def main():
parser = get_stimuli_compute_parser()
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
assert args.set_func, "Must specify set_func"
args = bs_fit.load_set_func(args)
compute_aggregate_dump(args, builder=StimuliCompute)
def main():
parser = get_compute_raw_response_parser()
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
assert args.set_func, "Must specify set_func"
args = bs_fit.load_set_func(args)
raw_stimuli_compute = RawStimuliCompute(args)
input_images = np.load(args.file_path)[:args.batch_size]
output = raw_stimuli_compute.get_output(input_images)
print(np.trace(output[:, :args.batch_size]))
def main():
parser = get_circular_var_parser()
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
assert args.set_func, "Must specify set_func"
args = bs_fit.load_set_func(args)
cir_var_compute = CircularVarCompute(args)
layers = cir_var_compute.layers
for which_split in range(4):
cir_var_compute.which_split = which_split
for layer in tqdm(layers):
cir_var_compute.compute_save_cir_var(layer)
def main():
parser = get_lucid_stimuli_compute_parser()
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
assert args.set_func, "Must specify set_func"
args = bs_fit.load_set_func(args)
layers = bs_fit.TF_RES18_LAYERS
if getattr(args, 'layers', None) is not None:
layers = args.layers.split(',')
for layer in tqdm(layers):
layer, image, all_losses = get_vm_model_image_losses(args, layer=layer)
#layer, image, all_losses = get_lucid_orig_image_losses(args)
dump_results(args, layer, image, all_losses)
def main():
parser = get_raw_pt_stimuli_compute_parser()
args = parser.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
assert args.set_func, "Must specify set_func"
args = bs_fit.load_set_func(args)
batch_size = 16
num_batches = args.num_batches
#batch_size = 4
#num_batches = 16
#for layer in tqdm(['layer2.1.relu', 'layer4.1.relu']):
layer_list = bs_fit.PT_RES18_LAYERS
if args.layer_start_idx is not None:
layer_end_idx = args.layer_start_idx + args.layer_len_idx
layer_list = layer_list[\
args.layer_start_idx : layer_end_idx]
for layer in tqdm(layer_list):
all_save_results = []
for start_idx in range(0, num_batches * batch_size, batch_size):
#stimuli_compute = RawPtStimuliCompute(args, start_idx, batch_size)
stimuli_compute = FFTRawPtStimuliCompute(
args, start_idx + (args.global_start_idx or 0), batch_size)
save_result = stimuli_compute.compute_optimal_stimuli(layer)
all_save_results.append(save_result)
del stimuli_compute
images = []
loss_vals = []
actual_response = []
for save_result in all_save_results:
images.append(save_result['images'])
loss_vals.append(save_result['loss_vals'])
actual_response.append(save_result['actual_response'])
images = np.concatenate(images, axis=0)
loss_vals = np.sum(loss_vals, axis=0)
actual_response = np.concatenate(actual_response, axis=0)
final_save_result = {
'images': images,
'loss_vals': loss_vals,
'actual_response': actual_response,
}
dump_result(final_save_result, args.identifier, layer,
args.global_start_idx)