In my final a number of articles I talked about generative deep studying algorithms, which largely are associated to textual content technology duties. So, I feel it might be attention-grabbing to change to generative algorithms for picture technology now. We knew that these days there have been loads of deep studying fashions specialised for producing photographs on the market, resembling Autoencoder, Variational Autoencoder (VAE), Generative Adversarial Community (GAN) and Neural Fashion Switch (NST). I truly acquired a few of my writings about these subjects posted on Medium as effectively. I present you the hyperlinks on the finish of this text if you wish to learn them.
In immediately’s article, I wish to focus on the so-called diffusion mannequin — probably the most impactful fashions within the area of deep studying for picture technology. The concept of this algorithm was first proposed within the paper titled Deep Unsupervised Studying utilizing Nonequilibrium Thermodynamics written by Sohl-Dickstein et al. again in 2015 [1]. Their framework was then developed additional by Ho et al. in 2020 of their paper titled Denoising Diffusion Probabilistic Fashions [2]. DDPM was later tailored by OpenAI and Google to develop DALLE-2 and Imagen, which we knew that these fashions have spectacular capabilities to generate high-quality photographs.
How Diffusion Mannequin Works
Typically talking, diffusion mannequin works by producing picture from noise. We are able to consider it like an artist reworking a splash of paint on a canvas into an exquisite art work. So as to take action, the diffusion mannequin must be educated first. There are two fundamental steps required to be adopted to coach the mannequin, particularly ahead diffusion and backward diffusion.
As you may see within the above determine, ahead diffusion is a course of the place Gaussian noise is utilized to the unique picture iteratively. We preserve including the noise till the picture is totally unrecognizable, at which level we will say that the picture now lies within the latent house. Totally different from Autoencoders and GANs the place the latent house usually has a decrease dimension than the unique picture, the latent house in DDPM maintains the very same dimensionality as the unique one. This noising course of follows the precept of a Markov Chain, that means that the picture at timestep t is affected solely by timestep t-1. Ahead diffusion is taken into account straightforward since what we mainly do is simply including some noise step-by-step.
The second coaching section known as backward diffusion, which our goal right here is to take away the noise little by little till we acquire a transparent picture. This course of follows the precept of the reverse Markov Chain, the place the picture at timestep t-1 can solely be obtained based mostly on the picture at timestep t. Such a denoising course of is actually tough since we have to guess which pixels are noise and which of them belong to the precise picture content material. Thus, we have to make use of a neural community mannequin to take action.
DDPM makes use of U-Web as the idea of the deep studying structure for backward diffusion. Nevertheless, as an alternative of utilizing the unique U-Web mannequin [4], we have to make a number of modifications to it in order that it will likely be extra appropriate for our job. In a while, I’m going to coach this mannequin on the MNIST Handwritten Digit dataset [5], and we’ll see whether or not it may generate comparable photographs.
Properly, that was just about all the elemental ideas it’s essential to find out about diffusion fashions for now. Within the subsequent sections we’re going to get even deeper into the main points whereas implementing the algorithm from scratch.
PyTorch Implementation
We’re going to begin by importing the required modules. In case you’re not but aware of the imports beneath, each torch and torchvision are the libraries we’ll use for making ready the mannequin and the dataset. In the meantime, matplotlib and tqdm will assist us show photographs and progress bars.
# Codeblock 1
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from torch.optim import Adam
from torch.utils.knowledge import DataLoader
from torchvision import datasets, transforms
from tqdm import tqdmBecause the modules have been imported, the subsequent factor to do is to initialize some config parameters. Take a look at the Codeblock 2 beneath for the main points.
# Codeblock 2
IMAGE_SIZE = 28 #(1)
NUM_CHANNELS = 1 #(2)
BATCH_SIZE = 2
NUM_EPOCHS = 10
LEARNING_RATE = 0.001
NUM_TIMESTEPS = 1000 #(3)
BETA_START = 0.0001 #(4)
BETA_END = 0.02 #(5)
TIME_EMBED_DIM = 32 #(6)
DEVICE = torch.system("cuda" if torch.cuda.is_available else "cpu") #(7)
DEVICE# Codeblock 2 Output
system(kind='cuda')On the traces marked with #(1) and #(2) I set IMAGE_SIZE and NUM_CHANNELS to twenty-eight and 1, which these numbers are obtained from the picture dimension within the MNIST dataset. The BATCH_SIZE, NUM_EPOCHS, and LEARNING_RATE variables are fairly simple, so I don’t assume I want to elucidate them additional.
At line #(3), the variable NUM_TIMESTEPS denotes the variety of iterations within the ahead and backward diffusion course of. Timestep 0 is the situation the place the picture is in its authentic state (the leftmost picture in Determine 1). On this case, since we set this parameter to 1000, timestep quantity 999 goes to be the situation the place the picture is totally unrecognizable (the rightmost picture in Determine 1). You will need to take into account that the selection of the variety of timesteps includes a tradeoff between mannequin accuracy and computational value. If we assign a small worth for NUM_TIMESTEPS, the inference time goes to be shorter, but the ensuing picture may not be actually good for the reason that mannequin has fewer steps to refine the picture within the backward diffusion stage. However, growing NUM_TIMESTEPS will decelerate the inference course of, however we will anticipate the output picture to have higher high quality due to the gradual denoising course of which leads to a extra exact reconstruction.
