initialized

.gitignore

@@ -0,0 +1,18 @@
+# OS specific
+*.DS_Store
+
+# Python
+/build
+/dist
+__pycache__
+*.ipynb_checkpoints
+*.egg-info
+
+# Vim
+*.vim
+*.swk
+*.swl
+*.swm
+*.swn
+*.swo
+*.swp

LICENSE

@@ -0,0 +1,25 @@
+Modified MIT License
+
+Software Copyright (c) 2021 OpenAI
+
+We don’t claim ownership of the content you create with the DALL-E discrete VAE, so it is yours to
+do with as you please. We only ask that you use the model responsibly and clearly indicate that it
+was used.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
+associated documentation files (the "Software"), to deal in the Software without restriction,
+including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense,
+and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so,
+subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included
+in all copies or substantial portions of the Software.
+The above copyright notice and this permission notice need not be included
+with content created by the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
+INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE
+OR OTHER DEALINGS IN THE SOFTWARE.

README.md

@@ -0,0 +1,11 @@
+# Overview
+
+[[Blog]](https://openai.com/blog/dall-e/) [[Paper]](https://arxiv.org/abs/2102.12092) [[Model Card]](model_card.md) [[Usage]](notebooks/usage.ipynb)
+
+This is the official PyTorch package for the discrete VAE used for DALL·E.
+
+# Installation
+
+Before running [the example notebook](notebooks/usage.ipynb), you will need to install the package using
+
+	pip install git+https://github.com/openai/DALL-E.git

dall_e/__init__.py

@@ -0,0 +1,18 @@
+import io, requests
+import torch
+import torch.nn as nn
+
+from dall_e.encoder import Encoder
+from dall_e.decoder import Decoder
+from dall_e.utils   import map_pixels, unmap_pixels
+
+def load_model(path: str, device: torch.device = None) -> nn.Module:
+    if path.startswith('http://') or path.startswith('https://'):
+        resp = requests.get(path)
+        resp.raise_for_status()
+
+        with io.BytesIO(resp.content) as buf:
+            return torch.load(buf, map_location=device)
+    else:
+        with open(path, 'rb') as f:
+            return torch.load(f, map_location=device)

