, train_rel_err, val_mr_MSE, val_rel_err, time.time()-t0])
# plot
epochs.append(e)
plot_train_losses.append(train_mr_MSE)
plot_val_losses.append(val_mr_MSE)
sp_train.set_data(epochs, plot_train_losses)
sp_val.set_data(epochs, plot_val_losses)
axe.set_ylim(0,max(max(plot_train_losses), max(plot_val_losses))+3)
train_plot.canvas.draw()
train_plot.savefig(model_dir + "training_plot.png")
# save
if val_rel_err < best_val_rel_error:
best_val_rel_error = val_rel_err
best_epoch = e
torch.save(net.state_dict(), os.path.join(model_dir, 'best_model.torch'))
t0 = time.time()
if decay_lr:
scheduler.step()
logger.close()
torch.save(net.state_dict(), os.path.join(model_dir, 'final_model.torch'))
parameters['best_model_epochs'] = best_epoch
with open(config_file, "w") as json_file:
json.dump(parameters, json_file, indent=2)
print("Training complete, model saved. Best model after epoch " + str(best_epoch))
# now commence the fine tuning model if present on best model
if fine_tune:
print("Beginning fine tuning training step on device = ", device)
net = model.DeepSSMNet(config_file)
if parameters["use_best_model"]:
model_path = os.path.join(model_dir, 'best_model.torch')
else:
model_path = os.path.join(model_dir, 'final_model.torch')
net.load_state_dict(torch.load(model_path))
net.to(device)
logger = open(model_dir + "train_log_ft.csv", "w+")
log_print(logger, ["Epoch", "Train_Err_mdl", "Train_Rel_Err_mdl", "Val_Err_mdl", "Val_Rel_Err_mdl", "Sec"])
ft_epochs = parameters['fine_tune']['epochs']
learning_rate = parameters['fine_tune']['learning_rate']
eval_freq = parameters['fine_tune']['val_freq']
decay_lr = parameters['fine_tune']['decay_lr']
loss_func = method_to_call = getattr(losses, parameters['fine_tune']["loss"])
# free the last params
for param in net.decoder.fc_fine.parameters():
param.requires_grad=True
# train on the corr loss
best_ft_val_rel_error = np.Inf
for e in range(1, ft_epochs + 1):
if sw_check_abort():
sw_message("Aborted")
return
sw_message(f"Fine Tuning, Epoch {e}/{ft_epochs}")
sw_progress(e / (num_epochs+1))
torch.cuda.empty_cache()
# train
net.train()
train_losses = []
train_rel_losses = []
pred_particles = []
true_particles = []
for img, pca, mdl in train_loader:
opt.zero_grad()
img = img.to(device)
mdl = mdl.to(device)
[pred_pca, pred_mdl] = net(img)
loss = torch.mean((pred_mdl - mdl)**2)
loss.backward()
opt.step()
train_losses.append(loss.item())
train_rel_loss = F.mse_loss(pred_mdl, mdl) / F.mse_loss(pred_mdl*0, mdl)
train_rel_losses.append(train_rel_loss.item())
pred_particles.extend(pred_mdl.detach().cpu().numpy())
true_particles.extend(mdl.detach().cpu().numpy())
train_viz.write_examples(np.array(pred_particles), np.array(true_particles), model_dir + "examples/train_")
# test validation
pred_particles = []
true_particles = []
if ((e % eval_freq) == 0 or e == 1):
net.eval()
val_losses = []
val_rel_losses = []
for img, pca, mdl in val_loader:
opt.zero_grad()
img = img.to(device)
mdl = mdl.to(device)
[pred_pca, pred_mdl] = net(img)
v_loss = torch.mean((pred_mdl - mdl)**2)
val_losses.append(v_loss.item())
val_rel_loss = (F.mse_loss(pred_mdl, mdl) / F.mse_loss(pred_mdl*0, mdl)).item()
val_rel_losses.append(val_rel_loss)
if val_rel_loss < best_ft_val_rel_error:
best_ft_val_rel_error = val_rel_loss
best_ft_epoch = e
torch.save(net.state_dict(), os.path.join(model_dir, 'best_model_ft.torch'))
pred_particles.extend(pred_mdl.detach().cpu().numpy())
true_particles.extend(mdl.detach().cpu().numpy())
train_viz.write_examples(np.array(pred_particles), np.array(true_particles), model_dir + "examples/validation_")
# log
train_mr_MSE = np.mean(np.sqrt(train_losses))
val_mr_MSE = np.mean(np.sqrt(val_losses))
train_rel_err = np.mean(train_rel_losses)
val_rel_err = np.mean(val_rel_losses)
log_print(logger, [e, train_mr_MSE, train_rel_err, val_mr_MSE, val_rel_err, time.time()-t0])
t0 = time.time()
logger.close()
torch.save(net.state_dict(), os.path.join(model_dir, 'final_model_ft.torch'))
parameters['best_ft_model_epochs'] = best_ft_epoch
with open(config_file, "w") as json_file:
json.dump(parameters, json_file, indent=2)
print("Fine tuning complete, model saved. Best model after epoch " + str(best_ft_epoch))
def runDataAugmentation(out_dir, img_list, local_point_list, num_samples=3, num_dim=0, percent_variability=0.95, sampler_type="KDE", mixture_num=0, processes=1, world_point_list=None):
sw_message("Running point based data augmentation.")