Subsequent, the BETA_START (#(4)) and BETA_END (#(5)) variables are used to manage the quantity of Gaussian noise added at every timestep, whereas TIME_EMBED_DIM (#(6)) is employed to find out the function vector size for storing the timestep info. Lastly, at line #(7) I assign “cuda” to the DEVICE variable if Pytorch detects GPU put in in our machine. I extremely advocate you run this mission on GPU since coaching a diffusion mannequin is computationally costly. Along with the above parameters, the values set for NUM_TIMESTEPS, BETA_START and BETA_END are all adopted straight from the DDPM paper [2].
The whole implementation will probably be finished in a number of steps: developing the U-Web mannequin, making ready the dataset, defining noise scheduler for the diffusion course of, coaching, and inference. We’re going to focus on every of these phases within the following sub-sections.
The U-Web Structure: Time Embedding
As I’ve talked about earlier, the idea of a diffusion mannequin is U-Web. This structure is used as a result of its output layer is appropriate to signify a picture, which undoubtedly is sensible because it was initially launched for picture segmentation job on the first place. The next determine exhibits what the unique U-Web structure seems to be like.
Nevertheless, it’s essential to change this structure in order that it may additionally have in mind the timestep info. Not solely that, since we’ll solely use MNIST dataset, we additionally have to make the mannequin smaller. Simply bear in mind the conference in deep studying that easier fashions are sometimes more practical for easy duties.
Within the determine beneath I present you the whole U-Web mannequin that has been modified. Right here you may see that the time embedding tensor is injected to the mannequin at each stage, which is able to later be finished by element-wise summation, permitting the mannequin to seize the timestep info. Subsequent, as an alternative of repeating every of the downsampling and the upsampling phases 4 instances like the unique U-Web, on this case we’ll solely repeat every of them twice. Moreover, it’s value noting that the stack of downsampling phases is also referred to as the encoder, whereas the stack of upsampling phases is usually referred to as the decoder.
Now let’s begin developing the structure by creating a category for producing the time embedding tensor, which the concept is much like the positional embedding in Transformer. See the Codeblock 3 beneath for the main points.
# Codeblock 3
class TimeEmbedding(nn.Module):
def ahead(self):
time = torch.arange(NUM_TIMESTEPS, system=DEVICE).reshape(NUM_TIMESTEPS, 1) #(1)
print(f"timett: {time.form}")
i = torch.arange(0, TIME_EMBED_DIM, 2, system=DEVICE)
denominator = torch.pow(10000, i/TIME_EMBED_DIM)
print(f"denominatort: {denominator.form}")
even_time_embed = torch.sin(time/denominator) #(1)
odd_time_embed = torch.cos(time/denominator) #(2)
print(f"even_time_embedt: {even_time_embed.form}")
print(f"odd_time_embedt: {odd_time_embed.form}")
stacked = torch.stack([even_time_embed, odd_time_embed], dim=2) #(3)
print(f"stackedtt: {stacked.form}")
time_embed = torch.flatten(stacked, start_dim=1, end_dim=2) #(4)
print(f"time_embedt: {time_embed.form}")
return time_embedWhat we mainly do within the above code is to create a tensor of measurement NUM_TIMESTEPS × TIME_EMBED_DIM (1000×32), the place each single row of this tensor will comprise the timestep info. In a while, every of the 1000 timesteps will probably be represented by a function vector of size 32. The values within the tensor themselves are obtained based mostly on the 2 equations in Determine 4. Within the Codeblock 3 above, these two equations are carried out at line #(1) and #(2), every forming a tensor having the dimensions of 1000×16. Subsequent, these tensors are mixed utilizing the code at line #(3) and #(4).
Right here I additionally print out each single step finished within the above codeblock so to get a greater understanding of what’s truly being finished within the TimeEmbedding class. For those who nonetheless need extra rationalization in regards to the above code, be at liberty to learn my earlier publish about Transformer which you’ll entry by way of the hyperlink on the finish of this text. When you clicked the hyperlink, you may simply scroll all the way in which all the way down to the Positional Encoding part.
Now let’s verify if the TimeEmbedding class works correctly utilizing the next testing code. The ensuing output exhibits that it efficiently produced a tensor of measurement 1000×32, which is strictly what we anticipated earlier.
# Codeblock 4
time_embed_test = TimeEmbedding()
out_test = time_embed_test()# Codeblock 4 Output
time : torch.Measurement([1000, 1])
denominator : torch.Measurement([16])
even_time_embed : torch.Measurement([1000, 16])
odd_time_embed : torch.Measurement([1000, 16])
stacked : torch.Measurement([1000, 16, 2])
time_embed : torch.Measurement([1000, 32])The U-Web Structure: DoubleConv
For those who take a better have a look at the modified structure, you will notice that we truly acquired a lot of repeating patterns, resembling those highlighted in yellow packing containers within the following determine.
