def main(params):
with open(params.lowshotmeta, 'r') as f:
lowshotmeta = json.load(f)
accs = np.zeros(6)
with open(params.experimentpath.format(params.experimentid), 'r') as f:
exp = json.load(f)
novel_idx = np.array(exp)[:, :params.lowshotn]
if params.testsetup:
novel_classes = lowshotmeta['novel_classes_2']
base_classes = lowshotmeta['base_classes_2']
else:
novel_classes = lowshotmeta['novel_classes_1']
base_classes = lowshotmeta['base_classes_1']
novel_idx = np.sort(novel_idx[novel_classes, :].reshape(-1))
generator = None
generator_fn = None
if params.generator_name != '':
generator_fn, generator = generation.get_generator(
params.generator_name, params.generator_file)
with h5py.File(params.trainfile, 'r') as f:
lowshot_dataset = LowShotDataset(f, base_classes, novel_classes,
novel_idx, params.max_per_label,
generator_fn, generator)
model = training_loop(lowshot_dataset, params.numclasses, params,
params.batchsize, params.maxiters)
print('trained')
with h5py.File(params.testfile, 'r') as f:
test_loader = get_test_loader(f)
accs, results = eval_loop(test_loader, model, base_classes,
novel_classes)
modelrootdir = os.path.basename(os.path.dirname(params.trainfile))
outpath = os.path.join(
params.outdir, modelrootdir +
'_lr_{:.3f}_wd_{:.3f}_expid_{:d}_lowshotn_{:d}_maxgen_{:d}.json'.
format(params.lr, params.wd, params.experimentid, params.lowshotn,
params.max_per_label))
with open(outpath, 'w+') as f:
json.dump(
dict(lr=params.lr,
wd=params.wd,
expid=params.experimentid,
lowshotn=params.lowshotn,
accs=accs.tolist()), f)
return results
exp = json.load(f)
novel_idx = np.array(exp)[:, :params.lowshotn]
if params.testsetup:
novel_classes = lowshotmeta['novel_classes_2'][:params.n_way]
base_classes = lowshotmeta['base_classes_2']
else:
novel_classes = lowshotmeta['novel_classes_1'][:params.n_way]
base_classes = lowshotmeta['base_classes_1']
novel_idx = np.sort(novel_idx[novel_classes, :].reshape(-1))
generator = None
generator_fn = None
if params.generator_name != '':
generator_fn, generator = generation.get_generator(
params.generator_name, params.generator_file)
with h5py.File(params.trainfile, 'r') as f:
lowshot_dataset = LowShotDataset(f, base_classes, novel_classes,
novel_idx, params.max_per_label,
generator_fn, generator)
model = training_loop(lowshot_dataset, params.numclasses, params,
params.batchsize, params.maxiters)
print('trained')
with h5py.File(params.testfile, 'r') as f:
test_loader = get_test_loader(f)
accs = eval_loop(test_loader, model, base_classes, novel_classes)
modelrootdir = os.path.basename(os.path.dirname(params.trainfile))
outpath = os.path.join(
def test(test_loader, test_dataset, model, cfg):
model.eval()
generator = get_generator(model)
model_counter = defaultdict(int)
if cfg['mode'] == 'eval':
eval_meshes(test_loader, test_dataset, model, cfg)
return 0
for it, data in enumerate(tqdm(test_loader)):
# Output folders
mesh_dir = os.path.join(cfg['out']['out_dir'], 'meshes')
pointcloud_dir = os.path.join(cfg['out']['out_dir'], 'pointcloud')
in_dir = os.path.join(cfg['out']['out_dir'], 'input')
generation_vis_dir = os.path.join(
cfg['out']['out_dir'],
'vis',
)
