def test_determinism(self):
"""Test whether training and scoring is deterministic given seed"""
em = EndModel(
seed=123,
batchnorm=True,
dropout=0.1,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model((Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=1,
checkpoint=False)
score_1 = em.score((Xs[2], Ys[2]), verbose=False)
# Test scoring determinism
score_2 = em.score((Xs[2], Ys[2]), verbose=False)
self.assertEqual(score_1, score_2)
# Test training determinism
em_2 = EndModel(
seed=123,
batchnorm=True,
dropout=0.1,
layer_out_dims=[2, 10, 2],
verbose=False,
)
em_2.train_model((Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=1,
checkpoint=False)
score_3 = em_2.score((Xs[2], Ys[2]), verbose=False)
self.assertEqual(score_1, score_3)
def test_lstm_memorize_marker(self):
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = torch.zeros(n).long()
needles = np.random.randint(1, SEQ_LEN - 1, n)
for i in range(n):
X[i, needles[i]] = MAX_INT + 1
Y[i] = X[i, needles[i] + 1]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 2
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size,
bidirectional=True,
verbose=False,
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size * 2, MAX_INT],
batchnorm=True,
seed=1,
verbose=False,
)
em.train_model(Xs[0], Ys[0], Xs[1], Ys[1], n_epochs=10, verbose=False)
score = em.score(Xs[2], Ys[2], verbose=False)
self.assertGreater(score, 0.95)
def test_lstm_memorize_first(self):
"""Confirm that lstm can memorize the first token in a long sequence"""
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = X[:, 0]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 1
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size=vocab_size,
bidirectional=False,
verbose=False,
lstm_reduction="attention",
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size, MAX_INT],
optimizer="adam",
batchnorm=True,
seed=1,
verbose=False,
)
em.train_model((Xs[0], Ys[0]), dev_data=(Xs[1], Ys[1]), n_epochs=10)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def test_lstm_memorize_first(self):
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = X[:, 0]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 1
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size,
bidirectional=False,
verbose=False,
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size, MAX_INT],
optimizer="adam",
batchnorm=True,
seed=1,
verbose=False,
)
em.train_model(Xs[0], Ys[0], Xs[1], Ys[1], n_epochs=5, verbose=True)
score = em.score(Xs[2], Ys[2], verbose=False)
self.assertGreater(score, 0.95)
def test_logwriter(self):
"""Test the basic LogWriter class"""
writer_kwargs = {
"log_dir": self.log_dir,
"run_dir": "test_dir",
"run_name": "test",
}
em = EndModel(
seed=1,
input_batchnorm=False,
middle_batchnorm=False,
input_dropout=0.0,
middle_dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model(
(Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=7,
checkpoint=False,
writer="json",
**writer_kwargs,
)
# Load the log
with open(em.writer.log_path, "r") as f:
log_dict = json.load(f)
self.assertEqual(log_dict["config"]["train_config"]["n_epochs"], 7)
self.assertEqual(len(log_dict["run_log"]["train/loss"]), 7)
def test_logging(self):
"""Test the basic LogWriter class"""
log_writer = LogWriter(run_dir="test_dir", run_name="test")
em = EndModel(
seed=1,
input_batchnorm=False,
middle_batchnorm=False,
input_dropout=0.0,
middle_dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model(
(Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=7,
log_writer=log_writer,
)
# Load the log
with open(log_writer.log_path, "r") as f:
run_log = json.load(f)
self.assertEqual(run_log["config"]["train_config"]["n_epochs"], 7)
self.assertEqual(len(run_log["run-log"]["train-loss"]), 7)
# Clean up
rmtree(log_writer.log_subdir)
def test_checkpointing(self):
"""Confirm that different checkpoints are being saved with checkpoint_every on"""
em = EndModel(
seed=1,
batchnorm=False,
dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model(
(Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=5,
checkpoint=True,
checkpoint_every=1,
)
test_model = copy.deepcopy(em.state_dict())
new_model = torch.load("checkpoints/model_checkpoint_4.pth")