DoubleConv class [3].These 5 yellow packing containers share the identical construction, the place they encompass two convolution layers with the time embedding tensor injected proper after the primary convolution operation is carried out. So, what we’re going to do now could be to create one other class named DoubleConv to breed this construction. Take a look at the Codeblock 5a and 5b beneath to see how I try this.
# Codeblock 5a
class DoubleConv(nn.Module):
def __init__(self, in_channels, out_channels): #(1)
tremendous().__init__()
self.conv_0 = nn.Conv2d(in_channels=in_channels, #(2)
out_channels=out_channels,
kernel_size=3,
bias=False,
padding=1)
self.bn_0 = nn.BatchNorm2d(num_features=out_channels) #(3)
self.time_embedding = TimeEmbedding() #(4)
self.linear = nn.Linear(in_features=TIME_EMBED_DIM, #(5)
out_features=out_channels)
self.conv_1 = nn.Conv2d(in_channels=out_channels, #(6)
out_channels=out_channels,
kernel_size=3,
bias=False,
padding=1)
self.bn_1 = nn.BatchNorm2d(num_features=out_channels) #(7)
self.relu = nn.ReLU(inplace=True) #(8)The 2 inputs of the __init__() methodology above offers us flexibility to configure the variety of enter and output channels (#(1)) in order that the DoubleConv class can be utilized to instantiate all of the 5 yellow packing containers just by adjusting its enter arguments. Because the identify suggests, right here we initialize two convolution layers (line #(2) and #(6)), every adopted by a batch normalization layer and a ReLU activation operate. Take into account that the 2 normalization layers have to be initialized individually (line #(3) and #(7)) since every of them has their very own trainable normalization parameters. In the meantime, the ReLU activation operate ought to solely be initialized as soon as (#(8)) as a result of it incorporates no parameters, permitting it for use a number of instances in several components of the community. At line #(4), we initialize the TimeEmbedding layer we created earlier, which is able to later be linked to a normal linear layer (#(5)). This linear layer is accountable to regulate the dimension of the time embedding tensor in order that the ensuing output may be summed with the output from the primary convolution layer in an element-wise method.
Now let’s check out the Codeblock 5b beneath to raised perceive the move of the DoubleConv block. Right here you may see that the ahead() methodology accepts two inputs: the uncooked picture x and the timestep info t as proven at line #(1). We initially course of the picture with the primary Conv-BN-ReLU sequence (#(2–4)). This Conv-BN-ReLU construction is usually used when working with CNN-based fashions, even when the illustration doesn’t explicitly present the batch normalization and the ReLU layers. Other than the picture, we then take the t-th timestep info from our embedding tensor of the corresponding picture (#(5)) and cross it by way of the linear layer (#(6)). We nonetheless have to develop the dimension of the ensuing tensor utilizing the code at line #(7) earlier than performing element-wise summation at line #(8). Lastly, we course of the ensuing tensor with the second Conv-BN-ReLU sequence (#(9–11)).
# Codeblock 5b
def ahead(self, x, t): #(1)
print(f'imagesttt: {x.measurement()}')
print(f'timestepstt: {t.measurement()}, {t}')
x = self.conv_0(x) #(2)
x = self.bn_0(x) #(3)
x = self.relu(x) #(4)
print(f'nafter first convt: {x.measurement()}')
time_embed = self.time_embedding()[t] #(5)
print(f'ntime_embedtt: {time_embed.measurement()}')
time_embed = self.linear(time_embed) #(6)
print(f'time_embed after lineart: {time_embed.measurement()}')
time_embed = time_embed[:, :, None, None] #(7)
print(f'time_embed expandedt: {time_embed.measurement()}')
x = x + time_embed #(8)
print(f'nafter summationtt: {x.measurement()}')
x = self.conv_1(x) #(9)
x = self.bn_1(x) #(10)
x = self.relu(x) #(11)
print(f'after second convt: {x.measurement()}')
return xTo see if our DoubleConv implementation works correctly, we’re going to take a look at it with the Codeblock 6 beneath. Right here I wish to simulate the very first occasion of this block, which corresponds to the leftmost yellow field in Determine 5. To take action, we have to we have to set the in_channels and out_channels parameters to 1 and 64, respectively (#(1)). Subsequent, we initialize two enter tensors, particularly x_test and t_test. The x_test tensor has the dimensions of two×1×28×28, representing a batch of two grayscale photographs having the dimensions of 28×28 (#(2)). Take into account that that is only a dummy tensor of random values which will probably be changed with the precise photographs from MNIST dataset later within the coaching section. In the meantime, t_test is a tensor containing the timestep numbers of the corresponding photographs (#(3)). The values for this tensor are randomly chosen between 0 and NUM_TIMESTEPS (1000). Word that the datatype of this tensor have to be an integer for the reason that numbers will probably be used for indexing, as proven at line #(5) again in Codeblock 5b. Lastly, at line #(4) we cross each x_test and t_test tensors to the double_conv_test layer.
By the way in which, I re-run the earlier codeblocks with the print() capabilities eliminated previous to operating the next code in order that the outputs will look neater.