dall_e/decoder.py

@@ -0,0 +1,94 @@
+import attr
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from collections  import OrderedDict
+from functools    import partial
+from dall_e.utils import Conv2d
+
+@attr.s(eq=False, repr=False)
+class DecoderBlock(nn.Module):
+	n_in:     int = attr.ib(validator=lambda i, a, x: x >= 1)
+	n_out:    int = attr.ib(validator=lambda i, a, x: x >= 1 and x % 4 ==0)
+	n_layers: int = attr.ib(validator=lambda i, a, x: x >= 1)
+
+	device:        torch.device = attr.ib(default=None)
+	requires_grad: bool         = attr.ib(default=False)
+
+	def __attrs_post_init__(self) -> None:
+		super().__init__()
+		self.n_hid = self.n_out // 4
+		self.post_gain = 1 / (self.n_layers ** 2)
+
+		make_conv     = partial(Conv2d, device=self.device, requires_grad=self.requires_grad)
+		self.id_path  = make_conv(self.n_in, self.n_out, 1) if self.n_in != self.n_out else nn.Identity()
+		self.res_path = nn.Sequential(OrderedDict([
+				('relu_1', nn.ReLU()),
+				('conv_1', make_conv(self.n_in,  self.n_hid, 1)),
+				('relu_2', nn.ReLU()),
+				('conv_2', make_conv(self.n_hid, self.n_hid, 3)),
+				('relu_3', nn.ReLU()),
+				('conv_3', make_conv(self.n_hid, self.n_hid, 3)),
+				('relu_4', nn.ReLU()),
+				('conv_4', make_conv(self.n_hid, self.n_out, 3)),]))
+
+	def forward(self, x: torch.Tensor) -> torch.Tensor:
+		return self.id_path(x) + self.post_gain * self.res_path(x)
+
+@attr.s(eq=False, repr=False)
+class Decoder(nn.Module):
+	group_count:     int = 4
+	n_init:          int = attr.ib(default=128,  validator=lambda i, a, x: x >= 8)
+	n_hid:           int = attr.ib(default=256,  validator=lambda i, a, x: x >= 64)
+	n_blk_per_group: int = attr.ib(default=2,    validator=lambda i, a, x: x >= 1)
+	output_channels: int = attr.ib(default=3,    validator=lambda i, a, x: x >= 1)
+	vocab_size:      int = attr.ib(default=8192, validator=lambda i, a, x: x >= 512)
+
+	device:              torch.device = attr.ib(default=torch.device('cpu'))
+	requires_grad:       bool         = attr.ib(default=False)
+	use_mixed_precision: bool         = attr.ib(default=True)
+
+	def __attrs_post_init__(self) -> None:
+		super().__init__()
+
+		blk_range  = range(self.n_blk_per_group)
+		n_layers   = self.group_count * self.n_blk_per_group
+		make_conv  = partial(Conv2d, device=self.device, requires_grad=self.requires_grad)
+		make_blk   = partial(DecoderBlock, n_layers=n_layers, device=self.device,
+				requires_grad=self.requires_grad)
+
+		self.blocks = nn.Sequential(OrderedDict([
+			('input', make_conv(self.vocab_size, self.n_init, 1, use_float16=False)),
+			('group_1', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(self.n_init if i == 0 else 8 * self.n_hid, 8 * self.n_hid)) for i in blk_range],
+				('upsample', nn.Upsample(scale_factor=2, mode='nearest')),
+			]))),
+			('group_2', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(8 * self.n_hid if i == 0 else 4 * self.n_hid, 4 * self.n_hid)) for i in blk_range],
+				('upsample', nn.Upsample(scale_factor=2, mode='nearest')),
+			]))),
+			('group_3', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(4 * self.n_hid if i == 0 else 2 * self.n_hid, 2 * self.n_hid)) for i in blk_range],
+				('upsample', nn.Upsample(scale_factor=2, mode='nearest')),
+			]))),
+			('group_4', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(2 * self.n_hid if i == 0 else 1 * self.n_hid, 1 * self.n_hid)) for i in blk_range],
+			]))),
+			('output', nn.Sequential(OrderedDict([
+				('relu', nn.ReLU()),
+				('conv', make_conv(1 * self.n_hid, 2 * self.output_channels, 1)),
+			]))),
+		]))
+
+	def forward(self, x: torch.Tensor) -> torch.Tensor:
+		if len(x.shape) != 4:
+			raise ValueError(f'input shape {x.shape} is not 4d')
+		if x.shape[1] != self.vocab_size:
+			raise ValueError(f'input has {x.shape[1]} channels but model built for {self.vocab_size}')
+		if x.dtype != torch.float32:
+			raise ValueError('input must have dtype torch.float32')
+
+		return self.blocks(x)