self.assertFalse(
torch.all(
torch.eq(
test_model["network.1.0.weight"],
new_model["model"]["network.1.0.weight"],
)))
new_model = torch.load("checkpoints/model_checkpoint_5.pth")
self.assertTrue(
torch.all(
torch.eq(
test_model["network.1.0.weight"],
new_model["model"]["network.1.0.weight"],
)))
def test_save_and_load(self):
"""Test basic saving and loading"""
em = EndModel(
seed=1337,
input_batchnorm=False,
middle_batchnorm=False,
input_dropout=0.0,
middle_dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model((Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=3,
checkpoint=False)
score = em.score((Xs[2], Ys[2]), verbose=False)
# Save model
SAVE_PATH = "test_save_model.pkl"
em.save(SAVE_PATH)
# Reload and make sure (a) score and (b) non-buffer, non-Parameter
# attributes are the same
em_2 = EndModel.load(SAVE_PATH)
self.assertEqual(em.seed, em_2.seed)
score_2 = em_2.score((Xs[2], Ys[2]), verbose=False)
self.assertEqual(score, score_2)
# Clean up
os.remove(SAVE_PATH)
def test_scoring(self):
"""Test the metrics whole way through"""
em = EndModel(
seed=1,
batchnorm=False,
dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model((Xs[0], Ys[0]), dev_data=(Xs[1], Ys[1]), n_epochs=5)
metrics = list(METRICS.keys())
scores = em.score((Xs[2], Ys[2]), metric=metrics, verbose=True)
for i, metric in enumerate(metrics):
self.assertGreater(scores[i], 0.95)
def test_singletask(self):
"""Test basic single-task end model"""
em = EndModel(
seed=1,
input_batchnorm=False,
middle_batchnorm=False,
input_dropout=0.0,
middle_dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model((Xs[0], Ys[0]), dev_data=(Xs[1], Ys[1]), n_epochs=5)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def test_lstm_embeddings_freeze(self):
"""Confirm that if embeddings are frozen, they do not change during training"""
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = torch.zeros(n).long()
needles = np.random.randint(1, SEQ_LEN - 1, n)
for i in range(n):
X[i, needles[i]] = MAX_INT + 1
Y[i] = X[i, needles[i] + 1]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 2
for freeze_embs in [True, False]:
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size=vocab_size,
freeze=freeze_embs,
verbose=False,
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size * 2, MAX_INT],
verbose=False,
)
before = lstm_module.embeddings.weight.clone()
em.train_model(
(Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=15,
verbose=False,
)
after = lstm_module.embeddings.weight.clone()
if freeze_embs:
self.assertEqual(torch.abs(before - after).sum().item(), 0.0)
else:
self.assertNotEqual(
torch.abs(before - after).sum().item(), 0.0)
def test_singletask_extras(self):
"""Test batchnorm and dropout"""
em = EndModel(
seed=1,
input_batchnorm=True,
middle_batchnorm=True,
input_dropout=0.01,
middle_dropout=0.01,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model((Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=5,
checkpoint=False)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def test_gpustorage(self):
# Running basics tutorial problem
with open("tutorials/data/basics_tutorial.pkl", "rb") as f:
X, Y, L, D = pickle.load(f)
Xs, Ys, Ls, Ds = split_data(X,
Y,
L,
D,
splits=[0.8, 0.1, 0.1],
stratify_by=Y,
seed=123)
label_model = LabelModel(k=2, seed=123)
label_model.train_model(Ls[0],
Y_dev=Ys[1],
n_epochs=500,
log_train_every=25)
Y_train_ps = label_model.predict_proba(Ls[0])
# Creating a really large end model to use lots of memory
end_model = EndModel([1000, 100000, 2], seed=123, device="cuda")
# Getting initial GPU storage use
initial_gpu_mem = GPUtil.getGPUs()[0].memoryUsed
# Training model
end_model.train_model(
(Xs[0], Y_train_ps),
valid_data=(Xs[1], Ys[1]),
l2=0.1,
batch_size=256,
n_epochs=3,
log_train_every=1,
validation_metric="f1",
)
# Final GPU storage use
final_gpu_mem = GPUtil.getGPUs()[0].memoryUsed
# On a Titan X, this model uses ~ 3 GB of memory
gpu_mem_difference = final_gpu_mem - initial_gpu_mem
self.assertGreater(gpu_mem_difference, 1000)
def test_lstm_direct_features(self):
"""Confirm that lstm can work over features passed in directly (rather
than embedded)."""