# Codeblock 6
double_conv_test = DoubleConv(in_channels=1, out_channels=64).to(DEVICE) #(1)
x_test = torch.randn((BATCH_SIZE, NUM_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)).to(DEVICE) #(2)
t_test = torch.randint(0, NUM_TIMESTEPS, (BATCH_SIZE,)).to(DEVICE) #(3)
out_test = double_conv_test(x_test, t_test) #(4)# Codeblock 6 Output
photographs : torch.Measurement([2, 1, 28, 28]) #(1)
timesteps : torch.Measurement([2]), tensor([468, 304], system='cuda:0') #(2)
after first conv : torch.Measurement([2, 64, 28, 28]) #(3)
time_embed : torch.Measurement([2, 32]) #(4)
time_embed after linear : torch.Measurement([2, 64])
time_embed expanded : torch.Measurement([2, 64, 1, 1]) #(5)
after summation : torch.Measurement([2, 64, 28, 28]) #(6)
after second conv : torch.Measurement([2, 64, 28, 28]) #(7)The form of our authentic enter tensors may be seen at traces #(1) and #(2) within the above output. Particularly at line #(2), I additionally print out the 2 timesteps that we chosen randomly. On this instance we assume that every of the 2 photographs within the x tensor are already noised with the noise stage from 468-th and 304-th timesteps previous to being fed into the community. We are able to see that the form of the picture tensor x adjustments to 2×64×28×28 after being handed by way of the primary convolution layer (#(3)). In the meantime, the dimensions of our time embedding tensor turns into 2×32 (#(4)), which is obtained by extracting rows 468 and 304 from the unique embedding of measurement 1000×32. With a view to enable element-wise summation to be carried out (#(6)), we have to map the 32-dimensional time embedding vectors into 64 and develop their axes, leading to a tensor of measurement 2×64×1×1 (#(5)) in order that it may be broadcast to the two×64×28×28 tensor. After the summation is finished, we then cross the tensor by way of the second convolution layer, at which level the tensor dimension doesn’t change in any respect (#(7)).
The U-Web Structure: Encoder
As now we have efficiently carried out the DoubleConv block, the subsequent step to do is to implement the so-called DownSample block. In Determine 6 beneath, this corresponds to the components enclosed within the crimson field.
DownSample blocks [3].The aim of a DownSample block is to scale back the spatial dimension of a picture, however you will need to word that on the identical time it will increase the variety of channels. With a view to obtain this, we will merely stack a DoubleConv block and a maxpooling operation. On this case the pooling makes use of 2×2 kernel measurement with the stride of two, inflicting the spatial dimension of the picture to be twice as small because the enter. The implementation of this block may be seen in Codeblock 7 beneath.
# Codeblock 7
class DownSample(nn.Module):
def __init__(self, in_channels, out_channels): #(1)
tremendous().__init__()
self.double_conv = DoubleConv(in_channels=in_channels, #(2)
out_channels=out_channels)
self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2) #(3)
def ahead(self, x, t): #(4)
print(f'originaltt: {x.measurement()}')
print(f'timestepstt: {t.measurement()}, {t}')
convolved = self.double_conv(x, t) #(5)
print(f'nafter double convt: {convolved.measurement()}')
maxpooled = self.maxpool(convolved) #(6)
print(f'after poolingtt: {maxpooled.measurement()}')
return convolved, maxpooled #(7)Right here I set the __init__() methodology to take variety of enter and output channels in order that we will use it for creating the 2 DownSample blocks highlighted in Determine 6 without having to put in writing them in separate lessons (#(1)). Subsequent, the DoubleConv and the maxpooling layers are initialized at line #(2) and #(3), respectively. Do not forget that for the reason that DoubleConv block accepts picture x and the corresponding timestep t because the inputs, we additionally have to set the ahead() methodology of this DownSample block such that it accepts each of them as effectively (#(4)). The data contained in x and t are then mixed as the 2 tensors are processed by the double_conv layer, which the output is saved within the variable named convolved (#(5)). Afterwards, we now truly carry out the downsampling with the maxpooling operation at line #(6), producing a tensor named maxpooled. You will need to word that each the convolved and maxpooled tensors are going to be returned, which is basically finished as a result of we’ll later convey maxpooled to the subsequent downsampling stage, whereas the convolved tensor will probably be transferred on to the upsampling stage within the decoder by way of skip-connections.
Now let’s take a look at the DownSample class utilizing the Codeblock 8 beneath. The enter tensors used listed here are precisely the identical as those in Codeblock 6. Primarily based on the ensuing output, we will see that the pooling operation efficiently transformed the output of the DoubleConv block from 2×64×28×28 (#(1)) to 2×64×14×14 (#(2)), indicating that our DownSample class works correctly.
