X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=minidiffusion.py;h=a3ffda008ef7672ebbe5d98246a21113cd2030b1;hb=HEAD;hp=037ef11f0b3b5ebb312d35b04f690d686e249706;hpb=f27d6083fbe7243f5896ddd49587fe1923fe9a79;p=pytorch.git

diff --git a/minidiffusion.py b/minidiffusion.py
index 037ef11..a3ffda0 100755
--- a/minidiffusion.py
+++ b/minidiffusion.py
@@ -5,100 +5,402 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
-# Minimal implementation of Jonathan Ho, Ajay Jain, Pieter Abbeel
-# "Denoising Diffusion Probabilistic Models" (2020)
-#
-# https://arxiv.org/abs/2006.11239
+import math, argparse
 
 import matplotlib.pyplot as plt
-import torch
+
+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_phi(nb):
+
+def sample_gaussian_mixture(nb):
     p, std = 0.3, 0.2
-    result = torch.empty(nb).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()
+    q = (torch.rand(nb) <= 0.5).long()
+    b = b * (0.3 + 0.2 * q)
+    result = torch.empty(nb, 2)
+    result[:, 0] = a.cos() * b - 0.5 + q
+    result[:, 1] = a.sin() * b - 0.5 + q
+    return result
+
+
+def sample_disc_grid(nb):
+    a = torch.rand(nb) * math.pi * 2
+    b = torch.rand(nb).sqrt()
+    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
+    result[:, 1] = a.sin() * b + r
+    return result
+
+
+def sample_spiral(nb):
+    u = torch.rand(nb)
+    rho = u * 0.65 + 0.25 + torch.rand(nb) * 0.15
+    theta = u * math.pi * 3
+    result = torch.empty(nb, 2)
+    result[:, 0] = theta.cos() * rho
+    result[:, 1] = theta.sin() * rho
+    return result
+
+
+def sample_mnist(nb):
+    train_set = torchvision.datasets.MNIST(root="./data/", train=True, download=True)
+    result = train_set.data[:nb].to(device).view(-1, 1, 28, 28).float()
+    return result
+
+
+samplers = {
+    f.__name__.removeprefix("sample_"): f
+    for f in [
+        sample_gaussian_mixture,
+        sample_ramp,
+        sample_two_discs,
+        sample_disc_grid,
+        sample_spiral,
+        sample_mnist,
+    ]
+}
+
 ######################################################################
 
-model = nn.Sequential(
-    nn.Linear(2, 32),
-    nn.ReLU(),
-    nn.Linear(32, 32),
-    nn.ReLU(),
-    nn.Linear(32, 1),
+parser = argparse.ArgumentParser(
+    description="""A minimal implementation of Jonathan Ho, Ajay Jain, Pieter Abbeel
+"Denoising Diffusion Probabilistic Models" (2020)
+https://arxiv.org/abs/2006.11239""",
+    formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+)
+
+parser.add_argument(
+    "--seed", type=int, default=0, help="Random seed, < 0 is no seeding"
+)
+
+parser.add_argument("--nb_epochs", type=int, default=100, help="How many epochs")
+
+parser.add_argument("--batch_size", type=int, default=25, help="Batch size")
+
+parser.add_argument(
+    "--nb_samples", type=int, default=25000, help="Number of training examples"
 )
 
+parser.add_argument("--learning_rate", type=float, default=1e-3, help="Learning rate")
+
+parser.add_argument(
+    "--ema_decay", type=float, default=0.9999, help="EMA decay, <= 0 is no EMA"
+)
+
+data_list = ", ".join([str(k) for k in samplers])
+
+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:
+    # torch.backends.cudnn.deterministic = True
+    # torch.backends.cudnn.benchmark = False
+    # torch.use_deterministic_algorithms(True)
+    torch.manual_seed(args.seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed_all(args.seed)
+
 ######################################################################
-# Train
 
-nb_samples = 25000
-nb_epochs = 250
-batch_size = 100
 
-train_input = sample_phi(nb_samples)[:, None]
+class EMA:
+    def __init__(self, model, decay):
+        self.model = model
+        self.decay = decay
+        self.mem = {}
+        with torch.no_grad():
+            for p in model.parameters():
+                self.mem[p] = p.clone()
+
+    def step(self):
+        with torch.no_grad():
+            for p in self.model.parameters():
+                self.mem[p].copy_(self.decay * self.mem[p] + (1 - self.decay) * p)
+
+    def copy_to_model(self):
+        with torch.no_grad():
+            for p in self.model.parameters():
+                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__()
+
+        ks, nc = 5, 64
+
+        self.core = nn.Sequential(
+            TimeAppender(),
+            nn.Conv2d(in_channels + 1, nc, ks, padding=ks // 2),
+            nn.ReLU(),
+            nn.Conv2d(nc, nc, ks, padding=ks // 2),
+            nn.ReLU(),
+            nn.Conv2d(nc, nc, ks, padding=ks // 2),
+            nn.ReLU(),
+            nn.Conv2d(nc, nc, ks, padding=ks // 2),
+            nn.ReLU(),
+            nn.Conv2d(nc, nc, ks, padding=ks // 2),
+            nn.ReLU(),
+            nn.Conv2d(nc, out_channels, ks, padding=ks // 2),
+        )
+
+    def forward(self, u):
+        return self.core(u)
+
+
+######################################################################
+# Data
+
+try:
+    train_input = samplers[args.data](args.nb_samples).to(device)
+except KeyError:
+    print(f"unknown data {args.data}")
+    exit(1)
+
+train_mean, train_std = train_input.mean(), train_input.std()
+
+######################################################################
+# Model
+
+if train_input.dim() == 2:
+    nh = 256
+
+    model = nn.Sequential(
+        TimeAppender(),
+        nn.Linear(train_input.size(1) + 1, nh),
+        nn.ReLU(),
+        nn.Linear(nh, nh),
+        nn.ReLU(),
+        nn.Linear(nh, nh),
+        nn.ReLU(),
+        nn.Linear(nh, train_input.size(1)),
+    )
+
+elif train_input.dim() == 4:
+    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
 
