for gcn in self.gcns:
x = gcn(x, edge_index)
x = F.relu(x)
x, _, _, batch, _, _ = self.graph_pooling(x, edge_index, None, batch)
x = global_max_pool(x, batch)
outputs = self.dense(x)
outputs = F.relu(outputs)
outputs = self.out(outputs)
return outputs
model = Model(dataset.item_num, 2)
opt = optim.SGD(model.parameters(), lr=0.5)
# 3. learner
learner = Learner(model, opt, F.cross_entropy, train_db, metrics=[accuracy])
# 4. fit
learner.fit(3)
# 5. test
learner.test(test_dl)
# 6. predict
pred = learner.predict(test_dl)
print(pred.size())
# 7. plot
learner.recorder.plot_metrics()
plt.show()
in_dim = data.train_ds.x.shape[1]
h_dim = 128
model = nn.Sequential(nn.Linear(in_dim, h_dim), nn.ReLU(),
nn.Linear(h_dim, 10))
opt = optim.SGD(model.parameters(), lr=1.5)
loss_func = F.cross_entropy
learner = Learner(model,
opt,
loss_func,
data,
metrics=[accuracy],
callbacks=EarlyStopping(patience=3))
learner.fit(10)
learner.test(test_dl)
pred = learner.predict(x_valid)
print(pred.size())
learner.recorder.plot_loss()
learner.recorder.plot_metrics()
plt.show()
# # 1. ------ 数据
# x_train, y_train, x_valid, y_valid = mnist_data()
# train_ds, valid_ds = Dataset(x_train, y_train), Dataset(x_valid, y_valid)
# bs = 128
# train_dl = DataLoader(train_ds, batch_size=bs, shuffle=True)
# valid_dl = DataLoader(valid_ds, batch_size=bs, shuffle=True)
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.gcn1 = GCN(1433, 16, F.relu)
self.gcn2 = GCN(16, 7, None)
def forward(self, g, features):
x = self.gcn1(g, features)
x = self.gcn2(g, x)
return x
net = Net()
optimizer = th.optim.Adam(net.parameters(), lr=1e-3)
# 3. learner
learner = Learner(net, optimizer, masked_cross_entropy, data, metrics=masked_accuracy)
# 4. fit
learner.fit(50)
# 5. test
learner.test(test_dl)
# 6. predict
learner.predict(test_dl)
# 7. plot
learner.recorder.plot_metrics()
plt.show()
nn.Conv2d(8, 16, 3, padding=1, stride=2),
nn.ReLU(), # 7
nn.Conv2d(16, 32, 3, padding=1, stride=2),
nn.ReLU(), # 4
nn.Conv2d(32, 32, 3, padding=1, stride=2),
nn.ReLU(), # 2
nn.AdaptiveAvgPool2d(1),
Lambda(dp.flatten),
nn.Linear(32, out_dim))
x_train, y_train, x_valid, y_valid, x_test, y_test = mnist()
x_train, x_valid, x_test = dp.normalize_to(x_train, x_valid, x_test)
train_ds, valid_ds, test_ds = Dataset(x_train, y_train), Dataset(
x_valid, y_valid), Dataset(x_test, y_test)
bs = 512
train_dl = DataLoader(train_ds, batch_size=bs, shuffle=True)
valid_dl = DataLoader(valid_ds, batch_size=bs)
test_dl = DataLoader(test_ds, batch_size=bs)
data = DataBunch(train_dl, valid_dl)
model = cnn_model(10)
opt = optim.Adam(model.parameters(), lr=0.005)
loss_func = F.cross_entropy
learner = Learner(model, opt, loss_func, data, metrics=accuracy)
learner.fit(5)
plt.show()
model = Classifier(1, 256, train_ds.num_classes)
optimizer = optim.Adam(model.parameters(), lr=0.001)
# 3. learne
loss_func = nn.CrossEntropyLoss()
learner = Learner(model,
optimizer,
loss_func,
data,
metrics=accuracy,
callbacks=DGLInterpreter)
# 4. fit
learner.fit(80)
# 5. test
learner.test(test_dl)
loss = nn.CrossEntropyLoss(reduction='none')
scores, xs, ys, preds, indecies = learner.interpreter.top_data(loss,
k=10,
phase='train',
largest=True)
plt.figure(figsize=[10, 5])
for i in range(10):
plt.subplot(2, 5, i + 1)
learner.interpreter.plot_graph(xs[i], node_size=50)
plt.title(f'loss-{scores[i]:0.6f}')
train_dl = DataLoader(train_ds, batch_size=bs, shuffle=True)
valid_dl = DataLoader(valid_ds, batch_size=bs)
test_dl = DataLoader(test_ds, batch_size=bs)
data = DataBunch(train_dl, valid_dl)
in_dim = data.train_ds.x.shape[1]
h_dim = 128
model = nn.Sequential(nn.Linear(in_dim, h_dim), nn.ReLU(),
nn.Linear(h_dim, 10))
opt = optim.SGD(model.parameters(), lr=0.35)
loss_func = F.cross_entropy
cbks = Checkpoints(3) # save checkpoint each 3 epochs
learner = Learner(model,
opt,
loss_func,
data,
metrics=[accuracy],
callbacks=cbks)
learner.fit(3)
learner.load_checkpoint() # load the latest checkpoint
learner.fit(2) # go on training
learner.test(test_dl)
pred = learner.predict(x_valid)
print(pred.size())
learner.recorder.plot_loss()
plt.show()
