ds_train, ds_test = data_utils.get_cached_dataset_with_normalized_inputs(
args.ds, trainval_test=args.trainval)
model_ds = ds_train
if args.test_dataset != "":
ds_train, ds_test = data_utils.get_cached_dataset_with_normalized_inputs(
args.test_dataset, trainval_test=args.trainval)
# Model
print("Initializing model and loading state...")
model = model_utils.get_model(
net_name=args.net,
ds_train=model_ds,
depth=args.depth,
width=args.width, # width factor for WRN, base_width for others
epoch_count=1,
dropout=args.dropout or args.mcdropout)
model.load_state(args.saved_path)
if args.view or args.hard or args.view2:
ds_disp = ds_train if args.test_on_training_set else ds_test
if args.hard:
ds_disp = training.get_hard_examples(model, ds_disp)
def predict(x):
out_names = ['output', 'probs', 'probs_entropy']
output, probs, probs_entropy = model.predict(x,
single_input=True,
'tinyimagenet-c': r'TinyImageNet-C',
'tinyimagenet-r': r'TinyImageNet-R',
'lsun-c': r'LSUN-C',
'lsun-r': r'LSUN-R',
'isun-r': r'iSUN',
'gaussian': r'Gaussiov šum',
'uniform': r'Uniformni šum',
}[name]
# Model
print("Initializing model and loading state...")
model = model_utils.get_model(net_name=args.net,
ds_train=ds_id_to_ds[args.ds],
depth=args.depth,
width=args.width,
epoch_count=1,
dropout=args.dropout or args.mcdropout)
model.load_state(args.saved_path)
temp = 0 if args.notemp else 1000
odin = Odin(model, temp=temp)
# Missclassification
if args.misclassified:
ds = ds_id_to_ds[args.ds]
corr = [model.predict(x, single_input=True) == y for x, y in tqdm(ds)]
correct = []
misclassified = []
for i, c in enumerate(corr):
dsid_to_ds = {
'CIFAR-10': ds,
'CIFAR-10-UD': ds_ud,
'random': ds_rand,
'CamVid': ds_camvid
}
print("Loading model")
#ds_id, net_name, depth, width = 'cifar', 'wrn', 28, 10
#saved_path = dirs.SAVED_NETS + '/cifar-trainval/wrn-28-10/2018-04-28-1926/Model'
ds_id, net_name, depth, width = 'mozgalo', 'rn', 18, 64
saved_path = dirs.SAVED_NETS + '/mozgalo-trainval/rn-18-64-e10/2018-05-13-1747/Model'
model = model_utils.get_model(
net_name=net_name, ds_train=ds, depth=depth, width=width)
model.load_state(saved_path)
print("Printing logit biases")
with model._graph.as_default():
vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
logit_biases_var = list(
filter(lambda x: x.name == 'conv_logits/bias:0', vars))[0]
logit_biases = model._sess.run(logit_biases_var)
print("Logit biases:", logit_biases)
print("Logit biases sum:", logit_biases.sum())
print("Logit biases mean:", logit_biases.mean())
print("Collecting logit statistics")