X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=minidiffusion.py;h=066cbbbe1fa365458ea163bea0e64e1cd2787c74;hb=cfa42f2006a72819ad507f60429ea5e26308e5d0;hp=2c54d196062ee1775385335d67c03ba29a34b3ca;hpb=a3c7617d0b5770edf6030502e4eac477a7218820;p=pytorch.git

diff --git a/minidiffusion.py b/minidiffusion.py
index 2c54d19..066cbbb 100755
--- a/minidiffusion.py
+++ b/minidiffusion.py
@@ -11,17 +11,24 @@ import matplotlib.pyplot as plt
 
 import torch, torchvision
 from torch import nn
+from torch.nn import functional as F
 
 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
+print(f'device {device}')
+
 ######################################################################
 
 def sample_gaussian_mixture(nb):
     p, std = 0.3, 0.2
-    result = torch.empty(nb, 1).normal_(0, std)
+    result = torch.randn(nb, 1) * std
     result = result + torch.sign(torch.rand(result.size()) - p) / 2
     return result
 
+def sample_ramp(nb):
+    result = torch.min(torch.rand(nb, 1), torch.rand(nb, 1))
+    return result
+
 def sample_two_discs(nb):
     a = torch.rand(nb) * math.pi * 2
     b = torch.rand(nb).sqrt()
@@ -35,8 +42,9 @@ def sample_two_discs(nb):
 def sample_disc_grid(nb):
     a = torch.rand(nb) * math.pi * 2
     b = torch.rand(nb).sqrt()
-    q = torch.randint(5, (nb,)) / 2.5 - 2 / 2.5
-    r = torch.randint(5, (nb,)) / 2.5 - 2 / 2.5
+    N = 4
+    q = (torch.randint(N, (nb,)) - (N - 1) / 2) / ((N - 1) / 2)
+    r = (torch.randint(N, (nb,)) - (N - 1) / 2) / ((N - 1) / 2)
     b = b * 0.1
     result = torch.empty(nb, 2)
     result[:, 0] = a.cos() * b + q
@@ -58,11 +66,14 @@ def sample_mnist(nb):
     return result
 
 samplers = {
-    'gaussian_mixture': sample_gaussian_mixture,
-    'two_discs': sample_two_discs,
-    'disc_grid': sample_disc_grid,
-    'spiral': sample_spiral,
-    'mnist': sample_mnist,
+    f.__name__.removeprefix('sample_') : f for f in [
+        sample_gaussian_mixture,
+        sample_ramp,
+        sample_two_discs,
+        sample_disc_grid,
+        sample_spiral,
+        sample_mnist,
+    ]
 }
 
 ######################################################################
@@ -97,7 +108,7 @@ parser.add_argument('--learning_rate',
 
 parser.add_argument('--ema_decay',
                     type = float, default = 0.9999,
-                    help = 'EMA decay, < 0 is no EMA')
+                    help = 'EMA decay, <= 0 is no EMA')
 
 data_list = ', '.join( [ str(k) for k in samplers ])
 
@@ -105,6 +116,9 @@ parser.add_argument('--data',
                     type = str, default = 'gaussian_mixture',
                     help = f'Toy data-set to use: {data_list}')
 
+parser.add_argument('--no_window',
+                    action='store_true', default = False)
+
 args = parser.parse_args()
 
 if args.seed >= 0:
@@ -121,26 +135,37 @@ class EMA:
     def __init__(self, model, decay):
         self.model = model
         self.decay = decay
-        if self.decay < 0: return
-        self.ema = { }
+        self.mem = { }
         with torch.no_grad():
             for p in model.parameters():
-                self.ema[p] = p.clone()
+                self.mem[p] = p.clone()
 
     def step(self):
-        if self.decay < 0: return
         with torch.no_grad():
             for p in self.model.parameters():
-                self.ema[p].copy_(self.decay * self.ema[p] + (1 - self.decay) * p)
+                self.mem[p].copy_(self.decay * self.mem[p] + (1 - self.decay) * p)
 
-    def copy(self):
-        if self.decay < 0: return
+    def copy_to_model(self):
         with torch.no_grad():
             for p in self.model.parameters():
-                p.copy_(self.ema[p])
+                p.copy_(self.mem[p])
 
 ######################################################################
 
+# Gets a pair (x, t) and appends t (scalar or 1d tensor) to x as an
+# additional dimension / channel
+
+class TimeAppender(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, u):
+        x, t = u
+        if not torch.is_tensor(t):
+            t = x.new_full((x.size(0),), t)
+        t = t.view((-1,) + (1,) * (x.dim() - 1)).expand_as(x[:,:1])
+        return torch.cat((x, t), 1)
+
 class ConvNet(nn.Module):
     def __init__(self, in_channels, out_channels):
         super().__init__()
@@ -148,7 +173,8 @@ class ConvNet(nn.Module):
         ks, nc = 5, 64
 
         self.core = nn.Sequential(
-            nn.Conv2d(in_channels, nc, ks, padding = ks//2),
+            TimeAppender(),
+            nn.Conv2d(in_channels + 1, nc, ks, padding = ks//2),
             nn.ReLU(),
             nn.Conv2d(nc, nc, ks, padding = ks//2),
             nn.ReLU(),
@@ -161,8 +187,8 @@ class ConvNet(nn.Module):
             nn.Conv2d(nc, out_channels, ks, padding = ks//2),
         )
 