dall_e/encoder.py

@@ -0,0 +1,93 @@
+import attr
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from collections  import OrderedDict
+from functools    import partial
+from dall_e.utils import Conv2d
+
+@attr.s(eq=False, repr=False)
+class EncoderBlock(nn.Module):
+	n_in:     int = attr.ib(validator=lambda i, a, x: x >= 1)
+	n_out:    int = attr.ib(validator=lambda i, a, x: x >= 1 and x % 4 ==0)
+	n_layers: int = attr.ib(validator=lambda i, a, x: x >= 1)
+
+	device:        torch.device = attr.ib(default=None)
+	requires_grad: bool         = attr.ib(default=False)
+
+	def __attrs_post_init__(self) -> None:
+		super().__init__()
+		self.n_hid = self.n_out // 4
+		self.post_gain = 1 / (self.n_layers ** 2)
+
+		make_conv     = partial(Conv2d, device=self.device, requires_grad=self.requires_grad)
+		self.id_path  = make_conv(self.n_in, self.n_out, 1) if self.n_in != self.n_out else nn.Identity()
+		self.res_path = nn.Sequential(OrderedDict([
+				('relu_1', nn.ReLU()),
+				('conv_1', make_conv(self.n_in,  self.n_hid, 3)),
+				('relu_2', nn.ReLU()),
+				('conv_2', make_conv(self.n_hid, self.n_hid, 3)),
+				('relu_3', nn.ReLU()),
+				('conv_3', make_conv(self.n_hid, self.n_hid, 3)),
+				('relu_4', nn.ReLU()),
+				('conv_4', make_conv(self.n_hid, self.n_out, 1)),]))
+
+	def forward(self, x: torch.Tensor) -> torch.Tensor:
+		return self.id_path(x) + self.post_gain * self.res_path(x)
+
+@attr.s(eq=False, repr=False)
+class Encoder(nn.Module):
+	group_count:     int = 4
+	n_hid:           int = attr.ib(default=256,  validator=lambda i, a, x: x >= 64)
+	n_blk_per_group: int = attr.ib(default=2,    validator=lambda i, a, x: x >= 1)
+	input_channels:  int = attr.ib(default=3,    validator=lambda i, a, x: x >= 1)
+	vocab_size:      int = attr.ib(default=8192, validator=lambda i, a, x: x >= 512)
+
+	device:              torch.device = attr.ib(default=torch.device('cpu'))
+	requires_grad:       bool         = attr.ib(default=False)
+	use_mixed_precision: bool         = attr.ib(default=True)
+
+	def __attrs_post_init__(self) -> None:
+		super().__init__()
+
+		blk_range  = range(self.n_blk_per_group)
+		n_layers   = self.group_count * self.n_blk_per_group
+		make_conv  = partial(Conv2d, device=self.device, requires_grad=self.requires_grad)
+		make_blk   = partial(EncoderBlock, n_layers=n_layers, device=self.device,
+				requires_grad=self.requires_grad)
+
+		self.blocks = nn.Sequential(OrderedDict([
+			('input', make_conv(self.input_channels, 1 * self.n_hid, 7)),
+			('group_1', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(1 * self.n_hid, 1 * self.n_hid)) for i in blk_range],
+				('pool', nn.MaxPool2d(kernel_size=2)),
+			]))),
+			('group_2', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(1 * self.n_hid if i == 0 else 2 * self.n_hid, 2 * self.n_hid)) for i in blk_range],
+				('pool', nn.MaxPool2d(kernel_size=2)),
+			]))),
+			('group_3', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(2 * self.n_hid if i == 0 else 4 * self.n_hid, 4 * self.n_hid)) for i in blk_range],
+				('pool', nn.MaxPool2d(kernel_size=2)),
+			]))),
+			('group_4', nn.Sequential(OrderedDict([
+				*[(f'block_{i + 1}', make_blk(4 * self.n_hid if i == 0 else 8 * self.n_hid, 8 * self.n_hid)) for i in blk_range],
+			]))),
+			('output', nn.Sequential(OrderedDict([
+				('relu', nn.ReLU()),
+				('conv', make_conv(8 * self.n_hid, self.vocab_size, 1, use_float16=False)),
+			]))),
+		]))
+
+	def forward(self, x: torch.Tensor) -> torch.Tensor:
+		if len(x.shape) != 4:
+			raise ValueError(f'input shape {x.shape} is not 4d')
+		if x.shape[1] != self.input_channels:
+			raise ValueError(f'input has {x.shape[1]} channels but model built for {self.input_channels}')
+		if x.dtype != torch.float32:
+			raise ValueError('input must have dtype torch.float32')
+
+		return self.blocks(x)