# Codeblock 8
down_sample_test = DownSample(in_channels=1, out_channels=64).to(DEVICE)
x_test = torch.randn((BATCH_SIZE, NUM_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)).to(DEVICE)
t_test = torch.randint(0, NUM_TIMESTEPS, (BATCH_SIZE,)).to(DEVICE)
out_test = down_sample_test(x_test, t_test)# Codeblock 8 Output
authentic : torch.Measurement([2, 1, 28, 28])
timesteps : torch.Measurement([2]), tensor([468, 304], system='cuda:0')
after double conv : torch.Measurement([2, 64, 28, 28]) #(1)
after pooling : torch.Measurement([2, 64, 14, 14]) #(2)The U-Web Structure: Decoder
We have to introduce the so-called UpSample block within the decoder, which is liable for reverting the tensor within the intermediate layers to the unique picture dimension. With a view to preserve a symmetrical construction, the variety of UpSample blocks should match that of the DownSample blocks. Take a look at the Determine 7 beneath to see the place the 2 UpSample blocks are positioned.
UpSample blocks [3].Since each UpSample blocks are structurally similar, we will simply initialize a single class for them, similar to the DownSample class we created earlier. Take a look at the Codeblock 9 beneath to see how I implement it.
# Codeblock 9
class UpSample(nn.Module):
def __init__(self, in_channels, out_channels):
tremendous().__init__()
self.conv_transpose = nn.ConvTranspose2d(in_channels=in_channels, #(1)
out_channels=out_channels,
kernel_size=2, stride=2) #(2)
self.double_conv = DoubleConv(in_channels=in_channels, #(3)
out_channels=out_channels)
def ahead(self, x, t, connection): #(4)
print(f'originaltt: {x.measurement()}')
print(f'timestepstt: {t.measurement()}, {t}')
print(f'connectiontt: {connection.measurement()}')
x = self.conv_transpose(x) #(5)
print(f'nafter conv transposet: {x.measurement()}')
x = torch.cat([x, connection], dim=1) #(6)
print(f'after concattt: {x.measurement()}')
x = self.double_conv(x, t) #(7)
print(f'after double convt: {x.measurement()}')
return xWithin the __init__() methodology, we use nn.ConvTranspose2d to upsample the spatial dimension (#(1)). Each the kernel measurement and stride are set to 2 in order that the output will probably be twice as giant (#(2)). Subsequent, the DoubleConv block will probably be employed to scale back the variety of channels, whereas on the identical time combining the timestep info from the time embedding tensor (#(3)).
The move of this UpSample class is a little more sophisticated than the DownSample class. If we take a better have a look at the structure, we’ll see that that we even have a skip-connection coming straight from the encoder. Thus, we’d like the ahead() methodology to just accept one other argument along with the unique picture x and the timestep t, particularly the residual tensor connection (#(4)). The very first thing we do inside this methodology is to course of the unique picture x with the transpose convolution layer (#(5)). In truth, not solely upsampling the spatial measurement, however this layer additionally reduces the variety of channels on the identical time. Nevertheless, the ensuing tensor is then straight concatenated with connection in a channel-wise method (#(6)), inflicting it to look like no channel discount is carried out. You will need to know that at this level these two tensors are simply concatenated, that means that the knowledge from the 2 usually are not but mixed. We lastly feed these concatenated tensors to the double_conv layer (#(7)), permitting them to share info to one another by way of the learnable parameters contained in the convolution layers.
The Codeblock 10 beneath exhibits how I take a look at the UpSample class. The scale of the tensors to be handed by way of are set based on the second upsampling block, i.e., the rightmost blue field in Determine 7.
# Codeblock 10
up_sample_test = UpSample(in_channels=128, out_channels=64).to(DEVICE)
x_test = torch.randn((BATCH_SIZE, 128, 14, 14)).to(DEVICE)
t_test = torch.randint(0, NUM_TIMESTEPS, (BATCH_SIZE,)).to(DEVICE)
connection_test = torch.randn((BATCH_SIZE, 64, 28, 28)).to(DEVICE)
out_test = up_sample_test(x_test, t_test, connection_test)Within the ensuing output beneath, if we evaluate the enter tensor (#(1)) with the ultimate tensor form (#(2)), we will clearly see that the variety of channels efficiently lowered from 128 to 64, whereas on the identical time the spatial dimension elevated from 14×14 to twenty-eight×28. This primarily implies that our UpSample class is now prepared for use in the primary U-Web structure.
# Codeblock 10 Output
authentic : torch.Measurement([2, 128, 14, 14]) #(1)
timesteps : torch.Measurement([2]), tensor([468, 304], system='cuda:0')
connection : torch.Measurement([2, 64, 28, 28])
after conv transpose : torch.Measurement([2, 64, 28, 28])
after concat : torch.Measurement([2, 128, 28, 28])
after double conv : torch.Measurement([2, 64, 28, 28]) #(2)The U-Web Structure: Placing All Parts Collectively
As soon as all U-Web elements have been created, what we’re going to do subsequent is to wrap them collectively right into a single class. Take a look at the Codeblock 11a and 11b beneath for the main points.