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = X[:, 0]
# Convert X to one-hot features
Xf = torch.zeros((n, SEQ_LEN, MAX_INT)).long()
for i in range(n):
for j in range(SEQ_LEN):
Xf[i, j, X[i, j] - 1] = 1
X = Xf
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = MAX_INT
hidden_size = 10
lstm_module = LSTMModule(
embed_size,
hidden_size,
skip_embeddings=True, # This is where we configure for this setting
bidirectional=False,
verbose=False,
lstm_reduction="attention",
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size, MAX_INT],
optimizer="adam",
batchnorm=True,
seed=1,
verbose=False,
)
em.train_model((Xs[0], Ys[0]), dev_data=(Xs[1], Ys[1]), n_epochs=15)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def test_resume_training(self):
"""Confirm that a checkpoint can be saved and reloaded without throwing error"""
em = EndModel(
seed=1,
batchnorm=False,
dropout=0.0,
layer_out_dims=[2, 10, 2],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model(
(Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
n_epochs=5,
checkpoint=True,
checkpoint_every=1,
)
em.resume_training(
(Xs[0], Ys[0]),
valid_data=(Xs[1], Ys[1]),
model_path="checkpoints/model_checkpoint_2.pth",
)
def test_lstm_memorize_marker(self):
"""Confirm that lstm can return the token that comes after a special marker"""
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = torch.zeros(n).long()
needles = np.random.randint(1, SEQ_LEN - 1, n)
for i in range(n):
X[i, needles[i]] = MAX_INT + 1
Y[i] = X[i, needles[i] + 1]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 2
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size=vocab_size,
bidirectional=True,
verbose=False,
lstm_reduction="attention",
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size * 2, MAX_INT],
batchnorm=True,
seed=1,
verbose=False,
)
em.train_model((Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=15,
verbose=False)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def test_custom_modules(self):
"""Test custom input/head modules"""
input_module = nn.Sequential(IdentityModule(), nn.Linear(2, 10))
middle_modules = [nn.Linear(10, 8), IdentityModule()]
head_module = nn.Sequential(nn.Linear(8, 2), IdentityModule())
em = EndModel(
seed=1,
input_module=input_module,
middle_modules=middle_modules,
head_module=head_module,
layer_out_dims=[10, 8, 8],
verbose=False,
)
Xs, Ys = self.single_problem
em.train_model(
(Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=5,
verbose=False,
show_plots=False,
)
score = em.score((Xs[2], Ys[2]), verbose=False)
self.assertGreater(score, 0.95)
def train_model(args):
# Create datasets and dataloaders
train, dev, test = load_dataset(args)
data_loader = get_data_loader(train, dev, test, args.batch_size,
args.num_workers)
hidden_size = 128
num_classes = 2
encode_dim = 1000 # using get_frm_output_size()
# Define input encoder
cnn_encoder = FrameEncoderOC
if (torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
#import ipdb; ipdb.set_trace()
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction=args.lstm_reduction,
encoder_class=cnn_encoder,
)
init_kwargs = {
"layer_out_dims": [hidden_size, num_classes],
"input_module": lstm_module,
"optimizer": "adam",
"verbose": False,
"input_batchnorm": False,
"use_cuda": cuda,
'seed': args.seed,
'device': device
}
end_model = EndModel(**init_kwargs)
if not os.path.exists(args.checkpoint_dir):
os.mkdir(args.checkpoint_dir)
with open(args.checkpoint_dir + '/init_kwargs.pickle', "wb") as f:
pickle.dump(init_kwargs, f, protocol=pickle.HIGHEST_PROTOCOL)
dropout = 0.4
# Train end model
end_model.train_model(
train_data=data_loader["train"],
valid_data=data_loader["dev"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar=True,
#loss_weights = [0.55,0.45],
batchnorm=False,
input_dropout=0.1,
middle_dropout=dropout,
checkpoint_dir=args.checkpoint_dir,
log_valid_metrics=['accuracy', 'f1'],
