def func(c: int, d):
with torch.no_grad():
all_thresholds = {}
model.eval()
hap_list = []
for idx in range(d.shape[0]):
# for j in range(n_layers):
# hap_list_j = []
out_j = model.conv(d[idx].unsqueeze(0))
t_jk = find_t(model.conv(data)) # shape == [channels, L']
for k in range(n_features):
han_jk, max_size = han(out_j, k, t_jk)
irf_j = int(receptive_field_dict["1"]["r"])
tmp = generate_indices_for_hap(han_jk, irf_j, max_size)
# hap_list_j.append(sorted(set(tmp)))
# hap_list.append(hap_list_j)
hap_list.append(d[idx, c, tmp]) # List[List[int]]
# data.shape == [batch, in_features, seq_len]
# hap_list[0].shape == [in_features, samples, receptive_filed_size_of_layer_j]
return torch.cat([i for i in hap_list])
som = SOM(8, 8, hap_lists[idx].shape[1], niter=NITER, device=device)
learning_error = som.fit(hap_lists[idx])
return som
soms = Parallel(n_jobs=-1)([delayed(train_som)(i) for i in range(in_features)])
x = torch.tensor(test_ds.data[0], dtype=torch.float32).unsqueeze(0)
model.eval()
with torch.no_grad():
t_jk = find_t(model.conv(torch.tensor(
test_ds.data, dtype=torch.float32))) # shape == [channels, L']
out = model.conv(x)
h, m = han(out, 0, t_jk)
rf = int(receptive_field_dict["1"]["r"])
hap = generate_indices_for_hap(h, rf, m)
ps = []
for h in hap:
ps.append(x[0][0][h])
ps = torch.stack(ps)
predicts, _ = soms[0].predict(ps)
tmp = torch.cat(hap_lists)
for_mean_variance: torch.Tensor = torch.stack([
p for p in tmp
if (soms[0].predict(p.unsqueeze(0))[0] == predicts[0]).all()
])
# TODO: predicts は複数個含まれることが考えられる。それに対する処理を入れる。
cphap = for_mean_variance.mean(0)
uncertainty = torch.var(for_mean_variance, dim=0)
def main():
experiment = Experiment(project_name="cphap", auto_output_logging=False)
train_loader, test_loader, (n_targets, n_features) = fetch_loader(batch_size=32)
model = CNN(n_features, 32, n_targets)
optimizer = optim.Adam(model.parameters())
runner = TimeSeriesUnsupervisedRunner(
model, optimizer, experiment, evaluator=SVC(probability=True)
)
runner.add_loader("train", train_loader).add_loader("test", test_loader)
runner.train_config(epochs=1000)
runner.run()
receptive_field_dict = receptive_field(model.conv, (n_features, 30))
data = test_loader.dataset.data
max_size = data.shape[-1]
model = runner.model.cpu()
runner.device = torch.device("cpu")
def func(c: int, d, padding: int):
with torch.no_grad():
all_thresholds = {}
model.eval()
hap_list = []
d = nn.functional.pad(d, [0, padding])
for idx in range(d.shape[0]):
# for j in range(n_layers):
# hap_list_j = []
out_j = model.conv(d[idx].unsqueeze(0))
t_jk = find_t(model.conv(data)) # shape == [channels, L']
for k in range(n_features):
han_jk, _ = han(out_j, k, t_jk)
irf_j = int(receptive_field_dict["1"]["r"])
tmp = generate_indices_for_hap(han_jk, irf_j, max_size)
# hap_list_j.append(sorted(set(tmp)))
# hap_list.append(hap_list_j)
hap_list.append(d[idx, c, tmp]) # List[List[int]]
# data.shape == [batch, in_features, seq_len]
# hap_list[0].shape == [in_features, samples, receptive_filed_size_of_layer_j]
return torch.cat([i for i in hap_list])
hap_lists = Parallel(n_jobs=-1)([delayed(func)(i, data, 2) for i in range(n_features)])
def train_som(niter, idx):
print("IDX: ", idx)
som = SOM(8, 8, hap_lists[idx].shape[1], niter=niter, device=runner.device)
learning_error = som.fit(hap_lists[idx].to(runner.device))
return som
soms = []
for i in range(n_features):
soms.append(train_som(1000, i))
for i in np.random.choice(list(range(len(test_loader.dataset.data))), size=20):
x = test_loader.dataset.data[i].unsqueeze(0)
x = nn.functional.pad(x, [0, 2])
y = test_loader.dataset.target[i]
model.eval()
with torch.no_grad():
model_logit = runner.encode(x)
model_pred = runner.evaluator.predict(model_logit)
model_logit = runner.evaluator.predict_proba(model_logit)[0][model_pred]
t_jk = find_t(
model.conv(test_loader.dataset.data)
) # shape == [channels, L']
out = model.conv(x)
h, m = han(out, 0, t_jk)
rf = int(receptive_field_dict["1"]["r"])
hap = generate_indices_for_hap(h, rf, m)
print("HAP: {} | Idx: {}".format(hap, i))
ps = []
cphaps = []
uncertainties = []
clusters = []
for h in hap:
ps.append(x[0][0][h])
try:
ps = torch.stack(ps).to(runner.device)
except RuntimeError:
pass
else:
predicts, _ = soms[0].predict(ps)
tmp = torch.cat(hap_lists)
for predict in predicts:
for_mean_variance: torch.Tensor = torch.stack(
[p for p in tmp if (soms[0].predict(p.unsqueeze(0).to(runner.device))[0] == predict).all()]
)
cphaps.append(for_mean_variance.mean(0))
uncertainties.append(torch.var(for_mean_variance, dim=0))
clusters.append(predict)
def plot(channel: int):
cm = plt.cm.get_cmap("hsv")
table = gen_table()
plt.figure(figsize=(10, 5))
plt.plot(x[0][channel])
title = "Label: {} | Channel : {} | SVM Pred: {} ({:.4f})".format(
y.cpu().numpy(), channel, model_pred[0], model_logit[0]
)
if len(cphaps) != 0:
for idx in range(len(cphaps)):
color = cm((table == clusters[idx]).all(1).argmax() / 64)
plt.fill_between(
hap[idx], cphaps[idx] - uncertainties[idx], cphaps[idx] + uncertainties[idx],
alpha=0.3, color=color, label=str(clusters[idx])
)
plt.legend()
plt.title(title)
plot(0)
plt.show()
runner.quite()