# Codeblock 11a
class UNet(nn.Module):
def __init__(self):
tremendous().__init__()
self.downsample_0 = DownSample(in_channels=NUM_CHANNELS, #(1)
out_channels=64)
self.downsample_1 = DownSample(in_channels=64, #(2)
out_channels=128)
self.bottleneck = DoubleConv(in_channels=128, #(3)
out_channels=256)
self.upsample_0 = UpSample(in_channels=256, #(4)
out_channels=128)
self.upsample_1 = UpSample(in_channels=128, #(5)
out_channels=64)
self.output = nn.Conv2d(in_channels=64, #(6)
out_channels=NUM_CHANNELS,
kernel_size=1)You may see within the __init__() methodology above that we initialize two downsampling (#(1–2)) and two upsampling (#(4–5)) blocks, which the variety of enter and output channels are set based on the structure proven within the illustration. There are literally two further elements I haven’t defined but, particularly the bottleneck (#(3)) and the output layer (#(6)). The previous is basically only a DoubleConv block, which acts as the primary connection between the encoder and the decoder. Take a look at the Determine 8 beneath to see which elements of the community belong to the bottleneck layer. Subsequent, the output layer is a normal convolution layer which is accountable to show the 64-channel picture produced by the final UpSampling stage into 1-channel solely. This operation is finished utilizing a kernel of measurement 1×1, that means that it combines info throughout all channels whereas working independently at every pixel place.
I assume the ahead() methodology of the whole U-Web within the following codeblock is fairly simple, as what we primarily do right here is cross the tensors from one layer to a different — simply don’t neglect to incorporate the skip connections between the downsampling and upsampling blocks.
# Codeblock 11b
def ahead(self, x, t): #(1)
print(f'originaltt: {x.measurement()}')
print(f'timestepstt: {t.measurement()}, {t}')
convolved_0, maxpooled_0 = self.downsample_0(x, t)
print(f'nmaxpooled_0tt: {maxpooled_0.measurement()}')
convolved_1, maxpooled_1 = self.downsample_1(maxpooled_0, t)
print(f'maxpooled_1tt: {maxpooled_1.measurement()}')
x = self.bottleneck(maxpooled_1, t)
print(f'after bottleneckt: {x.measurement()}')
upsampled_0 = self.upsample_0(x, t, convolved_1)
print(f'upsampled_0tt: {upsampled_0.measurement()}')
upsampled_1 = self.upsample_1(upsampled_0, t, convolved_0)
print(f'upsampled_1tt: {upsampled_1.measurement()}')
x = self.output(upsampled_1)
print(f'closing outputtt: {x.measurement()}')
return xNow let’s see whether or not now we have accurately constructed the U-Web class above by operating the next testing code.
# Codeblock 12
unet_test = UNet().to(DEVICE)
x_test = torch.randn((BATCH_SIZE, NUM_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)).to(DEVICE)
t_test = torch.randint(0, NUM_TIMESTEPS, (BATCH_SIZE,)).to(DEVICE)
out_test = unet_test(x_test, t_test)# Codeblock 12 Output
authentic : torch.Measurement([2, 1, 28, 28]) #(1)
timesteps : torch.Measurement([2]), tensor([468, 304], system='cuda:0')
maxpooled_0 : torch.Measurement([2, 64, 14, 14]) #(2)
maxpooled_1 : torch.Measurement([2, 128, 7, 7]) #(3)
after bottleneck : torch.Measurement([2, 256, 7, 7]) #(4)
upsampled_0 : torch.Measurement([2, 128, 14, 14])
upsampled_1 : torch.Measurement([2, 64, 28, 28])
closing output : torch.Measurement([2, 1, 28, 28]) #(5)We are able to see within the above output that the 2 downsampling phases efficiently transformed the unique tensor of measurement 1×28×28 (#(1)) into 64×14×14 (#(2)) and 128×7×7 (#(3)), respectively. This tensor is then handed by way of the bottleneck layer, inflicting its variety of channels to develop to 256 with out altering the spatial dimension (#(4)). Lastly, we upsample the tensor twice earlier than finally shrinking the variety of channels to 1 (#(5)). Primarily based on this output, it seems to be like our mannequin is working correctly. Thus, it’s now able to be educated for our diffusion job.
Dataset Preparation
As now we have efficiently created the whole U-Web structure, the subsequent factor to do is to organize the MNIST Handwritten Digit dataset. Earlier than truly loading it, we have to outline the preprocessing steps first utilizing the transforms.Compose() methodology from Torchvision, as proven at line #(1) in Codeblock 13. There are two issues we do right here: changing the pictures into PyTorch tensors which additionally scales the pixel values from 0–255 to 0–1 (#(2)), and normalize them in order that the ultimate pixel values ranging between -1 and 1 (#(3)). Subsequent, we obtain the dataset utilizing datasets.MNIST(). On this case, we’re going to take the pictures from the coaching knowledge, therefore we use prepare=True (#(5)). Don’t neglect to cross the rework variable we initialized earlier to the rework parameter (rework=rework) so that it’ll mechanically preprocess the pictures as we load them (#(6)). Lastly, we have to make use of DataLoader to load the pictures from mnist_dataset (#(7)). The arguments I exploit for the enter parameters are meant to randomly decide BATCH_SIZE (2) photographs from the dataset in every iteration.
