4 import matplotlib.pyplot as plt
9 ######################################################################
12 a = torch.empty(nb).uniform_(0, 1).view(-1, 1)
14 x = 0.4 * ((a-0.5) * 5 * math.pi).cos()
16 data = torch.cat((y, x), 1)
17 data = data @ torch.tensor([[1., -1.], [1., 1.]])
21 a = torch.empty(nb).uniform_(0, 1).view(-1, 1)
22 x = (a * 2.25 * math.pi).cos() * (a * 0.8 + 0.5)
23 y = (a * 2.25 * math.pi).sin() * (a * 0.8 + 0.5)
24 data = torch.cat((y, x), 1)
28 a = (torch.randint(5, (nb,)).float() / 5 * 2 * math.pi).view(-1, 1)
31 data = torch.cat((y, x), 1)
32 data = data + data.new(data.size()).normal_(0, 0.05)
35 ######################################################################
37 def train_model(data):
38 model = nn.Sequential(
44 batch_size, nb_epochs = 100, 1000
45 optimizer = torch.optim.Adam(model.parameters(), lr = 1e-3)
46 criterion = nn.MSELoss()
48 for e in range(nb_epochs):
50 for input in data.split(batch_size):
51 noise = input.new(input.size()).normal_(0, 0.1)
52 output = model(input + noise)
53 loss = criterion(output, input)
54 acc_loss += loss.item()
58 if (e+1)%100 == 0: print(e+1, acc_loss)
62 ######################################################################
64 def save_image(data_name, model, data):
65 a = torch.linspace(-1.5, 1.5, 30)
66 x = a.view( 1, -1, 1).expand(a.size(0), a.size(0), 1)
67 y = a.view(-1, 1, 1).expand(a.size(0), a.size(0), 1)
68 grid = torch.cat((y, x), 2).view(-1, 2)
70 # Take the origins of the arrows on the part of the grid closer than
71 # sqrt(0.1) to the data points
72 dist = (grid.view(-1, 1, 2) - data.view(1, -1, 2)).pow(2).sum(2).min(1)[0]
73 origins = grid[torch.arange(grid.size(0)).masked_select(dist < 0.1)]
75 field = model(origins).detach() - origins
78 ax = fig.add_subplot(1, 1, 1)
81 ax.set_xlim(-1.6, 1.6)
82 ax.set_ylim(-1.6, 1.6)
85 plot_field = ax.quiver(
86 origins[:, 0].numpy(), origins[:, 1].numpy(),
87 field[:, 0].numpy(), field[:, 1].numpy(),
88 units = 'xy', scale = 1,
89 width = 3e-3, headwidth = 25, headlength = 25
92 plot_data = ax.scatter(
93 data[:, 0].numpy(), data[:, 1].numpy(),
94 s = 1, color = 'tab:blue'
97 filename = f'denoising_field_{data_name}.pdf'
98 print(f'Saving {filename}')
99 fig.savefig(filename, bbox_inches='tight')
101 ######################################################################
103 for data_source in [ data_zigzag, data_spiral, data_penta ]:
104 data, data_name = data_source(1000)
105 data = data - data.mean(0)
106 model = train_model(data)
107 save_image(data_name, model, data)