dall_e/utils.py

@@ -0,0 +1,59 @@
+import attr
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+logit_laplace_eps: float = 0.1
+
+@attr.s(eq=False)
+class Conv2d(nn.Module):
+	n_in:  int = attr.ib(validator=lambda i, a, x: x >= 1)
+	n_out: int = attr.ib(validator=lambda i, a, x: x >= 1)
+	kw:    int = attr.ib(validator=lambda i, a, x: x >= 1 and x % 2 == 1)
+
+	use_float16:   bool         = attr.ib(default=True)
+	device:        torch.device = attr.ib(default=torch.device('cpu'))
+	requires_grad: bool         = attr.ib(default=False)
+
+	def __attrs_post_init__(self) -> None:
+		super().__init__()
+
+		w = torch.empty((self.n_out, self.n_in, self.kw, self.kw), dtype=torch.float32,
+			device=self.device, requires_grad=self.requires_grad)
+		w.normal_(std=1 / math.sqrt(self.n_in * self.kw ** 2))
+
+		b = torch.zeros((self.n_out,), dtype=torch.float32, device=self.device,
+			requires_grad=self.requires_grad)
+		self.w, self.b = nn.Parameter(w), nn.Parameter(b)
+
+	def forward(self, x: torch.Tensor) -> torch.Tensor:
+		if self.use_float16 and 'cuda' in self.w.device.type:
+			if x.dtype != torch.float16:
+				x = x.half()
+
+			w, b = self.w.half(), self.b.half()
+		else:
+			if x.dtype != torch.float32:
+				x = x.float()
+
+			w, b = self.w, self.b
+
+		return F.conv2d(x, w, b, padding=(self.kw - 1) // 2)
+
+def map_pixels(x: torch.Tensor) -> torch.Tensor:
+	if len(x.shape) != 4:
+		raise ValueError('expected input to be 4d')
+	if x.dtype != torch.float:
+		raise ValueError('expected input to have type float')
+
+	return (1 - 2 * logit_laplace_eps) * x + logit_laplace_eps
+
+def unmap_pixels(x: torch.Tensor) -> torch.Tensor:
+	if len(x.shape) != 4:
+		raise ValueError('expected input to be 4d')
+	if x.dtype != torch.float:
+		raise ValueError('expected input to have type float')
+
+	return torch.clamp((x - logit_laplace_eps) / (1 - 2 * logit_laplace_eps), 0, 1)

model_card.md

@@ -0,0 +1,41 @@
+# Model Card: DALL·E dVAE
+
+Following [Model Cards for Model Reporting (Mitchell et al.)](https://arxiv.org/abs/1810.03993) and [Lessons from
+Archives (Jo & Gebru)](https://arxiv.org/pdf/1912.10389.pdf), we're providing some information about about the discrete
+VAE (dVAE) that was used to train DALL·E.
+
+## Model Details
+
+The dVAE was developed by researchers at OpenAI to reduce the memory footprint of the transformer trained on the
+text-to-image generation task. The details involved in training the dVAE are described in [the paper][dalle_paper]. This
+model card describes the first version of the model, released in February 2021. The model consists of a convolutional
+encoder and decoder whose architectures are described [here](dall_e/encoder.py) and [here](dall_e/decoder.py), respectively.
+For questions or comments about the models or the code release, please file a Github issue.
+
+## Model Use
+
+### Intended Use
+
+The model is intended for others to use for training their own generative models.
+
+### Out-of-Scope Use Cases
+
+This model is inappropriate for high-fidelity image processing applications. We also do not recommend its use as a
+general-purpose image compressor.
+
+## Training Data
+
+The model was trained on publicly available text-image pairs collected from the internet. This data consists partly of
+[Conceptual Captions][cc] and a filtered subset of [YFCC100M][yfcc100m]. We used a subset of the filters described in
+[Sharma et al.][cc_paper] to construct this dataset; further details are described in [our paper][dalle_paper]. We will
+not be releasing the dataset.
+
+## Performance and Limitations
+
+The heavy compression from the encoding process results in a noticeable loss of detail in the reconstructed images. This
+renders it inappropriate for applications that require fine-grained details of the image to be preserved.
+
+[dalle_paper]: https://arxiv.org/abs/2102.12092
+[cc]: https://ai.google.com/research/ConceptualCaptions
+[cc_paper]: https://www.aclweb.org/anthology/P18-1238/
+[yfcc100m]: http://projects.dfki.uni-kl.de/yfcc100m/