def main(args):
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
my_experiment = experiment(args.name, args, "/data5/jlindsey/continual/results", args.commit)
writer = SummaryWriter(my_experiment.path + "tensorboard")
logger = logging.getLogger('experiment')
logger.setLevel(logging.INFO)
total_clases = 10
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
logger.info("Frozen layers = %s", " ".join(frozen_layers))
#
final_results_all = []
total_clases = [10, 50, 75, 100, 150, 200]
if args.twentyclass:
total_clases = [20, 50]
if args.fiveclass:
total_clases = [5]
for tot_class in total_clases:
avg_perf = 0.0
lr_list = [0.03]
for aoo in range(0, args.runs):
keep = np.random.choice(list(range(200)), tot_class, replace=False)
print('keep', keep)
if args.dataset == "omniglot":
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot", train=True, background=False), keep)
print('lenbefore', len(dataset.data))
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter_omni(dataset, False, classes=total_clases),
batch_size=1,
shuffle=args.iid, num_workers=2)
print("LEN", len(iterator_sorted), len(dataset.data))
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot", train=not args.test, background=False), keep)
iterator = torch.utils.data.DataLoader(dataset, batch_size=32,
shuffle=False, num_workers=1)
elif args.dataset == "CIFAR100":
keep = np.random.choice(list(range(50, 100)), tot_class)
dataset = utils.remove_classes(df.DatasetFactory.get_dataset(args.dataset, train=True), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter(dataset, False, classes=tot_class),
batch_size=16,
shuffle=args.iid, num_workers=2)
dataset = utils.remove_classes(df.DatasetFactory.get_dataset(args.dataset, train=False), keep)
iterator = torch.utils.data.DataLoader(dataset, batch_size=128,
shuffle=False, num_workers=1)
# sampler = ts.MNISTSampler(list(range(0, total_clases)), dataset)
#
print(args)
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
results_mem_size = {}
for mem_size in [args.memory]:
max_acc = -10
max_lr = -10
for lr in lr_list:
print(lr)
# for lr in [0.001, 0.0003, 0.0001, 0.00003, 0.00001]:
maml = torch.load(args.model, map_location='cpu')
if args.scratch:
config = mf.ModelFactory.get_model("na", args.dataset)
maml = learner.Learner(config)
# maml = MetaLearingClassification(args, config).to(device).net
maml = maml.to(device)
for name, param in maml.named_parameters():
param.learn = True
for name, param in maml.named_parameters():
# logger.info(name)
if name in frozen_layers:
# logger.info("Freeezing name %s", str(name))
param.learn = False
# logger.info(str(param.requires_grad))
else:
if args.reset:
w = nn.Parameter(torch.ones_like(param))
# logger.info("W shape = %s", str(len(w.shape)))
if len(w.shape) > 1:
torch.nn.init.kaiming_normal_(w)
else:
w = nn.Parameter(torch.zeros_like(param))
param.data = w
param.learn = True
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
'''
torch.nn.init.kaiming_normal_(maml.parameters()[-2])
w = nn.Parameter(torch.zeros_like(maml.parameters()[-1]))
maml.parameters()[-1].data = w
for n, a in maml.named_parameters():
n = n.replace(".", "_")
# logger.info("Name = %s", n)
if n == "vars_14":
w = nn.Parameter(torch.ones_like(a))
# logger.info("W shape = %s", str(w.shape))
torch.nn.init.kaiming_normal_(w)
a.data = w
if n == "vars_15":
w = nn.Parameter(torch.zeros_like(a))
a.data = w
'''
correct = 0
for img, target in iterator:
with torch.no_grad():
img = img.to(device)
target = target.to(device)
logits_q = maml(img, vars=None, bn_training=False, feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q, target).sum().item() / len(img)
logger.info("Pre-epoch accuracy %s", str(correct / len(iterator)))
filter_list = ["vars.0", "vars.1", "vars.2", "vars.3", "vars.4", "vars.5"]
logger.info("Filter list = %s", ",".join(filter_list))