-    def forward(self, x):
-        return self.core(x)
+    def forward(self, u):
+        return self.core(u)
 
 ######################################################################
 # Data
@@ -179,9 +205,10 @@ train_mean, train_std = train_input.mean(), train_input.std()
 # Model
 
 if train_input.dim() == 2:
-    nh = 64
+    nh = 256
 
     model = nn.Sequential(
+        TimeAppender(),
         nn.Linear(train_input.size(1) + 1, nh),
         nn.ReLU(),
         nn.Linear(nh, nh),
@@ -193,10 +220,32 @@ if train_input.dim() == 2:
 
 elif train_input.dim() == 4:
 
-    model = ConvNet(train_input.size(1) + 1, train_input.size(1))
+    model = ConvNet(train_input.size(1), train_input.size(1))
 
 model.to(device)
 
+print(f'nb_parameters {sum([ p.numel() for p in model.parameters() ])}')
+
+######################################################################
+# Generate
+
+def generate(size, T, alpha, alpha_bar, sigma, model, train_mean, train_std):
+
+    with torch.no_grad():
+
+        x = torch.randn(size, device = device)
+
+        for t in range(T-1, -1, -1):
+            output = model((x, t / (T - 1) - 0.5))
+            z = torch.zeros_like(x) if t == 0 else torch.randn_like(x)
+            x = 1/torch.sqrt(alpha[t]) \
+                * (x - (1-alpha[t]) / torch.sqrt(1-alpha_bar[t]) * output) \
+                + sigma[t] * z
+
+        x = x * train_std + train_mean
+
+        return x
+
 ######################################################################
 # Train
 
@@ -206,7 +255,7 @@ alpha = 1 - beta
 alpha_bar = alpha.log().cumsum(0).exp()
 sigma = beta.sqrt()
 
-ema = EMA(model, decay = args.ema_decay)
+ema = EMA(model, decay = args.ema_decay) if args.ema_decay > 0 else None
 
 for k in range(args.nb_epochs):
 
@@ -217,95 +266,123 @@ for k in range(args.nb_epochs):
         x0 = (x0 - train_mean) / train_std
         t = torch.randint(T, (x0.size(0),) + (1,) * (x0.dim() - 1), device = x0.device)
         eps = torch.randn_like(x0)
-        input = torch.sqrt(alpha_bar[t]) * x0 + torch.sqrt(1 - alpha_bar[t]) * eps
-        input = torch.cat((input, t.expand_as(x0[:,:1]) / (T - 1) - 0.5), 1)
-        loss = (eps - model(input)).pow(2).mean()
+        xt = torch.sqrt(alpha_bar[t]) * x0 + torch.sqrt(1 - alpha_bar[t]) * eps
+        output = model((xt, t / (T - 1) - 0.5))
+        loss = (eps - output).pow(2).mean()
         acc_loss += loss.item() * x0.size(0)
 
         optimizer.zero_grad()
         loss.backward()
         optimizer.step()
 
-        ema.step()
+        if ema is not None: ema.step()
 
-    if k%10 == 0: print(f'{k} {acc_loss / train_input.size(0)}')
+    print(f'{k} {acc_loss / train_input.size(0)}')
 
-ema.copy()
+if ema is not None: ema.copy_to_model()
 
 ######################################################################
-# Generate
+# Plot
 
-def generate(size, model):
-    with torch.no_grad():
-        x = torch.randn(size, device = device)
+model.eval()
 
-        for t in range(T-1, -1, -1):
-            z = torch.zeros_like(x) if t == 0 else torch.randn_like(x)
-            input = torch.cat((x, torch.full_like(x[:,:1], t / (T - 1) - 0.5)), 1)
-            x = 1/torch.sqrt(alpha[t]) \
-                * (x - (1-alpha[t]) / torch.sqrt(1-alpha_bar[t]) * model(input)) \
-                + sigma[t] * z
+########################################
+# Nx1 -> histogram
+if train_input.dim() == 2 and train_input.size(1) == 1:
 
-        x = x * train_std + train_mean
+    fig = plt.figure()
+    fig.set_figheight(5)
+    fig.set_figwidth(8)
 
-        return x
+    ax = fig.add_subplot(1, 1, 1)
 
-######################################################################
-# Plot
+    x = generate((10000, 1), T, alpha, alpha_bar, sigma,
+                 model, train_mean, train_std)
 
-model.eval()
+    ax.set_xlim(-1.25, 1.25)
+    ax.spines.right.set_visible(False)
+    ax.spines.top.set_visible(False)
 