# Codeblock 13
rework = transforms.Compose([ #(1)
transforms.ToTensor(), #(2)
transforms.Normalize((0.5,), (0.5,)) #(3)
])
mnist_dataset = datasets.MNIST( #(4)
root='./knowledge',
prepare=True, #(5)
obtain=True,
rework=rework #(6)
)
loader = DataLoader(mnist_dataset, #(7)
batch_size=BATCH_SIZE,
drop_last=True,
shuffle=True)Within the following codeblock, I attempt to load a batch of photographs from the dataset. In each iteration, loader supplies each the pictures and the corresponding labels, therefore we have to retailer them in two separate variables: photographs and labels.
# Codeblock 14
photographs, labels = subsequent(iter(loader))
print('imagestt:', photographs.form)
print('labelstt:', labels.form)
print('min valuet:', photographs.min())
print('max valuet:', photographs.max())We are able to see within the ensuing output beneath that the photographs tensor has the dimensions of two×1×28×28 (#(1)), indicating that two grayscale photographs of measurement 28×28 have been efficiently loaded. Right here we will additionally see that the size of the labels tensor is 2, which matches the variety of the loaded photographs (#(2)). Word that on this case the labels are going to be utterly ignored. My plan right here is that I simply need the mannequin to generate any quantity it beforehand seen from the whole coaching dataset with out even understanding what quantity it truly is. Lastly, this output additionally exhibits that the preprocessing works correctly, because the pixel values now vary between -1 and 1.
# Codeblock 14 Output
photographs : torch.Measurement([2, 1, 28, 28]) #(1)
labels : torch.Measurement([2]) #(2)
min worth : tensor(-1.)
max worth : tensor(1.)Run the next code if you wish to see what the picture we simply loaded seems to be like.
# Codeblock 15
plt.imshow(photographs[0].squeeze(), cmap='grey')
plt.present()
Noise Scheduler
On this part we’re going to speak about how the ahead and backward diffusion are carried out, which the method primarily includes including or eradicating noise little by little at every timestep. It’s essential to know that we mainly need a uniform quantity of noise throughout all timesteps, the place within the ahead diffusion the picture must be utterly filled with noise precisely at timestep 1000, whereas within the backward diffusion, now we have to get the utterly clear picture at timestep 0. Therefore, we’d like one thing to manage the noise quantity for every timestep. Later on this part, I’m going to implement a category named NoiseScheduler to take action. — This may most likely be probably the most mathy part of this text, as I’ll show many equations right here. However don’t fear about that since we’ll deal with implementing these equations quite than discussing the mathematical derivations.
Now let’s check out the equations in Determine 10 which I’ll implement within the __init__() methodology of the NoiseScheduler class beneath.
NoiseScheduler class [3].# Codeblock 16a
class NoiseScheduler:
def __init__(self):
self.betas = torch.linspace(BETA_START, BETA_END, NUM_TIMESTEPS) #(1)
self.alphas = 1. - self.betas
self.alphas_cum_prod = torch.cumprod(self.alphas, dim=0)
self.sqrt_alphas_cum_prod = torch.sqrt(self.alphas_cum_prod)
self.sqrt_one_minus_alphas_cum_prod = torch.sqrt(1. - self.alphas_cum_prod)The above code works by creating a number of sequences of numbers, all of them are mainly managed by BETA_START (0.0001), BETA_END (0.02), and NUM_TIMESTEPS (1000). The primary sequence we have to instantiate is the betas itself, which is finished utilizing torch.linspace() (#(1)). What it primarily does is that it generates a 1-dimensional tensor of size 1000 ranging from 0.0001 to 0.02, the place each single ingredient on this tensor corresponds to a single timestep. The interval between every ingredient is uniform, permitting us to generate uniform quantity of noise all through all timesteps as effectively. With this betas tensor, we then compute alphas, alphas_cum_prod, sqrt_alphas_cum_prod and sqrt_one_minus_alphas_cum_prod based mostly on the 4 equations in Determine 10. In a while, these tensors will act as the idea of how the noise is generated or eliminated through the diffusion course of.
Diffusion is often finished in a sequential method. Nevertheless, the ahead diffusion course of is deterministic, therefore we will derive the unique equation right into a closed type in order that we will acquire the noise at a selected timestep with out having to iteratively add noise from the very starting. The Determine 11 beneath exhibits what the closed type of the ahead diffusion seems to be like, the place x₀ represents the unique picture whereas epsilon (ϵ) denotes a picture made up of random Gaussian noise. We are able to consider this equation as a weighted mixture, the place we mix the clear picture and the noise based on weights decided by the timestep, leading to a picture with a certain amount of noise.
The implementation of this equation may be seen in Codeblock 16b. On this forward_diffusion() methodology, x₀ and ϵ are denoted as authentic and noise. Right here it’s essential to take into account that these two enter variables are photographs, whereas sqrt_alphas_cum_prod_t and sqrt_one_minus_alphas_cum_prod_t are scalars. Thus, we have to regulate the form of those two scalars (#(1) and #(2)) in order that the operation at line #(3) may be carried out. The noisy_image variable goes to be the output of this operate, which I assume the identify is self-explanatory.