num_dim = DataAugmentation.point_based_aug(out_dir, img_list, local_point_list, num_samples, num_dim, percent_variability, sampler_type, mixture_num, processes, world_point_list)
sw_message("Done.")
return num_dim
def point_based_aug(out_dir, orig_img_list, orig_point_list, num_samples, num_dim=0, percent_variability=0.95, sampler_type="kde", mixture_num=0, processes=1, orig_world_point_list=None):
# Get Embedder
point_matrix = Utils.create_data_matrix(orig_point_list)
if orig_world_point_list is not None:
world_point_matrix = Utils.create_data_matrix(orig_world_point_list)
world_get_local = [Utils.estimate_homogeneous_similar_transform(
x=p.reshape((-1,3)).T,
y=point_matrix[i].reshape((-1,3)).T,
) for i,p in enumerate(world_point_matrix)]
world_get_local_info = {
'world_get_local_list':[
{
'homogeneous_transformation': m.tolist(),
'origi_world_point': orig_world_point_list[i],
'origi_local_point': orig_point_list[i],
} for i, m in enumerate(world_get_local)
],
'generated_particle__world_get_local_index':[],
}
PointEmbedder = Embedder.PCA_Embbeder(world_point_matrix, num_dim, percent_variability)
else:
PointEmbedder = Embedder.PCA_Embbeder(point_matrix, num_dim, percent_variability)
num_dim = PointEmbedder.num_dim
PointEmbedder.write_PCA(out_dir + "PCA_Particle_Info/", "particles") # write PCA info for DeepSSM testing
embedded_matrix = PointEmbedder.getEmbeddedMatrix()
# Get sampler
if sampler_type == "gaussian":
PointSampler = Sampler.Gaussian_Sampler()
PointSampler.fit(embedded_matrix)
elif sampler_type == "mixture":
PointSampler = Sampler.Mixture_Sampler()
PointSampler.fit(embedded_matrix, mixture_num)
elif sampler_type == "kde":
PointSampler = Sampler.KDE_Sampler()
PointSampler.fit(embedded_matrix)
else:
sw_message("Error sampler_type unrecognized.")
sw_message("Gaussian, mixture, and KDE currently supported.")
return 0
# Initialize output folders and lists
gen_point_dir = out_dir + "Generated-Particles/"
if not os.path.exists(gen_point_dir):
os.makedirs(gen_point_dir)
gen_image_dir = out_dir + "Generated-Images/"
if not os.path.exists(gen_image_dir):
os.makedirs(gen_image_dir)
gen_embeddings = []
gen_points_paths = []
gen_image_paths = []
if processes != 1:
generate_image_params_list = []
# Sample to generate new examples
for index in range(1, num_samples+1):
if sw_check_abort():
sw_message("Aborted")
return 0
sw_message("Generating " +str(index)+'/'+str(num_samples))
sw_progress(index / (num_samples+1))
name = 'Generated_sample_' + Utils.pad_index(index)
# Generate embedding
sampled_embedding, base_index = PointSampler.sample()
gen_embeddings.append(sampled_embedding)
# Generate particles
if orig_world_point_list is not None:
p = PointEmbedder.project(sampled_embedding)
i = int(np.nanargmin(np.linalg.norm(p-world_point_matrix, axis=(1,2))))
# TODO add Randomness in the transformation from world to local particles
gen_points = (world_get_local[i] @ Utils.get_homogeneous_coordinates(p.reshape((-1,3)).T))[:3,:].T
world_get_local_info['generated_particle__world_get_local_index'].append(i)
else:
gen_points = PointEmbedder.project(sampled_embedding)
gen_points_path = gen_point_dir + name + ".particles"
np.savetxt(gen_points_path, gen_points)
gen_points_paths.append(gen_points_path)
# Generate image
base_image_path = orig_img_list[base_index]
base_particles_path = orig_point_list[base_index]
if processes==1:
gen_image_path = Utils.generate_image(out_dir, gen_points_path, base_image_path, base_particles_path)
gen_image_paths.append(gen_image_path)
else:
generate_image_params_list.append({
'out_dir':out_dir,
'gen_points_path':gen_points_path,
'base_image_path':base_image_path,
'base_particles_path':base_particles_path,
})
if orig_world_point_list is not None:
# write world to local transformation information for generated particles
with open(out_dir + '/world_get_local_info.json', 'w') as f:
json.dump(world_get_local_info, f)
if processes!=1:
with mtps.Pool(processes=processes) as p:
gen_image_paths = p.map(generate_image, generate_image_params_list)
csv_file = out_dir + "TotalData.csv"
Utils.make_CSV(out_dir + "TotalData.csv", orig_img_list, orig_point_list, embedded_matrix, gen_image_paths, gen_points_paths, gen_embeddings)
return num_dim
def fit(self, embedded_matrix):
sw_message("Fitting Gaussian distribution...")
self.embedded_matrix = embedded_matrix
self.mean = np.mean(embedded_matrix, axis=0)
self.cov = np.cov(embedded_matrix, rowvar=0)