up

run_inference.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+# !pip install diffusers
+from diffusers import DiffusionPipeline
+import PIL.Image
+import numpy as np
+
+model_id = "fusing/ddpm-cifar10"
+model_id = "fusing/ddpm-lsun-bedroom"
+
+# load model and scheduler
+ddpm = DiffusionPipeline.from_pretrained(model_id)
+
+# run pipeline in inference (sample random noise and denoise)
+image = ddpm()
+
+# process image to PIL
+image_processed = image.cpu().permute(0, 2, 3, 1)
+image_processed = (image_processed + 1.0) * 127.5
+image_processed = image_processed.numpy().astype(np.uint8)
+image_pil = PIL.Image.fromarray(image_processed[0])
+
+# save image
+image_pil.save("/home/patrick/images/show.png")

src/diffusers/schedulers/ddim.py

@@ -0,0 +1,102 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import math
+from torch import nn
+
+from ..configuration_utils import ConfigMixin
+from .schedulers_utils import linear_beta_schedule, betas_for_alpha_bar
+
+
+SAMPLING_CONFIG_NAME = "scheduler_config.json"
+
+
+class GaussianDDPMScheduler(nn.Module, ConfigMixin):
+
+    config_name = SAMPLING_CONFIG_NAME
+
+    def __init__(
+        self,
+        timesteps=1000,
+        beta_start=0.0001,
+        beta_end=0.02,
+        beta_schedule="linear",
+        variance_type="fixed_small",
+    ):
+        super().__init__()
+        self.register(
+            timesteps=timesteps,
+            beta_start=beta_start,
+            beta_end=beta_end,
+            beta_schedule=beta_schedule,
+            variance_type=variance_type,
+        )
+        self.num_timesteps = int(timesteps)
+
+        if beta_schedule == "linear":
+            betas = linear_beta_schedule(timesteps, beta_start=beta_start, beta_end=beta_end)
+        elif beta_schedule == "squaredcos_cap_v2":
+            # GLIDE cosine schedule
+            betas = betas_for_alpha_bar(
+                timesteps,
+                lambda t: math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2,
+            )
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        alphas = 1.0 - betas
+        alphas_cumprod = torch.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = torch.nn.functional.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
+
+        variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
+
+        if variance_type == "fixed_small":
+            log_variance = torch.log(variance.clamp(min=1e-20))
+        elif variance_type == "fixed_large":
+            log_variance = torch.log(torch.cat([variance[1:2], betas[1:]], dim=0))
+
+        self.register_buffer("betas", betas.to(torch.float32))
+        self.register_buffer("alphas", alphas.to(torch.float32))
+        self.register_buffer("alphas_cumprod", alphas_cumprod.to(torch.float32))
+
+        self.register_buffer("log_variance", log_variance.to(torch.float32))
+
+    def get_alpha(self, time_step):
+        return self.alphas[time_step]
+
+    def get_beta(self, time_step):
+        return self.betas[time_step]
+
+    def get_alpha_prod(self, time_step):
+        if time_step < 0:
+            return torch.tensor(1.0)
+        return self.alphas_cumprod[time_step]
+
+    def sample_variance(self, time_step, shape, device, generator=None):
+        variance = self.log_variance[time_step]
+        nonzero_mask = torch.tensor([1 - (time_step == 0)], device=device).float()[None, :]
+
+        noise = self.sample_noise(shape, device=device, generator=generator)
+
+        sampled_variance = nonzero_mask * (0.5 * variance).exp()
+        sampled_variance = sampled_variance * noise
+
+        return sampled_variance
+
+    def sample_noise(self, shape, device, generator=None):
+        # always sample on CPU to be deterministic
+        return torch.randn(shape, generator=generator).to(device)
+
+    def __len__(self):
+        return self.num_timesteps

src/diffusers/schedulers/gaussian_ddpm.py

@@ -16,35 +16,12 @@ import math
 from torch import nn
 
 from ..configuration_utils import ConfigMixin
+from .schedulers_utils import linear_beta_schedule, betas_for_alpha_bar
 
 
 SAMPLING_CONFIG_NAME = "scheduler_config.json"
 
 
-def linear_beta_schedule(timesteps, beta_start, beta_end):
-    return torch.linspace(beta_start, beta_end, timesteps, dtype=torch.float64)
-
-
-def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
-    """
-    Create a beta schedule that discretizes the given alpha_t_bar function,
-    which defines the cumulative product of (1-beta) over time from t = [0,1].
-
-    :param num_diffusion_timesteps: the number of betas to produce.
-    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
-                      produces the cumulative product of (1-beta) up to that
-                      part of the diffusion process.
-    :param max_beta: the maximum beta to use; use values lower than 1 to
-                     prevent singularities.
-    """
-    betas = []
-    for i in range(num_diffusion_timesteps):
-        t1 = i / num_diffusion_timesteps
-        t2 = (i + 1) / num_diffusion_timesteps
-        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
-    return torch.tensor(betas, dtype=torch.float64)
-
-
 class GaussianDDPMScheduler(nn.Module, ConfigMixin):
 
     config_name = SAMPLING_CONFIG_NAME

src/diffusers/schedulers/schedulers_utils.py

@@ -0,0 +1,38 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+
+
+def linear_beta_schedule(timesteps, beta_start, beta_end):
+    return torch.linspace(beta_start, beta_end, timesteps, dtype=torch.float64)
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float64)