 T = 1000
-beta = torch.linspace(1e-4, 0.02, T)
+beta = torch.linspace(1e-4, 0.02, T, device=device)
 alpha = 1 - beta
 alpha_bar = alpha.log().cumsum(0).exp()
 sigma = beta.sqrt()
 
-for k in range(nb_epochs):
+ema = EMA(model, decay=args.ema_decay) if args.ema_decay > 0 else None
+
+for k in range(args.nb_epochs):
     acc_loss = 0
-    optimizer = torch.optim.Adam(model.parameters(), lr = 1e-4 * (1 - k / nb_epochs) )
-
-    for x0 in train_input.split(batch_size):
-        t = torch.randint(T, (x0.size(0), 1))
-        eps = torch.randn(x0.size())
-        input = alpha_bar[t].sqrt() * x0 + (1 - alpha_bar[t]).sqrt() * eps
-        input = torch.cat((input, 2 * t / T - 1), 1)
-        output = model(input)
+    optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)
+
+    for x0 in train_input.split(args.batch_size):
+        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)
+        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()
 
-        acc_loss += loss.item()
+        if ema is not None:
+            ema.step()
 
-    if k%10 == 0: print(k, loss.item())
+    print(f"{k} {acc_loss / train_input.size(0)}")
+
+if ema is not None:
+    ema.copy_to_model()
 
 ######################################################################
-# Generate
+# Plot
 
-x = torch.randn(10000, 1)
+model.eval()
 
-for t in range(T-1, -1, -1):
-    z = torch.zeros(x.size()) if t == 0 else torch.randn(x.size())
-    input = torch.cat((x, torch.ones(x.size(0), 1) * 2 * t / T - 1), 1)
-    x = 1 / alpha[t].sqrt() * (x - (1 - alpha[t])/(1 - alpha_bar[t]).sqrt() * model(input)) + sigma[t] * z
+########################################
+# Nx1 -> histogram
+if train_input.dim() == 2 and train_input.size(1) == 1:
+    fig = plt.figure()
+    fig.set_figheight(5)
+    fig.set_figwidth(8)
 
-######################################################################
-# Plot
+    ax = fig.add_subplot(1, 1, 1)
+
+    x = generate((10000, 1), T, alpha, alpha_bar, sigma, model, train_mean, train_std)
+
+    ax.set_xlim(-1.25, 1.25)
+    ax.spines.right.set_visible(False)
+    ax.spines.top.set_visible(False)
+
+    d = train_input.flatten().detach().to("cpu").numpy()
+    ax.hist(
+        d,
+        25,
+        (-1, 1),
+        density=True,
+        histtype="bar",
+        edgecolor="white",
+        color="lightblue",
+        label="Train",
+    )
+
+    d = x.flatten().detach().to("cpu").numpy()
+    ax.hist(
+        d, 25, (-1, 1), density=True, histtype="step", color="red", label="Synthesis"
+    )
+
+    ax.legend(frameon=False, loc=2)
+
+    filename = f"minidiffusion_{args.data}.pdf"
+    print(f"saving {filename}")
+    fig.savefig(filename, bbox_inches="tight")
+
+    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()
+
+########################################
+# Nx2 -> scatter plot
+elif train_input.dim() == 2 and train_input.size(1) == 2:
+    fig = plt.figure()
+    fig.set_figheight(6)
+    fig.set_figwidth(6)
+
+    ax = fig.add_subplot(1, 1, 1)
+
+    x = generate((1000, 2), T, alpha, alpha_bar, sigma, model, train_mean, train_std)
+
+    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], s=2.5, color="gray", label="Train")
+
+    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)
+
+    filename = f"minidiffusion_{args.data}.pdf"
+    print(f"saving {filename}")
+    fig.savefig(filename, bbox_inches="tight")
+
+    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()
 
-fig = plt.figure()
-ax = fig.add_subplot(1, 1, 1)
-ax.set_xlim(-1.25, 1.25)
+########################################
+# NxCxHxW -> image
+elif train_input.dim() == 4:
+    x = generate(
+        (128,) + train_input.size()[1:],
+        T,
+        alpha,
+        alpha_bar,
+        sigma,
+        model,
+        train_mean,
+        train_std,
+    )
 
-d = train_input.flatten().detach().numpy()
-ax.hist(d, 25, (-1, 1),
-        density = True,
-        histtype = 'stepfilled', color = 'lightblue', label = 'Train')
+    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]
 
-d = x.flatten().detach().numpy()
-ax.hist(d, 25, (-1, 1),
-        density = True,
-        histtype = 'step', color = 'red', label = 'Synthesis')
+    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]
 
-ax.legend(frameon = False, loc = 2)
+    result = 1 - torch.cat((t, x), 2) / 255
 
-filename = 'diffusion.pdf'
-print(f'saving {filename}')
-fig.savefig(filename, bbox_inches='tight')
+    filename = f"minidiffusion_{args.data}.png"
+    print(f"saving {filename}")
+    torchvision.utils.save_image(result, filename)
 
-plt.show()
+else:
+    print(f"cannot plot result of size {train_input.size()}")
 
 ######################################################################