checkpoint_metric='f1',
)
end_model.score(
data_loader["dev"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc'])
def train_model(args):
#global args
#args = parser.parse_args()
hidden_size = 128
num_classes = 2
encode_dim = 1000 # using get_frm_output_size()
L,Y = load_labels(args)
# Label Model
# labelling functions analysis
print(lf_summary(L["dev"], Y = Y["dev"]))
# training label model
label_model = LabelModel(k=num_classes, seed=123)
label_model.train_model(L["train"], Y["dev"], n_epochs = 500, log_train_every = 50)
# evaluating label model
print('Trained Label Model Metrics:')
label_model.score((L["dev"], Y["dev"]), metric=['accuracy','precision', 'recall', 'f1'])
# comparison with majority vote of LFs
mv = MajorityLabelVoter(seed=123)
print('Majority Label Voter Metrics:')
mv.score((L["dev"], Y["dev"]), metric=['accuracy','precision', 'recall', 'f1'])
Ytrain_p = label_model.predict_proba(L["train"])
#print(Ytrain_ps.shape) #(377*50,2)
#Ydev_p = label_model.predict_proba(L["dev"])
# test models
#label_model.score((Ltest,Ytest), metric=['accuracy','precision', 'recall', 'f1'])
# End Model
# Create datasets and dataloaders
train, dev, test = load_dataset(args, Ytrain_p, Y["dev"], Y["test"])
data_loader = get_data_loader(train, dev, test, args.batch_size, args.num_workers)
#print(len(data_loader["train"])) # 18850 / batch_size
#print(len(data_loader["dev"])) # 1500 / batch_size
#print(len(data_loader["test"])) # 1000 / batch_size
#import ipdb; ipdb.set_trace()
# Define input encoder
cnn_encoder = FrameEncoderOC
if(torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
#import ipdb; ipdb.set_trace()
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction="attention",
encoder_class=cnn_encoder,
)
# Define end model
end_model = EndModel(
input_module=lstm_module,
layer_out_dims=[hidden_size, num_classes],
optimizer="adam",
#use_cuda=cuda,
batchnorm=True,
seed=123,
verbose=False,
device = device,
)
#print('Training model')
#tic = time.time()
dropout = 0.4
# Train end model
end_model.train_model(
train_data=data_loader["train"],
valid_data=data_loader["dev"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar = True,
loss_weights = [0.45,0.55],
batchnorm = 'True',
input_dropout = dropout,
middle_dropout = dropout,
#validation_metric='f1',
)
#print('Time taken for training:')
#print(time.time() - tic)
# evaluate end model
end_model.score(data_loader["dev"], verbose=True, metric=['accuracy','precision', 'recall', 'f1'])
def train_model(args):
# Create datasets and dataloaders
train, dev, test = load_dataset(args)
print('train size:', len(train)) #
print('dev size:', len(dev)) #
print('test size:', len(test)) #
# data in tuple of the form (frame,label)
# frame shape (3,224,224)
#import pdb; pdb.set_trace()
data_loader = get_data_loader(train, dev, test, args.batch_size)
num_classes = 2
encode_dim = 1000
# Define input encoder - can use the same
#cnn_encoder = FrameEncoderOC
if (torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
model = torch_models.resnet34(pretrained=True)
model = model.double()
#model = model.float()
# Define end model
end_model = EndModel(
input_module=model,
layer_out_dims=[encode_dim, num_classes],
optimizer="adam",
use_cuda=cuda,
input_batchnorm=True,
seed=args.seed,
verbose=False,
)
dropout = 0.4
# Train end model
end_model.train_model(
train_data=data_loader["train"],
valid_data=data_loader["dev"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar=True,
#loss_weights = [0.9,0.1],
batchnorm='False',
log_valid_metrics=['accuracy', 'f1'],
checkpoint_metric='f1',
checkpoint_dir=args.checkpoint_dir,
#validation_metric='accuracy',
#input_dropout = 0.1,
middle_dropout=dropout,
)
end_model.score(
data_loader["dev"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc'])
#import ipdb; ipdb.set_trace()
# saving dev set performance
Y_p, Y, Y_s = end_model._get_predictions(data_loader["dev"],
