def get_config(return_unparsed=False):
"""Gets config and creates data_dir."""
config, unparsed = parse_config()
# If we have unparsed args, print usage and exit
if len(unparsed) > 0 and not return_unparsed:
print_usage()
exit(1)
def append_data_dir(p):
return os.path.join(config.data_dir, p)
# Append data_dir to all filepaths
config.pre_save_file = append_data_dir(config.pre_save_file)
config.raw_csv_file = append_data_dir(config.raw_csv_file)
config.embeddings_model = append_data_dir(config.embeddings_model)
config.embeddings_file = append_data_dir(config.embeddings_file)
# Create data_dir if it doesn't exist
if not os.path.exists(config.data_dir):
os.makedirs(config.data_dir)
if return_unparsed:
return config, unparsed
return config
def main(config):
if __name__ == "__main__":
# ----------------------------------------
# Parse configuration
config, unparsed = get_config()
# If we have unparsed arguments, print usage and exit
if len(unparsed) > 0:
print_usage()
exit(1)
print_config(config)
main(config)
timings_list = []
for i, pair in enumerate(pairs_per_th['0.0']):
model_dict[pair] = result[i][0]
inl_dict[pair] = result[i][1]
timings_list.append(result[i][2])
# Check model directory
if not os.path.exists(get_geom_path(cfg)):
os.makedirs(get_geom_path(cfg))
# Finally save packed models
save_h5(model_dict, get_geom_file(cfg))
save_h5(inl_dict, get_geom_inl_file(cfg))
# Save computational cost
save_h5({'cost': np.mean(timings_list)}, get_geom_cost_file(cfg))
print('Geometry cost (averaged over image pairs): {:0.2f} sec'.format(
np.mean(timings_list)))
if __name__ == '__main__':
cfg, unparsed = get_config()
# If we have unparsed arguments, print usage and exit
if len(unparsed) > 0:
print(unparsed)
print_usage()
exit(1)
main(cfg)
# Compute loss and store as numpy
loss = data_loss(logits, y)
te_loss += [loss.cpu().numpy()]
# Compute accuracy and store as numpy
pred = torch.argmax(logits, dim=1)
acc = torch.mean(torch.eq(pred, y).float()) * 100.0
te_acc += [acc.cpu().numpy()]
# Report Test loss and accuracy
print("Test Loss = {}".format(np.mean(te_loss)))
print("Test Accuracy = {}%".format(np.mean(te_acc)))
def main(config):
if config.mode == "train":
train(config)
elif config.mode == "test":
test(config)
else:
raise ValueError("Unknown run mode \"{}\"".format(config.mode))
if __name__ == "__main__":
config, unparsed = config_args.get()
# Verify all arguments are parsed before continuing.
if len(unparsed) == 0:
main(config.get)
else:
" Unparsed arguments "
config_args.print_usage()
def train():
# ----------------------------------------
# Parse configuration
config, unparsed = get_config()
# If we have unparsed arguments, print usage and exit
if len(unparsed) > 0:
print_usage()
exit(1)
print_config(config)
print("Number of train samples: ", len(train_data))
print("Number of test samples: ", len(test_data))
#print("Detected Classes are: ", train_data.class_to_idx) # classes are detected by folder structure
# Create log directory and save directory if it does not exist
if not os.path.exists(config.log_dir):
os.makedirs(config.log_dir)
if not os.path.exists(config.save_dir):
os.makedirs(config.save_dir)
# Initialize training
iter_idx = -1 # make counter start at zero
best_loss = -1 # to check if best loss
# Prepare checkpoint file and model file to save and load from
checkpoint_file = os.path.join(config.save_dir, "checkpoint.pth")
bestmodel_file = os.path.join(config.save_dir, "best_model.pth")
savemodel_file = os.path.join(config.save_dir, "save_model.pth")
model = sosnet_model.SOSNet32x32().cuda()
optimizer = optim.SGD(model.parameters(), lr=0.1, weight_decay=0.0001)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=10,
gamma=0.8)
seed = 42
torch.manual_seed(seed)
np.random.seed(seed)
# Create loss objects
data_loss = data_criterion(config)
model_loss = model_criterion(config)
#print("train_data: ", (train_data))
print("train_data_loader:", train_data_loader)
print("test_data_loader:", test_data_loader)
fpr_per_epoch = []
# Training loop
for epoch in range(config.num_epoch):
# For each iteration
prefix = "Training Epoch {:3d}: ".format(epoch)
print("len(train_data_loader):", len(train_data_loader))
for batch_idx, (data_a, data_p,
data_n) in tqdm(enumerate(train_data_loader)):
print("batch_idx:", batch_idx)
print("len(train_data_loader):", len(train_data_loader))
data_a = data_a.unsqueeze(1).float().cuda()
data_p = data_p.unsqueeze(1).float().cuda()
data_n = data_n.unsqueeze(1).float().cuda()
print("data_a.shape:", data_a.shape)
print("data_p.shape:", data_p.shape)
print("data_n.shape:", data_n.shape)
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
print("out_a:", out_a)
print("out_p:", out_p)
print("out_n:", out_n)
loss = F.triplet_margin_loss(out_a,
out_p,
out_n,
margin=2,
swap=True)
if best_loss == -1:
best_loss = loss
if loss < best_loss:
best_loss = loss
# Save
torch.save(
{
"iter_idx": iter_idx,
"best_loss": best_loss,
"model": model.state_dict(),
"optimizer": optimizer.state_dict(),
}, bestmodel_file)
# Save
torch.save(model.state_dict(), savemodel_file)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Save
torch.save(
{
"iter_idx": iter_idx,
"best_loss": best_loss,
"model": model.state_dict(),
"optimizer": optimizer.state_dict(),
}, checkpoint_file)
model.eval()
l = np.empty((0, ))
d = np.empty((0, ))
#evaluate the network after each epoch
for batch_idx, (data_l, data_r, lbls) in enumerate(test_data_loader):
data_l = data_l.unsqueeze(1).float().cuda()
data_r = data_r.unsqueeze(1).float().cuda()
out_l, out_r = model(data_l), model(data_r)
dists = torch.norm(out_l - out_r, 2, 1).detach().cpu().numpy()
l = np.hstack((l, lbls.numpy()))
d = np.hstack((d, dists))
# FPR95 code from Yurun Tian
d = torch.from_numpy(d)
l = torch.from_numpy(l)
dist_pos = d[l == 1]
dist_neg = d[l != 1]
dist_pos, indice = torch.sort(dist_pos)
loc_thr = int(np.ceil(dist_pos.numel() * 0.95))
thr = dist_pos[loc_thr]
fpr95 = float(dist_neg.le(thr).sum()) / dist_neg.numel()
print(epoch, fpr95)
fpr_per_epoch.append([epoch, fpr95])
scheduler.step()
np.savetxt('fpr.txt', np.array(fpr_per_epoch), delimiter=',')