list_of_names = list(
map(lambda x: x[1], list(filter(lambda x: x[0] not in filter_list, maml.named_parameters()))))
list_of_params = list(filter(lambda x: x.learn, maml.parameters()))
list_of_names = list(filter(lambda x: x[1].learn, maml.named_parameters()))
if args.scratch or args.no_freeze:
print("Empty filter list")
list_of_params = maml.parameters()
#
for x in list_of_names:
logger.info("Unfrozen layer = %s", str(x[0]))
opt = torch.optim.Adam(list_of_params, lr=lr)
fast_weights = maml.vars
if args.randomize_plastic_weights:
maml.randomize_plastic_weights()
if args.zero_plastic_weights:
maml.zero_plastic_weights()
for iter in range(0, args.epoch):
iter_count = 0
imgs = []
ys = []
for img, y in iterator_sorted:
#print(iter_count, y)
if iter_count % 15 >= args.shots:
iter_count += 1
continue
iter_count += 1
img = img.to(device)
y = y.to(device)
imgs.append(img)
ys.append(y)
if not args.batch_learning:
pred = maml(img, vars=fast_weights)
opt.zero_grad()
loss = F.cross_entropy(pred, y)
grad = torch.autograd.grad(loss, fast_weights)
# fast_weights = list(map(lambda p: p[1] - self.update_lr * p[0], zip(grad, self.net.parameters())))
if args.plastic_update:
fast_weights = list(
map(lambda p: p[1] - p[0] * p[2] if p[1].learn else p[1], zip(grad, fast_weights, maml.vars_plasticity)))
else:
fast_weights = list(
map(lambda p: p[1] - args.update_lr * p[0] if p[1].learn else p[1], zip(grad, fast_weights)))
for params_old, params_new in zip(maml.parameters(), fast_weights):
params_new.learn = params_old.learn
if args.batch_learning:
y = torch.cat(ys, 0)
img = torch.cat(imgs, 0)
pred = maml(img, vars=fast_weights)
opt.zero_grad()
loss = F.cross_entropy(pred, y)
grad = torch.autograd.grad(loss, fast_weights)
# fast_weights = list(map(lambda p: p[1] - self.update_lr * p[0], zip(grad, self.net.parameters())))
if args.plastic_update:
fast_weights = list(
map(lambda p: p[1] - p[0] * p[2] if p[1].learn else p[1], zip(grad, fast_weights, maml.vars_plasticity)))
else:
fast_weights = list(
map(lambda p: p[1] - args.update_lr * p[0] if p[1].learn else p[1], zip(grad, fast_weights)))
for params_old, params_new in zip(maml.parameters(), fast_weights):
params_new.learn = params_old.learn
#loss.backward()
#opt.step()
logger.info("Result after one epoch for LR = %f", lr)
correct = 0
for img, target in iterator:
img = img.to(device)
target = target.to(device)
logits_q = maml(img, vars=fast_weights, bn_training=False, feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q, target).sum().item() / len(img)
logger.info(str(correct / len(iterator)))
if (correct / len(iterator) > max_acc):
max_acc = correct / len(iterator)
max_lr = lr
lr_list = [max_lr]
results_mem_size[mem_size] = (max_acc, max_lr)
avg_perf += max_acc / args.runs
print('avg perf', avg_perf * args.runs / (1+aoo))
logger.info("Final Max Result = %s", str(max_acc))
writer.add_scalar('/finetune/best_' + str(aoo), max_acc, tot_class)
final_results_all.append((tot_class, results_mem_size))
print("A= ", results_mem_size)
logger.info("Final results = %s", str(results_mem_size))
my_experiment.results["Final Results"] = final_results_all
my_experiment.store_json()
np.save('evals/final_results_'+args.orig_name+'.npy', final_results_all)
print("FINAL RESULTS = ", final_results_all)
writer.close()
def main(args):
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
my_experiment = experiment(args.name, args, "../results/", args.commit)
writer = SummaryWriter(my_experiment.path + "tensorboard")
logger = logging.getLogger('experiment')
logger.setLevel(logging.INFO)
total_clases = 10
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
logger.info("Frozen layers = %s", " ".join(frozen_layers))
#for v in range(6):
# frozen_layers.append("vars_bn.{0}".format(v))
final_results_all = []