break_ties='random',
return_probs=True)
dev_labels = dev.labels
Y_s_0 = list(Y_s[:, 0])
Y_s_1 = list(Y_s[:, 1])
dev_ID = list(dev_labels["ID"])
dev_LABEL = list(dev_labels["LABEL"])
Y_p = list(Y_p)
Y = list(Y)
Y_p.insert(0, "Y_p"), Y.insert(0, "Y"),
Y_s_0.insert(0, "Y_s_0")
Y_s_1.insert(0, "Y_s_1")
dev_ID.insert(0, "ID")
dev_LABEL.insert(0, "LABEL")
np.save(args.mr_result_filename + "_dev",
np.column_stack((dev_ID, dev_LABEL, Y_p, Y, Y_s_0, Y_s_1)))
# saving test set performance
Y_p, Y, Y_s = end_model._get_predictions(data_loader["test"],
break_ties='random',
return_probs=True)
test_labels = test.labels
Y_s_0 = list(Y_s[:, 0])
Y_s_1 = list(Y_s[:, 1])
test_ID = list(test_labels["ID"])
test_LABEL = list(test_labels["LABEL"])
Y_p = list(Y_p)
Y = list(Y)
Y_p.insert(0, "Y_p"), Y.insert(0, "Y"),
Y_s_0.insert(0, "Y_s_0")
Y_s_1.insert(0, "Y_s_1")
test_ID.insert(0, "ID")
test_LABEL.insert(0, "LABEL")
np.save(args.mr_result_filename + "_test",
np.column_stack((test_ID, test_LABEL, Y_p, Y, Y_s_0, Y_s_1)))
def train_model(args):
#global args
#args = parser.parse_args()
hidden_size = 128
num_classes = 2
encode_dim = 1000 # using get_frm_output_size()
L,Y = load_labels(args)
data_list = {}
data_list["dev"] = glob(args.dev + '/la_4ch/*.npy')
data_list["test"] = glob(args.test + '/la_4ch/*.npy')
# End Model
# Create datasets and dataloaders
dev, test = load_dataset(data_list, Y)
data_loader = get_data_loader(dev, test, args.batch_size, args.num_workers)
#print(len(data_loader["dev"])) # 1500 / batch_size
#print(len(data_loader["test"])) # 1000 / batch_size
#import ipdb; ipdb.set_trace()
# Define input encoder
cnn_encoder = FrameEncoderOC
if(torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction="attention",
encoder_class=cnn_encoder
)
init_kwargs = {
"layer_out_dims":[hidden_size, num_classes],
"input_module": lstm_module,
"optimizer": "adam",
"verbose": False,
"input_batchnorm": False,
"use_cuda":cuda,
'seed':args.seed,
'device':device}
end_model = EndModel(**init_kwargs)
if not os.path.exists(args.checkpoint_dir):
os.mkdir(args.checkpoint_dir)
with open(args.checkpoint_dir+'/init_kwargs.pickle', "wb") as f:
pickle.dump(init_kwargs,f,protocol=pickle.HIGHEST_PROTOCOL)
# Train end model
end_model.train_model(
train_data=data_loader["dev"],
valid_data=data_loader["test"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar = True,
loss_weights = [0.55,0.45],
batchnorm = args.batchnorm,
middle_dropout = args.dropout,
checkpoint = False,
#checkpoint_every = args.n_epochs,
#checkpoint_best = False,
#checkpoint_dir = args.checkpoint_dir,
#validation_metric='f1',
)
# evaluate end model
end_model.score(data_loader["test"], verbose=True,metric=['accuracy','precision', 'recall', 'f1','roc-auc','ndcg'])
#end_model.score((Xtest,Ytest), verbose=True, metric=['accuracy','precision', 'recall', 'f1'])
# saving model
state = {
"model": end_model.state_dict(),
# "optimizer": optimizer.state_dict(),
# "lr_scheduler": lr_scheduler.state_dict() if lr_scheduler else None,
"score": end_model.score(data_loader["test"],verbose=False,metric=['accuracy','precision', 'recall', 'f1','roc-auc','ndcg'])
}
checkpoint_path = f"{args.checkpoint_dir}/best_model.pth"
torch.save(state, checkpoint_path)
def test_lstm_determinism(self):
"""Test whether training and scoring is deterministic given seed"""
X = torch.randint(1, MAX_INT + 1, (n, SEQ_LEN)).long()
Y = torch.zeros(n).long()
needles = np.random.randint(1, SEQ_LEN - 1, n)
for i in range(n):
X[i, needles[i]] = MAX_INT + 1
Y[i] = X[i, needles[i] + 1]
Xs = self._split_dataset(X)
Ys = self._split_dataset(Y)
embed_size = 4
hidden_size = 10
vocab_size = MAX_INT + 2
lstm_module = LSTMModule(
embed_size,
hidden_size,
vocab_size=vocab_size,
seed=123,
bidirectional=True,
verbose=False,
lstm_reduction="attention",
)
em = EndModel(
k=MAX_INT,
input_module=lstm_module,
