elif opt == '-c':
epoch_resume = int(arg)
if epoch_resume < 0:
error_and_exit()
elif opt == '-b':
batch_size = int(arg)
else:
error_and_exit()
# compulsory params
if loss is None or noise is None:
error_and_exit()
print("Params: loss=%s, noise=%s" % (loss, noise))
model = densenet.get_densenet(2)
train_loader, val_loader = utils.get_covid_loaders(float(noise),
batch_size=batch_size)
model_folder = './models/forwardt/'
model_path = fpath('./models', loss, noise)
if not os.path.exists(model_folder):
os.mkdir(model_folder)
# Check this.
filter_outlier = False
kerasModel = ChexpertModel(model,
train_loader,
val_loader,
def __init__(self,
encoder_name,
num_input_channels=1,
drop_rate=0,
img_size=(160, 160),
data_parallel=True,
continuous_latent_dim=100,
disc_latent_dim=10,
sample_temperature=0.67,
small_input=False):
super(VariationalAutoEncoder, self).__init__()
if "densenet" in encoder_name:
self.feature_extractor = get_densenet(
encoder_name,
drop_rate,
input_channels=num_input_channels,
small_input=small_input,
data_parallel=data_parallel)
elif "wideresnet" in encoder_name:
self.feature_extractor = get_wide_resnet(
encoder_name,
drop_rate,
input_channels=num_input_channels,
small_input=small_input,
data_parallel=data_parallel)
elif "preactresnet" in encoder_name:
self.feature_extractor = get_preact_resnet(
encoder_name,
drop_rate,
input_channels=num_input_channels,
small_input=small_input,
data_parallel=data_parallel)
else:
raise NotImplementedError(
"{} not implemented".format(encoder_name))
global_avg = nn.AdaptiveAvgPool2d(output_size=(1, 1))
if data_parallel:
global_avg = nn.DataParallel(global_avg)
self.global_avg = global_avg
self.continuous_inference = nn.Sequential()
self.disc_latent_inference = nn.Sequential()
conti_mean_inf_module = _Inference(
num_input_channels=self.feature_extractor.num_feature_channel,
latent_dim=continuous_latent_dim,
disc_variable=False)
conti_logsigma_inf_module = _Inference(
num_input_channels=self.feature_extractor.num_feature_channel,
latent_dim=continuous_latent_dim,
disc_variable=False)
if data_parallel:
conti_mean_inf_module = nn.DataParallel(conti_mean_inf_module)
conti_logsigma_inf_module = nn.DataParallel(
conti_logsigma_inf_module)
self.continuous_inference.add_module("mean", conti_mean_inf_module)
self.continuous_inference.add_module("log_sigma",
conti_logsigma_inf_module)
self._disc_latent_dim = disc_latent_dim
dic_inf = _Inference(
num_input_channels=self.feature_extractor.num_feature_channel,
latent_dim=disc_latent_dim,
disc_variable=True)
if data_parallel:
dic_inf = nn.DataParallel(dic_inf)
self.disc_latent_inference = dic_inf
sample = Sample(temperature=sample_temperature)
if data_parallel:
sample = nn.DataParallel(sample)
self.sample = sample
decoder_kernel_size = tuple([int(s / 32) for s in img_size])
self.feature_reconstructor = Decoder(
num_channel=num_input_channels,
latent_dim=int(continuous_latent_dim + np.sum(disc_latent_dim)),
data_parallel=data_parallel,
kernel_size=decoder_kernel_size)