print('Model loaded')
#for eval_thresh in reversed([0.9, 0.91, 0.92, 0.93, 0.94,
# 0.95, 0.96, 0.97, 0.98, 0.99,
# 0.992, 0.994, 0.996, 0.998, 0.999,
# 3.0]):
for eval_thresh in reversed([
0.9, 0.91, 0.92, 0.93, 0.94, 0.95, 0.96,
0.97, 0.98, 0.99, 0.992, 0.994, 0.996,
0.998, 0.999, 0.999, 0.9991, 0.9992,
0.9993, 0.9994, 0.9995, 0.9996, 0.9997,
0.9998, 0.9999, 3.0
]):
#for eval_thresh in reversed([0.9991, 0.9992, 0.9993, 0.9994, 0.9995, 0.9996, 0.9997, 0.9998, 0.9999, 3.0]):
#change_all_pca_layer_thresholds_and_inject_random_directions(eval_thresh, model, verbose=False)
sat, indims, fsdims, lnames = change_all_pca_layer_thresholds(
eval_thresh, network=model)
start = time()
print('Changed model threshold to',
eval_thresh)
#model = model.to(trainer.device)
#trainer.model = model
acc, loss = trainer.test(False)
print('InDims:', sum(indims), 'Acc:', acc,
'Loss:', loss, 'for', model.name,
'at threshold:', eval_thresh)
model_names.append(model.name)
accs.append(acc)
losses.append(loss)
dims.append(sum(indims))
inference_thresholds.append(eval_thresh)
print('Automatized experiment schedule enabled using', args.json_file)
config_dict = json.load(open(args.json_file, 'r'))
thresholds = [.99] if not 'threshs' in config_dict else config_dict['threshs']
dss = config_dict['dataset'] if isinstance(config_dict['dataset'], list) else [config_dict['dataset']]
optimizer = config_dict['optimizer']
run_num = 0
print(thresholds)
for dataset in dss:
for thresh in thresholds:
for batch_size in config_dict['batch_sizes']:
for model in config_dict['models']:
run_num += 1
print('Running Experiment', run_num, 'of', len(config_dict['batch_sizes'])*len(config_dict['models']*len(thresholds))*len(dss))
train_loader, test_loader, shape, num_classes = parse_dataset(dataset, batch_size)
model = parse_model(model, shape, num_classes)
change_all_pca_layer_thresholds(thresh, model, verbose=True)
if 'centering' in config_dict:
change_all_pca_layer_centering(centering=config_dict['centering'], network=model, verbose=True)
else:
change_all_pca_layer_centering(centering=False, network=model, verbose=True)
conv_method = 'channelwise' if 'conv_method' not in config_dict else config_dict['conv_method']
trainer = Trainer(model,
train_loader,
test_loader,
logs_dir=args.output,
device=args.device,
run_id=args.run_id,
epochs=config_dict['epochs'],
batch_size=batch_size,
optimizer=optimizer,
plot=True,
device = 'cuda:0'
torch.cuda.set_device(device)
cae = TinyCAE().to(device)
thresh = 10.0
_, test, _, _ = Food101(1, no_norm=True, shuffle_test=True)
path = './GoldFood101Samples'
if not exists(path):
mkdir(path=path)
cae.load_state_dict(torch.load(state_path)['model_state_dict'])
cae.eval()
change_all_pca_layer_thresholds(thresh, cae, verbose=True)
cae.to(device)
for i,(data, _) in enumerate(test):
data = data.to(device)
for thresh in [10.0, 0.9999, 0.999, 0.995, 0.99, 0.95, 0.9, 0.5]:
change_all_pca_layer_thresholds(thresh, cae)
d = data.squeeze().cpu().numpy().swapaxes(0, -1).swapaxes(0, 1)
imshow(d)
savefig(join(path, f'{i}sample-thresh{99999}-{8192}f-{8192}c-{100}s.jpg'))
pred = cae(data)
p = pred.squeeze().cpu().detach().numpy().swapaxes(0, -1).swapaxes(0, 1)
#im = np.append(d, p, axis=1)
im = p
imshow(im)
#title(f'Reconstruction with PCA-Threshold {round(thresh*100, 2) if thresh != 10 else 100}%')
savefig(join(path, f'{i}sample-thresh{round(thresh*10000) if thresh != 10 else 10000}-{cae.pca_layer.reduced_transformation_matrix.shape[-1]}f-{8192}c-{round(cae.pca_layer.reduced_transformation_matrix.shape[-1] / 8192*100, 2)}s.jpg'))
print(thresholds)
for dataset in dss:
for thresh in thresholds:
for batch_size in config_dict['batch_sizes']:
for model in config_dict['models']:
run_num += 1
print(
'Running Experiment', run_num, 'of',
len(config_dict['batch_sizes']) *
len(config_dict['models'] * len(thresholds)) *
len(dss))
train_loader, test_loader, shape, num_classes = parse_dataset(
dataset, batch_size)
model = parse_model(model, shape, num_classes)
change_all_pca_layer_thresholds(thresh,
model,
verbose=True)
if 'centering' in config_dict:
change_all_pca_layer_centering(
centering=config_dict['centering'],
network=model,
verbose=True)
else:
change_all_pca_layer_centering(centering=False,
network=model,
verbose=True)
conv_method = 'channelwise' if 'conv_method' not in config_dict else config_dict[
'conv_method']
trainer = Trainer(
model,
train_loader,