temp_result = []
total_clases = args.schedule
for tot_class in total_clases:
lr_list = [
0.001, 0.0006, 0.0004, 0.00035, 0.0003, 0.00025, 0.0002, 0.00015,
0.0001, 0.00009, 0.00008, 0.00006, 0.00003, 0.00001
]
lr_all = []
for lr_search in range(10):
keep = np.random.choice(list(range(650)), tot_class, replace=False)
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=True,
background=False,
path=args.dataset_path), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter_omni(dataset,
False,
classes=total_clases),
batch_size=1,
shuffle=args.iid,
num_workers=2)
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=not args.test,
background=False,
path=args.dataset_path), keep)
iterator = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=1)
print(args)
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
results_mem_size = {}
for mem_size in [args.memory]:
max_acc = -10
max_lr = -10
for lr in lr_list:
print(lr)
maml = torch.load(args.model, map_location='cpu')
if args.scratch:
config = mf.ModelFactory.get_model("OML", args.dataset)
maml = learner.Learner(config)
# maml = MetaLearingClassification(args, config).to(device).net
maml = maml.to(device)
for name, param in maml.named_parameters():
param.learn = True
for name, param in maml.named_parameters():
# logger.info(name)
if name in frozen_layers:
param.learn = False
else:
if args.reset:
w = nn.Parameter(torch.ones_like(param))
# logger.info("W shape = %s", str(len(w.shape)))
if len(w.shape) > 1:
torch.nn.init.kaiming_normal_(w)
else:
w = nn.Parameter(torch.zeros_like(param))
param.data = w
param.learn = True
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
torch.nn.init.kaiming_normal_(maml.parameters()[-2])
w = nn.Parameter(torch.zeros_like(maml.parameters()[-1]))
maml.parameters()[-1].data = w
if args.neuromodulation:
weights2reset = ["vars_26"]
biases2reset = ["vars_27"]
else:
weights2reset = ["vars_14"]
biases2reset = ["vars_15"]
for n, a in maml.named_parameters():
n = n.replace(".", "_")
if n in weights2reset:
w = nn.Parameter(torch.ones_like(a)).to(device)
torch.nn.init.kaiming_normal_(w)
a.data = w
if n in biases2reset:
w = nn.Parameter(torch.zeros_like(a)).to(device)
a.data = w
filter_list = ["vars.{0}".format(v) for v in range(6)]
logger.info("Filter list = %s", ",".join(filter_list))
list_of_names = list(
map(
lambda x: x[1],
list(
filter(lambda x: x[0] not in filter_list,
maml.named_parameters()))))
list_of_params = list(
filter(lambda x: x.learn, maml.parameters()))
list_of_names = list(
filter(lambda x: x[1].learn, maml.named_parameters()))
if args.scratch or args.no_freeze:
print("Empty filter list")
list_of_params = maml.parameters()
for x in list_of_names:
logger.info("Unfrozen layer = %s", str(x[0]))
opt = torch.optim.Adam(list_of_params, lr=lr)
for _ in range(0, args.epoch):
for img, y in iterator_sorted:
img = img.to(device)
y = y.to(device)
pred = maml(img)
opt.zero_grad()
loss = F.cross_entropy(pred, y)
loss.backward()
opt.step()
logger.info("Result after one epoch for LR = %f", lr)
correct = 0
for img, target in iterator:
img = img.to(device)
target = target.to(device)
logits_q = maml(img,
vars=None,
bn_training=False,
feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q,
target).sum().item() / len(img)
logger.info(str(correct / len(iterator)))
if (correct / len(iterator) > max_acc):
max_acc = correct / len(iterator)
max_lr = lr
lr_all.append(max_lr)
results_mem_size[mem_size] = (max_acc, max_lr)
logger.info("Final Max Result = %s", str(max_acc))
writer.add_scalar('/finetune/best_' + str(lr_search), max_acc,
tot_class)
temp_result.append((tot_class, results_mem_size))
print("A= ", results_mem_size)
logger.info("Temp Results = %s", str(results_mem_size))
my_experiment.results["Temp Results"] = temp_result
my_experiment.store_json()
print("LR RESULTS = ", temp_result)