layer_out_dims=[hidden_size * 2, MAX_INT],
batchnorm=True,
seed=123,
verbose=False,
)
em.train_model((Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=2,
verbose=False)
score_1 = em.score((Xs[2], Ys[2]), verbose=False)
# Test scoring determinism
score_2 = em.score((Xs[2], Ys[2]), verbose=False)
self.assertEqual(score_1, score_2)
# Test training determinism
lstm_module_2 = LSTMModule(
embed_size,
hidden_size,
vocab_size=vocab_size,
seed=123,
bidirectional=True,
verbose=False,
lstm_reduction="attention",
)
em_2 = EndModel(
k=MAX_INT,
input_module=lstm_module_2,
layer_out_dims=[hidden_size * 2, MAX_INT],
batchnorm=True,
seed=123,
verbose=False,
)
em_2.train_model((Xs[0], Ys[0]),
dev_data=(Xs[1], Ys[1]),
n_epochs=2,
verbose=False)
score_3 = em_2.score((Xs[2], Ys[2]), verbose=False)
self.assertEqual(score_1, score_3)
def train_model(args):
#global args
#args = parser.parse_args()
hidden_size = 128
num_classes = 2
encode_dim = 1000 # using get_frm_output_size()
L, Y = load_labels(args)
'''
# Label Model
# labelling functions analysis
print(lf_summary(L["dev"], Y = Y["dev"]))
# training label model
label_model = LabelModel(k=num_classes, seed=123)
label_model.train_model(L["train"], Y["dev"], n_epochs = 500, log_train_every = 50)
# evaluating label model
print('Trained Label Model Metrics:')
label_model.score((L["dev"], Y["dev"]), metric=['accuracy','precision', 'recall', 'f1'])
# comparison with majority vote of LFs
mv = MajorityLabelVoter(seed=123)
print('Majority Label Voter Metrics:')
mv.score((L["dev"], Y["dev"]), metric=['accuracy','precision', 'recall', 'f1'])
Ytrain_p = label_model.predict_proba(L["train"])
#print(Ytrain_ps.shape) #(377*50,2)
#Ydev_p = label_model.predict_proba(L["dev"])
# test models
#label_model.score((Ltest,Ytest), metric=['accuracy','precision', 'recall', 'f1'])
'''
# loading train model
#Ytrain = np.load(args.train_labels)
#Ytrain = 2 - Ytrain # only for mv_400
#Ytrain_p = np.zeros((len(Ytrain),2))
#Ytrain_p[:,1] = Ytrain
#Ytrain_p[:,0] = 1 - Ytrain
Ytrain_p = np.load(args.train_labels)
# End Model
# Create datasets and dataloaders
train, dev, test = load_dataset(args, Ytrain_p, Y["dev"], Y["test"])
data_loader = get_data_loader(train, dev, test, args.batch_size,
args.num_workers)
#print(len(data_loader["train"])) # 18850 / batch_size
#print(len(data_loader["dev"])) # 1500 / batch_size
#print(len(data_loader["test"])) # 1000 / batch_size
# Define input encoder
cnn_encoder = FrameEncoderOC
if (torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction=args.lstm_reduction,
encoder_class=cnn_encoder,
encoder_kwargs={"requires_grad": args.requires_grad})
'''
# Define end model
end_model = EndModel(
input_module=lstm_module,
layer_out_dims=[hidden_size, num_classes],
optimizer="adam",
#use_cuda=cuda,
batchnorm=False,
seed=args.seed,
verbose=False,
device = device,
)
'''
init_kwargs = {
"layer_out_dims": [hidden_size, num_classes],
"input_module": lstm_module,
"optimizer": "adam",
"verbose": False,
"input_batchnorm": False,
"use_cuda": cuda,
'seed': args.seed,
'device': device
}
end_model = EndModel(**init_kwargs)
if not os.path.exists(args.checkpoint_dir):
os.mkdir(args.checkpoint_dir)
with open(args.checkpoint_dir + '/init_kwargs.pickle', "wb") as f:
pickle.dump(init_kwargs, f, protocol=pickle.HIGHEST_PROTOCOL)
dropout = 0.4
# Train end model
end_model.train_model(
train_data=data_loader["train"],
valid_data=data_loader["dev"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar=True,
loss_weights=[0.55, 0.45],
batchnorm=True,
input_dropout=0.1,
middle_dropout=dropout,
checkpoint_dir=args.checkpoint_dir,
#validation_metric='f1',
)
# evaluate end model
print('Dev Set Performance')
end_model.score(
data_loader["dev"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc', 'ndcg'])
print('Test Set Performance')
end_model.score(
data_loader["test"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc', 'ndcg'])