-if train_input.dim() == 2:
-    fig = plt.figure()
-    ax = fig.add_subplot(1, 1, 1)
+    d = train_input.flatten().detach().to('cpu').numpy()
+    ax.hist(d, 25, (-1, 1),
+            density = True,
+            histtype = 'bar', edgecolor = 'white', color = 'lightblue', label = 'Train')
 
-    if train_input.size(1) == 1:
+    d = x.flatten().detach().to('cpu').numpy()
+    ax.hist(d, 25, (-1, 1),
+            density = True,
+            histtype = 'step', color = 'red', label = 'Synthesis')
 
-        x = generate((10000, 1), model)
+    ax.legend(frameon = False, loc = 2)
 
-        ax.set_xlim(-1.25, 1.25)
+    filename = f'minidiffusion_{args.data}.pdf'
+    print(f'saving {filename}')
+    fig.savefig(filename, bbox_inches='tight')
 
-        d = train_input.flatten().detach().to('cpu').numpy()
-        ax.hist(d, 25, (-1, 1),
-                density = True,
-                histtype = 'stepfilled', color = 'lightblue', label = 'Train')
+    if not args.no_window and hasattr(plt.get_current_fig_manager(), 'window'):
+        plt.get_current_fig_manager().window.setGeometry(2, 2, 1024, 768)
+        plt.show()
 
-        d = x.flatten().detach().to('cpu').numpy()
-        ax.hist(d, 25, (-1, 1),
-                density = True,
-                histtype = 'step', color = 'red', label = 'Synthesis')
+########################################
+# Nx2 -> scatter plot
+elif train_input.dim() == 2 and train_input.size(1) == 2:
 
-        ax.legend(frameon = False, loc = 2)
+    fig = plt.figure()
+    fig.set_figheight(6)
+    fig.set_figwidth(6)
 
-    elif train_input.size(1) == 2:
+    ax = fig.add_subplot(1, 1, 1)
 
-        x = generate((1000, 2), model)
+    x = generate((1000, 2), T, alpha, alpha_bar, sigma,
+                 model, train_mean, train_std)
 
-        ax.set_xlim(-1.25, 1.25)
-        ax.set_ylim(-1.25, 1.25)
-        ax.set(aspect = 1)
+    ax.set_xlim(-1.5, 1.5)
+    ax.set_ylim(-1.5, 1.5)
+    ax.set(aspect = 1)
+    ax.spines.right.set_visible(False)
+    ax.spines.top.set_visible(False)
 
-        d = train_input[:x.size(0)].detach().to('cpu').numpy()
-        ax.scatter(d[:, 0], d[:, 1],
-                   color = 'lightblue', label = 'Train')
+    d = train_input[:x.size(0)].detach().to('cpu').numpy()
+    ax.scatter(d[:, 0], d[:, 1],
+               s = 2.5, color = 'gray', label = 'Train')
 
-        d = x.detach().to('cpu').numpy()
-        ax.scatter(d[:, 0], d[:, 1],
-                   facecolors = 'none', color = 'red', label = 'Synthesis')
+    d = x.detach().to('cpu').numpy()
+    ax.scatter(d[:, 0], d[:, 1],
+               s = 2.0, color = 'red', label = 'Synthesis')
 
-        ax.legend(frameon = False, loc = 2)
+    ax.legend(frameon = False, loc = 2)
 
-    filename = f'diffusion_{args.data}.pdf'
+    filename = f'minidiffusion_{args.data}.pdf'
     print(f'saving {filename}')
     fig.savefig(filename, bbox_inches='tight')
 
-    if hasattr(plt.get_current_fig_manager(), 'window'):
+    if not args.no_window and hasattr(plt.get_current_fig_manager(), 'window'):
         plt.get_current_fig_manager().window.setGeometry(2, 2, 1024, 768)
         plt.show()
 
+########################################
+# NxCxHxW -> image
 elif train_input.dim() == 4:
-    x = generate((128,) + train_input.size()[1:], model)
-    x = 1 - x.clamp(min = 0, max = 255) / 255
-    torchvision.utils.save_image(x, f'diffusion_{args.data}.png', nrow = 16, pad_value = 0.8)
+
+    x = generate((128,) + train_input.size()[1:], T, alpha, alpha_bar, sigma,
+                 model, train_mean, train_std)
+
+    x = torchvision.utils.make_grid(x.clamp(min = 0, max = 255),
+                                    nrow = 16, padding = 1, pad_value = 64)
+    x = F.pad(x, pad = (2, 2, 2, 2), value = 64)[None]
+
+    t = torchvision.utils.make_grid(train_input[:128],
+                                    nrow = 16, padding = 1, pad_value = 64)
+    t = F.pad(t, pad = (2, 2, 2, 2), value = 64)[None]
+
+    result = 1 - torch.cat((t, x), 2) / 255
+
+    filename = f'minidiffusion_{args.data}.png'
+    print(f'saving {filename}')
+    torchvision.utils.save_image(result, filename)
+
+else:
+
+    print(f'cannot plot result of size {train_input.size()}')
 
 ######################################################################