from scipy import stats
best_lr = float(stats.mode(lr_all)[0][0])
logger.info("BEST LR %s= ", str(best_lr))
for aoo in range(args.runs):
keep = np.random.choice(list(range(650)), tot_class, replace=False)
if args.dataset == "omniglot":
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=True,
background=False), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter_omni(dataset,
False,
classes=total_clases),
batch_size=1,
shuffle=args.iid,
num_workers=2)
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=not args.test,
background=False), keep)
iterator = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=1)
elif args.dataset == "CIFAR100":
keep = np.random.choice(list(range(50, 100)), tot_class)
dataset = utils.remove_classes(
df.DatasetFactory.get_dataset(args.dataset, train=True),
keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter(dataset, False, classes=tot_class),
batch_size=16,
shuffle=args.iid,
num_workers=2)
dataset = utils.remove_classes(
df.DatasetFactory.get_dataset(args.dataset, train=False),
keep)
iterator = torch.utils.data.DataLoader(dataset,
batch_size=128,
shuffle=False,
num_workers=1)
print(args)
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
results_mem_size = {}
for mem_size in [args.memory]:
max_acc = -10
max_lr = -10
lr = best_lr
maml = torch.load(args.model, map_location='cpu')
if args.scratch:
config = mf.ModelFactory.get_model("MRCL", args.dataset)
maml = learner.Learner(config)
maml = maml.to(device)
for name, param in maml.named_parameters():
param.learn = True
for name, param in maml.named_parameters():
# logger.info(name)
if name in frozen_layers:
param.learn = False
else:
if args.reset:
w = nn.Parameter(torch.ones_like(param))
if len(w.shape) > 1:
torch.nn.init.kaiming_normal_(w)
else:
w = nn.Parameter(torch.zeros_like(param))
param.data = w
param.learn = True
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
torch.nn.init.kaiming_normal_(maml.parameters()[-2])
w = nn.Parameter(torch.zeros_like(maml.parameters()[-1]))
maml.parameters()[-1].data = w
for n, a in maml.named_parameters():
n = n.replace(".", "_")
if args.neuromodulation:
weights2reset = ["vars_26"]
biases2reset = ["vars_27"]
else:
weights2reset = ["vars_14"]
biases2reset = ["vars_15"]
for n, a in maml.named_parameters():
n = n.replace(".", "_")
if n in weights2reset:
w = nn.Parameter(torch.ones_like(a)).to(device)
torch.nn.init.kaiming_normal_(w)
a.data = w
if n in biases2reset:
w = nn.Parameter(torch.zeros_like(a)).to(device)
a.data = w
correct = 0
for img, target in iterator:
with torch.no_grad():
img = img.to(device)
target = target.to(device)
logits_q = maml(img,
vars=None,
bn_training=False,
feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q,
target).sum().item() / len(img)
logger.info("Pre-epoch accuracy %s",
str(correct / len(iterator)))
filter_list = ["vars.{0}".format(v) for v in range(6)]
logger.info("Filter list = %s", ",".join(filter_list))
list_of_names = list(
map(
lambda x: x[1],
list(
filter(lambda x: x[0] not in filter_list,
maml.named_parameters()))))
list_of_params = list(
filter(lambda x: x.learn, maml.parameters()))
list_of_names = list(
filter(lambda x: x[1].learn, maml.named_parameters()))
if args.scratch or args.no_freeze:
print("Empty filter list")
list_of_params = maml.parameters()
for x in list_of_names:
logger.info("Unfrozen layer = %s", str(x[0]))
opt = torch.optim.Adam(list_of_params, lr=lr)
for _ in range(0, args.epoch):
for img, y in iterator_sorted:
img = img.to(device)
y = y.to(device)
pred = maml(img)
opt.zero_grad()
loss = F.cross_entropy(pred, y)
loss.backward()
opt.step()
logger.info("Result after one epoch for LR = %f", lr)
correct = 0
for img, target in iterator:
img = img.to(device)
target = target.to(device)
logits_q = maml(img,
vars=None,
bn_training=False,
feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q, target).sum().item() / len(img)
logger.info(str(correct / len(iterator)))
if (correct / len(iterator) > max_acc):
max_acc = correct / len(iterator)
max_lr = lr
lr_list = [max_lr]
results_mem_size[mem_size] = (max_acc, max_lr)
logger.info("Final Max Result = %s", str(max_acc))
writer.add_scalar('/finetune/best_' + str(aoo), max_acc,
tot_class)
final_results_all.append((tot_class, results_mem_size))
print("A= ", results_mem_size)
logger.info("Final results = %s", str(results_mem_size))
my_experiment.results["Final Results"] = final_results_all
my_experiment.store_json()
print("FINAL RESULTS = ", final_results_all)
writer.close()
def main(args):
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
my_experiment = experiment(args.name, args, "../results/", args.commit)
writer = SummaryWriter(my_experiment.path + "tensorboard")
logger = logging.getLogger('experiment')
logger.setLevel(logging.INFO)
total_clases = 10
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
logger.info("Frozen layers = %s", " ".join(frozen_layers))
#
final_results_all = []
total_clases = [10, 50, 75, 100, 150, 200]
for tot_class in total_clases:
lr_list = [0.03, 0.01, 0.003, 0.001, 0.0003, 0.0001, 0.00003, 0.00001]
for aoo in range(0, 20):
keep = np.random.choice(list(range(200)), tot_class)
if args.dataset == "omniglot":
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=True,
background=False), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter_omni(dataset,
False,
classes=total_clases),
batch_size=1,
shuffle=args.iid,
num_workers=2)
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot",
train=not args.test,
background=False), keep)
iterator = torch.utils.data.DataLoader(dataset,
batch_size=32,
shuffle=False,
num_workers=1)
elif args.dataset == "CIFAR100":
keep = np.random.choice(list(range(50, 100)), tot_class)
dataset = utils.remove_classes(
df.DatasetFactory.get_dataset(args.dataset, train=True),
keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter(dataset, False, classes=tot_class),
batch_size=16,
shuffle=args.iid,
num_workers=2)
dataset = utils.remove_classes(
df.DatasetFactory.get_dataset(args.dataset, train=False),
keep)
iterator = torch.utils.data.DataLoader(dataset,
batch_size=128,
shuffle=False,
num_workers=1)
# sampler = ts.MNISTSampler(list(range(0, total_clases)), dataset)
#
print(args)
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
results_mem_size = {}
for mem_size in [args.memory]:
max_acc = -10
max_lr = -10
for lr in lr_list:
print(lr)
# for lr in [0.001, 0.0003, 0.0001, 0.00003, 0.00001]:
maml = torch.load(args.model, map_location='cpu')
if args.scratch:
config = mf.ModelFactory.get_model("na", args.dataset)
maml = learner.Learner(config)
# maml = MetaLearingClassification(args, config).to(device).net
maml = maml.to(device)
for name, param in maml.named_parameters():
param.learn = True
for name, param in maml.named_parameters():
# logger.info(name)
if name in frozen_layers:
# logger.info("Freeezing name %s", str(name))
param.learn = False
# logger.info(str(param.requires_grad))
else:
if args.reset:
w = nn.Parameter(torch.ones_like(param))
# logger.info("W shape = %s", str(len(w.shape)))
if len(w.shape) > 1:
torch.nn.init.kaiming_normal_(w)
else:
w = nn.Parameter(torch.zeros_like(param))
param.data = w
param.learn = True
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
torch.nn.init.kaiming_normal_(maml.parameters()[-2])
w = nn.Parameter(torch.zeros_like(maml.parameters()[-1]))
maml.parameters()[-1].data = w
for n, a in maml.named_parameters():
n = n.replace(".", "_")
# logger.info("Name = %s", n)
if n == "vars_14":
w = nn.Parameter(torch.ones_like(a))
# logger.info("W shape = %s", str(w.shape))
torch.nn.init.kaiming_normal_(w)
a.data = w
if n == "vars_15":
