def evaluate():
current_dir = os.getcwd()
print("Current_dir = ", current_dir)
model_dir = "./saved_models"
save_files = os.listdir(model_dir)
# Filter only asymmetric versions:
save_files = [
x for x in save_files if (int(x.split("_")[1]) == test_for_G)
]
indexes = [int(x.split("_")[2]) for x in save_files]
save_files = [x for _, x in sorted(zip(indexes, save_files))]
indexes = sorted(indexes)
print("Save files found: ", save_files)
# TODO: Filter save files to only load symmetrically trained ones... i.e. indexes mus be equal.
print("Depths parsed: ", indexes)
l = len(save_files)
image_batch = np.zeros([256, 64, 64, 3])
DCG = DCGAN()
num_pool_workers = 1
i = 0
noise = tf.random_normal([n_samples, 128])
# print("Noise: ", noise)
for j, gen in enumerate(save_files):
DCG.set_G_dim(test_for_G)
print("G_dim set to ", test_for_G)
# test_function(DCG, gen, disc)
param_tuple = (DCG, gen)
with contextlib.closing(Pool(num_pool_workers)) as po:
pool_results = po.map_async(test_function, (param_tuple, ))
results_list = pool_results.get()
image_batch[i * n_samples:n_samples * (i + 1)] = results_list[0]
i += 1
# print(image_batch)
generate_png(image_batch)
print("Output saved")
def evaluate():
"""
Should be able to mix and match various versions of GANs with this function. Steps:
1. List all models available in the save_dir
2. Double for loop:
2.1 For each Generator
2.2 For load each discriminator and
3. Run 100k samples and see the discriminator output.
"""
current_dir = os.getcwd()
print("Current_dir = ", current_dir)
model_dir = "./saved_models"
save_files = os.listdir(model_dir)
# Filter only symmetric versions:
save_files = [x for x in save_files if (
x.split("_")[1] == x.split("_")[2])]
indexes = [int(x.split("_")[1]) for x in save_files]
save_files = [x for _, x in sorted(zip(indexes, save_files))]
indexes = sorted(indexes)
print("Save files found: ", save_files)
print("Depths parsed: ", indexes)
l = len(save_files)
DCG = DCGAN()
i = 0
images_to_explain = []
for normalize_by_mean in [True, False]:
for k, disc in enumerate(tqdm(save_files)):
DCG.set_D_dim(indexes[k])
param_tuple = (DCG, disc, images_to_explain, k, normalize_by_mean)
with contextlib.closing(Pool(1)) as po:
pool_results = po.map_async(
explain, (param_tuple,))
images_to_explain = pool_results.get()[0]
print("Evaluation finished")
def evaluate():
"""
Should be able to mix and match various versions of GANs with this function. Steps:
1. List all models available in the save_dir
2. Double for loop:
2.1 For each Generator
2.2 For load each discriminator and
3. Run 100k samples and see the discriminator output.
"""
current_dir = os.getcwd()
print("Current_dir = ", current_dir)
model_dir = "./saved_models"
save_files = os.listdir(model_dir)
save_files = [
x for x in save_files if (x.split("_")[1] == x.split("_")[2])
]
indexes = [int(x.split("_")[1]) for x in save_files]
save_files = [x for _, x in sorted(zip(indexes, save_files))]
indexes = sorted(indexes)
print("Save files found: ", save_files)
# TODO: Filter save files to only load symmetrically trained ones... i.e. indexes mus be equal.
print("Depths parsed: ", indexes)
l = len(save_files)
results_tensor = np.empty([l, no_samples, FLAGS.batch_size])
DCG = DCGAN()
num_pool_workers = 1
i = 0
for k, disc in enumerate(tqdm(save_files)):
DCG.set_D_dim(indexes[k])
print("D_dim set to ", indexes[k])
param_tuple = (DCG, disc)
with contextlib.closing(Pool(num_pool_workers)) as po:
pool_results = po.map_async(test_function,
(param_tuple for _ in range(1)))
results_list = pool_results.get()
results_tensor[i] = results_list[0]
i += 1
print("Evaluation finished")
output_path = './real_evaluation_stats.npz'
np.savez_compressed(output_path,
results_tensor=results_tensor,
save_files=save_files)
print("Output saved")
def launch_managed_training():
"""
Function which trains many models in series
"""
epochs = 10
DCG = DCGAN()
num_pool_workers = 1 # can be bigger than 1, to enable parallel execution
for entry in models_and_options:
print("Starting training of model: \n", entry)
DCG.set_G_dim(entry[1])
DCG.set_D_dim(entry[2])
if entry[3] == 1:
Generator = DCG.DCGANG_1
else:
Generator = DCG.DCGANG_2
if entry[4] == 1:
Discriminator = DCG.DCGAND_1
else:
Discriminator = DCG.DCGAND_2
param_tuple = (entry[0], Generator, Discriminator, epochs, entry[6])
with contextlib.closing(Pool(num_pool_workers)) as po:
pool_results = po.map_async(
begin_training, (param_tuple for _ in range(1)))
results_list = pool_results.get()
# This ensures that the processes get closed once they are done
return 0
def evaluate():
"""
Should be able to mix and match various versions of GANs with this function. Steps:
1. List all models available in the save_dir
2. Double for loop:
2.1 For each Generator
2.2 For load each discriminator and
3. Run 100k samples and see the discriminator output.
"""
current_dir = os.getcwd()
print("Current_dir = ", current_dir)
model_dir = "./saved_models"
save_files = os.listdir(model_dir)
# Filter only symmetric versions:
save_files = [x for x in save_files if (
x.split("_")[1] == x.split("_")[2])]
indexes = [int(x.split("_")[1]) for x in save_files]
save_files = [x for _, x in sorted(zip(indexes, save_files))]
indexes = sorted(indexes)
print("Save files found: ", save_files)
print("Depths parsed: ", indexes)
l = len(save_files)
results_tensor = np.empty([l * l, 6])
# I.e. for each of l architectures, show prediction mean for real, fake, mixed
DCG = DCGAN()
num_pool_workers = 1
i = 0
for j, gen in enumerate(tqdm(save_files)):
for k, disc in enumerate(tqdm(save_files)):
DCG.set_G_dim(indexes[j])
DCG.set_D_dim(indexes[k])
param_tuple = (DCG, gen, disc)
with contextlib.closing(Pool(num_pool_workers)) as po:
pool_results = po.map_async(
test_function, (param_tuple for _ in range(1)))
results_list = pool_results.get()
results_tensor[i] = results_list[0]
i += 1
print("Evaluation finished")
output_path = './real_fake_half.npz'
results_tensor.reshape([l, l, 6])
np.savez_compressed(
output_path, results_tensor=results_tensor, save_files=save_files)
print("Output saved")