# Get index etc.
idx = data['idx'].item()
try:
model_dict = test_dataset.get_model_dict(idx)
except AttributeError:
model_dict = {'model': str(idx), 'category': 'n/a'}
modelname = model_dict['model']
category_id = model_dict.get('category', 'n/a')
try:
category_name = test_dataset.metadata[category_id].get(
'name', 'n/a')
except AttributeError:
category_name = 'n/a'
if category_id != 'n/a':
mesh_dir = os.path.join(mesh_dir, str(category_id))
pointcloud_dir = os.path.join(pointcloud_dir, str(category_id))
in_dir = os.path.join(in_dir, str(category_id))
folder_name = str(category_id)
if category_name != 'n/a':
folder_name = str(folder_name) + '_' + category_name.split(
',')[0]
generation_vis_dir = os.path.join(generation_vis_dir, folder_name)
if not os.path.exists(generation_vis_dir):
os.makedirs(generation_vis_dir)
if not os.path.exists(mesh_dir):
os.makedirs(mesh_dir)
if not os.path.exists(pointcloud_dir):
os.makedirs(pointcloud_dir)
if not os.path.exists(in_dir):
os.makedirs(in_dir)
# Generate outputs
out_file_dict = {}
# Also copy ground truth
out = generator.generate_mesh(data)
# Get statistics
try:
mesh, stats_dict = out
except TypeError:
mesh, stats_dict = out, {}
#time_dict.update(stats_dict)
# Write output
mesh_out_file = os.path.join(mesh_dir, '%s.off' % modelname)
mesh.export(mesh_out_file)
out_file_dict['mesh'] = mesh_out_file
# Save inputs
inputs_path = os.path.join(in_dir, '%s.jpg' % modelname)
inputs = data['inputs'].squeeze(0).cpu()
visualize_data(inputs, 'img', inputs_path)
out_file_dict['in'] = inputs_path
# Copy to visualization directory for first vis_n_output samples
c_it = model_counter[category_id]
if c_it < cfg['test']['vis_n_outputs']:
# Save output files
img_name = '%02d.off' % c_it
for k, filepath in out_file_dict.items():
ext = os.path.splitext(filepath)[1]
out_file = os.path.join(generation_vis_dir,
'%02d_%s%s' % (c_it, k, ext))
shutil.copyfile(filepath, out_file)
model_counter[category_id] += 1
exp = json.load(f)
novel_idx = np.array(exp)[:,:params.lowshotn]
if params.testsetup:
novel_classes = lowshotmeta['novel_classes_2']
base_classes = lowshotmeta['base_classes_2']
else:
novel_classes = lowshotmeta['novel_classes_1']
base_classes = lowshotmeta['base_classes_1']
novel_idx = np.sort(novel_idx[novel_classes,:].reshape(-1))
generator=None
generator_fn=None
if params.generator_name!='':
generator_fn, generator = generation.get_generator(params.generator_name, params.generator_file)
with h5py.File(params.trainfile, 'r') as f:
lowshot_dataset = LowShotDataset(f, base_classes, novel_classes, novel_idx, params.max_per_label, generator_fn, generator)
model = training_loop(lowshot_dataset, params.numclasses, params, params.batchsize, params.maxiters)
print('trained')
with h5py.File(params.testfile, 'r') as f:
test_loader = get_test_loader(f)
accs = eval_loop(test_loader, model, base_classes, novel_classes)
modelrootdir = os.path.basename(os.path.dirname(params.trainfile))
outpath = os.path.join(params.outdir, modelrootdir+'_lr_{:.3f}_wd_{:.3f}_expid_{:d}_lowshotn_{:d}_maxgen_{:d}.json'.format(
params.lr, params.wd, params.experimentid, params.lowshotn, params.max_per_label))
with open(outpath, 'w') as f:
exp = json.load(f)
novel_idx = np.array(exp)[:, :params.lowshotn]
if params.testsetup:
novel_classes = lowshotmeta['novel_classes_2']
base_classes = lowshotmeta['base_classes_2']
else:
novel_classes = lowshotmeta['novel_classes_1']
base_classes = lowshotmeta['base_classes_1']
novel_idx = np.sort(novel_idx[novel_classes, :].reshape(-1))
generator = None
generator_fn = None
if params.generator_name != '':
generator_fn, generator, trial_gen = generation.get_generator(
params.generator_name, params.generator_file)
with h5py.File(params.trainfile, 'r') as f:
if params.generator_type == 1:
lowshot_dataset = LowShotDataset(
f,
base_classes,
novel_classes,
novel_idx,
params.max_per_label,
trial_gen,
generator,
centroid_file=params.centroid_file,
generator_index=params.generator_type)
else:
lowshot_dataset = LowShotDataset(