w = nn.Parameter(torch.zeros_like(a))
a.data = w
correct = 0
for img, target in iterator:
with torch.no_grad():
img = img.to(device)
target = target.to(device)
logits_q = maml(img,
vars=None,
bn_training=False,
feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q,
target).sum().item() / len(img)
logger.info("Pre-epoch accuracy %s",
str(correct / len(iterator)))
filter_list = [
"vars.0", "vars.1", "vars.2", "vars.3", "vars.4",
"vars.5"
]
logger.info("Filter list = %s", ",".join(filter_list))
list_of_names = list(
map(
lambda x: x[1],
list(
filter(lambda x: x[0] not in filter_list,
maml.named_parameters()))))
list_of_params = list(
filter(lambda x: x.learn, maml.parameters()))
list_of_names = list(
filter(lambda x: x[1].learn, maml.named_parameters()))
if args.scratch or args.no_freeze:
print("Empty filter list")
list_of_params = maml.parameters()
#
for x in list_of_names:
logger.info("Unfrozen layer = %s", str(x[0]))
opt = torch.optim.Adam(list_of_params, lr=lr)
import module.replay as rep
res_sampler = rep.ReservoirSampler(mem_size)
for _ in range(0, args.epoch):
for img, y in iterator_sorted:
if mem_size > 0:
res_sampler.update_buffer(zip(img, y))
res_sampler.update_observations(len(img))
img = img.to(device)
y = y.to(device)
img2, y2 = res_sampler.sample_buffer(16)
img2 = img2.to(device)
y2 = y2.to(device)
img = torch.cat([img, img2], dim=0)
y = torch.cat([y, y2], dim=0)
else:
img = img.to(device)
y = y.to(device)
pred = maml(img)
opt.zero_grad()
loss = F.cross_entropy(pred, y)
loss.backward()
opt.step()
logger.info("Result after one epoch for LR = %f", lr)
correct = 0
for img, target in iterator:
img = img.to(device)
target = target.to(device)
logits_q = maml(img,
vars=None,
bn_training=False,
feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q,
target).sum().item() / len(img)
logger.info(str(correct / len(iterator)))
if (correct / len(iterator) > max_acc):
max_acc = correct / len(iterator)
max_lr = lr
lr_list = [max_lr]
results_mem_size[mem_size] = (max_acc, max_lr)
logger.info("Final Max Result = %s", str(max_acc))
writer.add_scalar('/finetune/best_' + str(aoo), max_acc,
tot_class)
# quit()
final_results_all.append((tot_class, results_mem_size))
print("A= ", results_mem_size)
logger.info("Final results = %s", str(results_mem_size))
my_experiment.results["Final Results"] = final_results_all
my_experiment.store_json()
print("FINAL RESULTS = ", final_results_all)
writer.close()
def main(args):
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
my_experiment = experiment(args.name, args, "./evals/", args.commit)
writer = SummaryWriter(my_experiment.path + "tensorboard")
ver = 0
while os.path.exists(args.modelX + "_" + str(ver)):
ver += 1
args.modelX = args.modelX + "_" + str(ver-1) + "/learner.model"
logger = logging.getLogger('experiment')
logger.setLevel(logging.INFO)
total_clases = 10
total_ff_vars = 2*(6 + 2 + args.num_extra_dense_layers)
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
for temp in range(args.rln_end * 2):
frozen_layers.append("net.vars." + str(total_ff_vars - 1 - temp))
#logger.info("Frozen layers = %s", " ".join(frozen_layers))
#
final_results_all = []
total_clases = [5]
if args.twentyclass:
total_clases = [20]
if args.twotask:
total_clases = [2, 10]
if args.fiftyclass:
total_clases = [50]
if args.tenclass:
total_clases = [10]
if args.fiveclass:
total_clases = [5]
print('yooo', total_clases)
for tot_class in total_clases:
avg_perf = 0.0
print('TOT_CLASS', tot_class)
lr_list = [0]#[0.03, 0.01, 0.003, 0.001, 0.0003, 0.0001, 0.00003, 0.00001]
for aoo in range(0, args.runs):
#print('run', aoo)
keep = np.random.choice(list(range(650)), tot_class, replace=False)
if args.dataset == "imagenet":
keep = np.random.choice(list(range(20)), tot_class, replace=False)