def train_model(args):
# Create datasets and dataloaders
train, dev, test = load_dataset(args)
#print('train size:',len(train)) # 250
#print('dev size:',len(dev)) # 250
#print('test size:',len(test)) # 250
# data in tuple of the form (series,label)
# series shape (50,3,224,224)
#import pdb; pdb.set_trace()
data_loader = get_data_loader(train, dev, test, args.batch_size)
hidden_size = 128
num_classes = 2
encode_dim = 1000
# Define input encoder
cnn_encoder = FrameEncoderOC
if (torch.cuda.is_available()):
device = 'cuda'
else:
device = 'cpu'
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction="attention",
encoder_class=cnn_encoder,
)
# Define end model
end_model = EndModel(
input_module=lstm_module,
layer_out_dims=[hidden_size, num_classes],
optimizer="adam",
use_cuda=cuda,
batchnorm=True,
seed=args.seed,
verbose=False,
)
dropout = 0.4
# Train end model
end_model.train_model(
train_data=data_loader["train"],
valid_data=data_loader["dev"],
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
progress_bar=True,
#loss_weights = [0.9,0.1],
batchnorm='False',
log_valid_metrics=['accuracy', 'f1'],
checkpoint_metric='f1',
checkpoint_dir=args.checkpoint_dir,
#validation_metric='accuracy',
#input_dropout = 0.1,
middle_dropout=dropout,
)
print('Dev Set Performance')
end_model.score(
data_loader["dev"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc'])
print('Test Set Performance')
end_model.score(
data_loader["test"],
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc'])
else:
device = 'cpu'
end_model = EndModel([1000, 10, 2], seed=123, device=device)
#print(type(Xs[0]))
#print(Xs[0])
#print(type(Y_train_ps))
#end_model.train_model((Xs[0], Y_train_ps), valid_data=(Xs[1], Ys[1]), lr=0.01, l2=0.01, batch_size=256, n_epochs=5, checkpoint_metric='accuracy', checkpoint_metric_mode='max')
#Xs = tf.convert_to_tensor(Xs)
#Y_train_ps = tf.convert_to_tensor(Y_train_ps)
#Ys = tf.convert_to_tensor(Ys)
#Ls = tf.convert_to_tensor(Ls)
end_model.train_model((Xs[0], Y_train_ps),
valid_data=(Xs[1], Ys[1]),
lr=0.01,
l2=0.01,
batch_size=256,
n_epochs=5,
checkpoint_metric='accuracy',
checkpoint_metric_mode='max')
print("Label Model:")
score = label_model.score((Ls[2], Ys[2]),
metric=['accuracy', 'precision', 'recall', 'f1'])
print()
print("End Model:")
score = end_model.score((Xs[2], Ys[2]),
metric=['accuracy', 'precision', 'recall', 'f1'])
def train_model(args):
#global args
#args = parser.parse_args()
# Create datasets and dataloaders
train, dev, test, classes = load_dataset(args)
#print('train size:',len(train)) # 106
#print('dev size:',len(dev)) # 216
#print('test size:',len(test)) # 90
# data in tuple of the form (series,label)
# series shape [30,3,32,32]
#import pdb; pdb.set_trace()
train_loader, dev_loader, test_loader = data_loader(
train, dev, test, args.batch_size)
hidden_size = 128
num_classes = 2
# using get_frm_output_size()
encode_dim = 132
# Define input encoder
cnn_encoder = FrameEncoderBAV
# Define LSTM module
lstm_module = LSTMModule(
encode_dim,
hidden_size,
bidirectional=False,
verbose=False,
lstm_reduction="attention",
encoder_class=cnn_encoder,
)
# Define end model
end_model = EndModel(
input_module=lstm_module,
layer_out_dims=[hidden_size, num_classes],
optimizer="adam",
use_cuda=cuda,
batchnorm=False,
seed=args.seed,
verbose=False,
)
#end_model.config['train_config']['validation_metric'] = 'f1'
# Train end model
end_model.train_model(
train_data=train_loader,
valid_data=dev_loader,
l2=args.weight_decay,
lr=args.lr,
n_epochs=args.n_epochs,
log_train_every=1,
verbose=True,
#loss_weights = [0.96,0.04],
batchnorm='False',
checkpoint_metric='f1',
log_valid_metrics=['accuracy', 'f1'],
input_dropout=0.1,
middle_dropout=0.25,
validation_metric='f1',
)
end_model.score(
dev_loader,
verbose=True,
metric=['accuracy', 'precision', 'recall', 'f1', 'roc-auc'])
# Test end model
'''