def cross_val(dataset, nb_fold, batch_size, collate_fn):
"""Given a dataset yields data for cross-validation.
Parameters
-----------
dataset : torch.Dataset
The dataset used
nb_fold : int
Number of fold used for cross-validation
batch_size : int
The batch size
Yields
------
trainloader : torch.Dataloader
testloader : torch.Dataloader
"""
indices = list(range(len(dataset)))
random.shuffle(indices)
for trainindices, testindices in all_but_one(indices, k=nb_fold):
trainindices = chain(*trainindices)
trainset = Subset(dataset, list(trainindices))
testset = Subset(dataset, list(testindices))
trainloader = DataLoader(trainset,
batch_size,
shuffle=True,
collate_fn=collate_fn)
testloader = DataLoader(testset,
batch_size,
shuffle=True,
collate_fn=collate_fn)
yield trainloader, testloader
def learn(model,
model_args,
device,
k=5,
batch_size=32,
seed=666,
smt_epoch=100,
rl_epoch=1000):
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
# Le probleme vient du count vectorizer qui vire certains mots
print("Load Dataset")
dataset, dataclasses = load(torch_dataset=True, dataclasses=True).values()
dataclasses = {qt._id: qt for qt in dataclasses}
engine = build_ir_engine()
collate_fn = embedding_collate_decorator(sequence_collate_fn)
indices = list(range(len(dataset)))
random.shuffle(indices)
for i, (trainindices, testindices) in enumerate(all_but_one(indices, k=5)):
trainindices = chain(*trainindices)
trainset = Subset(dataset, list(trainindices))
testset = Subset(dataset, list(testindices))
trainloader = DataLoader(trainset, device, True, collate_fn=collate_fn)
testloader = DataLoader(testset, device, True, collate_fn=collate_fn)
print("Build model")
model = model(*model_args)
try:
model = model.to(device)
except RuntimeError:
print("cudnn error")
model = model.to(device)
optimizer = optim.Adam(model.parameters())
loss_function = nn.BCELoss()
print("Train")
best_model = 0
delay = 0
max_delay = 10
print("Supervised Machine Translation")
for epoch in range(smt_epoch):
model.train()
n, mean = 0, 0
train_predictions = []
train_ids = []
for x, y, q_id, qrels, _ in trainloader:
x = x.to(device)
y = y.to(device)
pred = model(x)
pred__ = pred > 0.5
pred_ = pred__.detach().cpu().long().t().numpy().tolist()
train_predictions.extend(pred_)
train_ids.extend(map(lambda x: x.long().tolist(), q_id))
loss = loss_function(pred, y.float())
n += 1
mean = ((n - 1) * mean + loss.item()) / n
print(f"\rFold {i}, Epoch {epoch}\tTrain : {mean}", end="")
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_queries = {
id_: dataclasses[str(id_)].get_text(pred)
for id_, pred in zip(train_ids, train_predictions)
}
train_qrel = {
id_: dataclasses[str(id_)].qrels
for id_, pred in zip(train_ids, train_predictions)
}
train_map = eval_queries(train_queries, train_qrel, engine)
print(
f"\rFold {i}, Epoch {epoch}\tTrain Loss: {mean}, Train MAP {train_map}",
end="")
model.eval()
train_mean = mean
n, mean = 0, 0
test_predictions = []
test_ids = []
for x, y, q_id, qrels, _ in testloader:
x = x.to(device)
y = y.to(device)
pred = model(x)
pre__ = pred > 0.5
pred_ = pred__.detach().cpu().long().t().numpy().tolist()
test_predictions.extend(pred_)
test_ids.extend(map(lambda x: x.long().tolist(), q_id))
loss = loss_function(pred, y.float())
n += 1
mean = ((n - 1) * mean + loss.item()) / n
print(
f"\rFold {i}, Epoch {epoch}\tTrain Loss: {train_mean}\tTest : {mean}",
end="")
test_queries = {
id_: dataclasses[str(id_)].get_text(pred)
for id_, pred in zip(test_ids, test_predictions)
}
test_qrel = {
id_: dataclasses[str(id_)].qrels
for id_, pred in zip(test_ids, test_predictions)
}
test_map = eval_queries(test_queries, test_qrel, engine)
dataset_queries = {**train_queries, **test_queries}
dataset_qrel = {**train_qrel, **test_qrel}
dataset_map = eval_queries(dataset_queries, dataset_qrel, engine)
print(
"\b" * 500 +
f"\nFold {i}, Epoch {epoch}\tTrain MAP {train_map}\tTest MAP : {test_map}\tDataset MAP : {dataset_map}"
)
if test_map > best_model:
best_model = test_map
delay = 0
elif test_map < best_model:
delay += 1
if delay > max_delay:
print(best_model)
break
print("Reinforcement Learning")
for epoch in range(rl_epoch):
model.train()
n, mean = 0, 0
train_predictions = []
train_ids = []
for x, y, q_id, qrels, seq_lens in trainloader:
x = x.to(device)
y = y.to(device)
pred = model(x)
sampler = Bernoulli(pred)
batch_pred = sampler.sample()
log_probs = sampler.log_prob(batch_pred)
loss = log_probs.sum()
batch_ids = list(map(lambda x: x.long().tolist(), q_id))
batch_queries = {
id_: dataclasses[str(id_)].get_text(pred)
for id_, pred in zip(batch_ids, batch_pred)
}
batch_qrel = {
id_: dataclasses[str(id_)].qrels
for id_, pred in zip(batch_ids, batch_pred)
}
batch_map = eval_queries(batch_queries, batch_qrel, engine)
print(f"\rTrain Map : {batch_map}", end="")
loss = batch_map * log_probs
optimizer.zero_grad()
loss.backward()
optimizer.step()
model.eval()
train_mean = mean
n, mean = 0, 0
test_predictions = []
test_ids = []
print()
for x, y, q_id, qrels, seq_lens in testloader:
x = x.to(device)
y = y.to(device)
pred = model(x)
sampler = Bernoulli(pred)
batch_pred = sampler.sample()
log_probs = sampler.log_prob(batch_pred)
loss = log_probs.sum()
batch_qrel = {
id_: dataclasses[str(id_)].qrels
for id_, pred in zip(batch_ids, batch_pred)
}
batch_map = eval_queries(batch_queries, batch_qrel, engine)
print(f"\rTest Map : {batch_map}", end="")
print()
dataset = load(torch_dataset=True)["torch"]
def embedding_collate_decorator(collate_fn):
def wrapper(batch):
x, y, id_ = collate_fn(batch)
return x, y
return wrapper
collate_fn = embedding_collate_decorator(sequence_collate_fn)
indices = list(range(len(dataset)))
random.shuffle(indices)
for i, (trainindices, testindices) in enumerate(all_but_one(indices, k=10)):
trainindices = chain(*trainindices)
trainset = Subset(dataset, list(trainindices))
testset = Subset(dataset, list(testindices))
trainloader = DataLoader(trainset, 32, True, collate_fn=collate_fn)
testloader = DataLoader(testset, 32, True, collate_fn=collate_fn)
print("Build model")
encoder_archi = {
"input_size": embedding_size,
"hidden_size": hidden_size,
"num_layers": num_layers,
"bidirectional": True,
"dropout": 0.2
}
decoder_archi = {