dataset = imgnet.MiniImagenet(args.imagenet_path, mode='test', elem_per_class=30, classes=keep, seed=aoo)
dataset_test = imgnet.MiniImagenet(args.imagenet_path, mode='test', elem_per_class=30, classes=keep, test=args.test, seed=aoo)
iterator = torch.utils.data.DataLoader(dataset_test, batch_size=128,
shuffle=True, num_workers=1)
iterator_sorted = torch.utils.data.DataLoader(dataset, batch_size=1,
shuffle=False, num_workers=1)
if args.dataset == "omniglot":
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot", train=True, background=False), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter_omni(dataset, False, classes=total_clases),
batch_size=1,
shuffle=False, num_workers=2)
dataset = utils.remove_classes_omni(
df.DatasetFactory.get_dataset("omniglot", train=not args.test, background=False), keep)
iterator = torch.utils.data.DataLoader(dataset, batch_size=32,
shuffle=False, num_workers=1)
elif args.dataset == "CIFAR100":
keep = np.random.choice(list(range(50, 100)), tot_class)
dataset = utils.remove_classes(df.DatasetFactory.get_dataset(args.dataset, train=True), keep)
iterator_sorted = torch.utils.data.DataLoader(
utils.iterator_sorter(dataset, False, classes=tot_class),
batch_size=16,
shuffle=False, num_workers=2)
dataset = utils.remove_classes(df.DatasetFactory.get_dataset(args.dataset, train=False), keep)
iterator = torch.utils.data.DataLoader(dataset, batch_size=128,
shuffle=False, num_workers=1)
# sampler = ts.MNISTSampler(list(range(0, total_clases)), dataset)
#
#print(args)
if torch.cuda.is_available():
device = torch.device('cuda')
else:
device = torch.device('cpu')
results_mem_size = {}
#print("LEN", len(iterator_sorted))
for mem_size in [args.memory]:
max_acc = -10
max_lr = -10
for lr in lr_list:
#torch.cuda.empty_cache()
#print(lr)
# for lr in [0.001, 0.0003, 0.0001, 0.00003, 0.00001]:
maml = torch.load(args.modelX, map_location='cpu')
if args.scratch:
config = mf.ModelFactory.get_model(args.model_type, args.dataset)
maml = learner.Learner(config, lr)
# maml = MetaLearingClassification(args, config).to(device).net
#maml.update_lr = lr
maml = maml.to(device)
for name, param in maml.named_parameters():
param.learn = True
for name, param in maml.named_parameters():
#if name.find("feedback_strength_vars") != -1:
# print(name, param)
if name in frozen_layers:
# logger.info("Freeezing name %s", str(name))
param.learn = False
# logger.info(str(param.requires_grad))
else:
if args.reset:
w = nn.Parameter(torch.ones_like(param))
# logger.info("W shape = %s", str(len(w.shape)))
if len(w.shape) > 1:
torch.nn.init.kaiming_normal_(w)
else:
w = nn.Parameter(torch.zeros_like(param))
param.data = w
param.learn = True
frozen_layers = []
for temp in range(args.rln * 2):
frozen_layers.append("vars." + str(temp))
#torch.nn.init.kaiming_normal_(maml.parameters()[-2])
#w = nn.Parameter(torch.zeros_like(maml.parameters()[-1]))
#maml.parameters()[-1].data = w
for n, a in maml.named_parameters():
n = n.replace(".", "_")
# logger.info("Name = %s", n)
if n == "vars_"+str(14+2*args.num_extra_dense_layers):
pass
#w = nn.Parameter(torch.ones_like(a))
# logger.info("W shape = %s", str(w.shape))
#torch.nn.init.kaiming_normal_(w)
#a.data = w
if n == "vars_"+str(15+2*args.num_extra_dense_layers):
pass
#w = nn.Parameter(torch.zeros_like(a))
#a.data = w
#for fv in maml.feedback_vars:
# w = nn.Parameter(torch.zeros_like(fv))
# fv.data = w
#for fv in maml.feedback_strength_vars:
# w = nn.Parameter(torch.ones_like(fv))
# fv.data = w
correct = 0
for img, target in iterator:
#print('size', target.size())
target = torch.tensor(np.array([list(keep).index(int(target.cpu().numpy()[i])) for i in range(target.size()[0])]))
with torch.no_grad():
img = img.to(device)
target = target.to(device)
logits_q = maml(img, vars=None, bn_training=False, feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q, target).sum().item() / len(img)
#logger.info("Pre-epoch accuracy %s", str(correct / len(iterator)))
filter_list = ["vars.0", "vars.1", "vars.2", "vars.3", "vars.4", "vars.5"]
#logger.info("Filter list = %s", ",".join(filter_list))
list_of_names = list(
map(lambda x: x[1], list(filter(lambda x: x[0] not in filter_list, maml.named_parameters()))))
list_of_params = list(filter(lambda x: x.learn, maml.parameters()))
list_of_names = list(filter(lambda x: x[1].learn, maml.named_parameters()))
if args.scratch or args.no_freeze:
print("Empty filter list")
list_of_params = maml.parameters()
#
#for x in list_of_names:
# logger.info("Unfrozen layer = %s", str(x[0]))
opt = torch.optim.Adam(list_of_params, lr=lr)
fast_weights = None
if args.randomize_plastic_weights:
maml.randomize_plastic_weights()
if args.zero_plastic_weights:
maml.zero_plastic_weights()
res_sampler = rep.ReservoirSampler(mem_size)
iterator_sorted_new = []
iter_count = 0
for img, y in iterator_sorted:
y = torch.tensor(np.array([list(keep).index(int(y.cpu().numpy()[i])) for i in range(y.size()[0])]))
if iter_count % 15 >= args.shots:
iter_count += 1
continue
iterator_sorted_new.append((img, y))
iter_count += 1
iterator_sorted = []
perm = np.random.permutation(len(iterator_sorted_new))
for i in range(len(iterator_sorted_new)):
if args.iid:
iterator_sorted.append(iterator_sorted_new[perm[i]])
else:
iterator_sorted.append(iterator_sorted_new[i])
for iter in range(0, args.epoch):
iter_count = 0
imgs = []
ys = []
for img, y in iterator_sorted:
#print('iter count', iter_count)
#print('y is', y)
#if iter_count % 15 >= args.shots:
# iter_count += 1
# continue
iter_count += 1
#with torch.no_grad():
if args.memory == 0:
img = img.to(device)
y = y.to(device)
else:
res_sampler.update_buffer(zip(img, y))
res_sampler.update_observations(len(img))
img = img.to(device)
y = y.to(device)
img2, y2 = res_sampler.sample_buffer(8)
img2 = img2.to(device)
y2 = y2.to(device)
img = torch.cat([img, img2], dim=0)
y = torch.cat([y, y2], dim=0)
#print('img size', img.size())
imgs.append(img)
ys.append(y)
if not args.batch_learning:
logits = maml(img, vars=fast_weights)
fast_weights = maml.getOjaUpdate(y, logits, fast_weights, hebbian=args.hebb)
if args.batch_learning:
y = torch.cat(ys, 0)
img = torch.cat(imgs, 0)
logits = maml(img, vars=fast_weights)
fast_weights = maml.getOjaUpdate(y, logits, fast_weights, hebbian=args.hebb)
#logger.info("Result after one epoch for LR = %f", lr)
correct = 0
for img, target in iterator:
target = torch.tensor(np.array([list(keep).index(int(target.cpu().numpy()[i])) for i in range(target.size()[0])]))
img = img.to(device)
target = target.to(device)
logits_q = maml(img, vars=fast_weights, bn_training=False, feature=False)
pred_q = (logits_q).argmax(dim=1)
correct += torch.eq(pred_q, target).sum().item() / len(img)
#logger.info(str(correct / len(iterator)))
if (correct / len(iterator) > max_acc):
max_acc = correct / len(iterator)
max_lr = lr
del maml
#del maml
#del fast_weights
lr_list = [max_lr]
#print('result', max_acc)
results_mem_size[mem_size] = (max_acc, max_lr)
#logger.info("Final Max Result = %s", str(max_acc))
writer.add_scalar('/finetune/best_' + str(aoo), max_acc, tot_class)
avg_perf += max_acc / args.runs #TODO: change this if/when I ever use memory -- can't choose max memory size differently for each run!
print('avg perf', avg_perf * args.runs / (1+aoo))
final_results_all.append((tot_class, results_mem_size))
#writer.add_scalar('performance', avg_perf, tot_class)
#print("A= ", results_mem_size)
#logger.info("Final results = %s", str(results_mem_size))
my_experiment.results["Final Results"] = final_results_all
my_experiment.store_json()
np.save('evals/final_results_'+args.orig_name+'.npy', final_results_all)
#print("FINAL RESULTS = ", final_results_all)
writer.close()