# Codeblock 16b
def forward_diffusion(self, authentic, noise, t):
sqrt_alphas_cum_prod_t = self.sqrt_alphas_cum_prod[t]
sqrt_alphas_cum_prod_t = sqrt_alphas_cum_prod_t.to(DEVICE).view(-1, 1, 1, 1) #(1)
sqrt_one_minus_alphas_cum_prod_t = self.sqrt_one_minus_alphas_cum_prod[t]
sqrt_one_minus_alphas_cum_prod_t = sqrt_one_minus_alphas_cum_prod_t.to(DEVICE).view(-1, 1, 1, 1) #(2)
noisy_image = sqrt_alphas_cum_prod_t * authentic + sqrt_one_minus_alphas_cum_prod_t * noise #(3)
return noisy_imageNow let’s speak about backward diffusion. In truth, this one is a little more sophisticated than the ahead diffusion since we’d like three extra equations right here. Earlier than I offer you these equations, let me present you the implementation first. See the Codeblock 16c beneath.
# Codeblock 16c
def backward_diffusion(self, current_image, predicted_noise, t): #(1)
denoised_image = (current_image - (self.sqrt_one_minus_alphas_cum_prod[t] * predicted_noise)) / self.sqrt_alphas_cum_prod[t] #(2)
denoised_image = 2 * (denoised_image - denoised_image.min()) / (denoised_image.max() - denoised_image.min()) - 1 #(3)
current_prediction = current_image - ((self.betas[t] * predicted_noise) / (self.sqrt_one_minus_alphas_cum_prod[t])) #(4)
current_prediction = current_prediction / torch.sqrt(self.alphas[t]) #(5)
if t == 0: #(6)
return current_prediction, denoised_image
else:
variance = (1 - self.alphas_cum_prod[t-1]) / (1. - self.alphas_cum_prod[t]) #(7)
variance = variance * self.betas[t] #(8)
sigma = variance ** 0.5
z = torch.randn(current_image.form).to(DEVICE)
current_prediction = current_prediction + sigma*z
return current_prediction, denoised_imageLater within the inference section, the backward_diffusion() methodology will probably be referred to as inside a loop that iterates NUM_TIMESTEPS (1000) instances, ranging from t = 999, continued with t = 998, and so forth all the way in which to t = 0. This operate is accountable to take away the noise from the picture iteratively based mostly on the current_image (the picture produced by the earlier denoising step), the predicted_noise (the noise predicted by U-Web within the earlier step), and the timestep info t (#(1)). In every iteration, noise removing is finished utilizing the equation proven in Determine 12, which in Codeblock 16c, this corresponds to traces #(4-5).
So long as we haven’t reached t = 0, we’ll compute the variance based mostly on the equation in Determine 13 (#(7–8)). This variance will then be used to introduce one other managed noise to simulate the stochasticity within the backward diffusion course of for the reason that noise removing equation in Determine 12 is a deterministic approximation. That is primarily additionally the rationale that we don’t calculate the variance as soon as we reached t = 0 (#(6)) since we not want so as to add extra noise because the picture is totally clear already.
Totally different from current_prediction which goals to estimate the picture of the earlier timestep (xₜ₋₁), the target of the denoised_image tensor is to reconstruct the unique picture (x₀). Thanks to those completely different aims, we’d like a separate equation to compute denoised_image, which may be seen in Determine 14 beneath. The implementation of the equation itself is written at line #(2–3).
Now let’s take a look at the NoiseScheduler class we created above. Within the following codeblock, I instantiate a NoiseScheduler object and print out the attributes related to it, that are all computed utilizing the equation in Determine 10 based mostly on the values saved within the betas attribute. Do not forget that the precise size of those tensors is NUM_TIMESTEPS (1000), however right here I solely print out the primary 6 parts.
# Codeblock 17
noise_scheduler = NoiseScheduler()
print(f'betastttt: {noise_scheduler.betas[:6]}')
print(f'alphastttt: {noise_scheduler.alphas[:6]}')
print(f'alphas_cum_prodttt: {noise_scheduler.alphas_cum_prod[:6]}')
print(f'sqrt_alphas_cum_prodtt: {noise_scheduler.sqrt_alphas_cum_prod[:6]}')
print(f'sqrt_one_minus_alphas_cum_prodt: {noise_scheduler.sqrt_one_minus_alphas_cum_prod[:6]}')# Codeblock 17 Output
betas : tensor([1.0000e-04, 1.1992e-04, 1.3984e-04, 1.5976e-04, 1.7968e-04, 1.9960e-04])
alphas : tensor([0.9999, 0.9999, 0.9999, 0.9998, 0.9998, 0.9998])
alphas_cum_prod : tensor([0.9999, 0.9998, 0.9996, 0.9995, 0.9993, 0.9991])
sqrt_alphas_cum_prod : tensor([0.9999, 0.9999, 0.9998, 0.9997, 0.9997, 0.9996])
sqrt_one_minus_alphas_cum_prod : tensor([0.0100, 0.0148, 0.0190, 0.0228, 0.0264, 0.0300])The above output signifies that our __init__() methodology works as anticipated. Subsequent, we’re going to take a look at the forward_diffusion() methodology. For those who return to Determine 16b, you will notice that forward_diffusion() accepts three inputs: authentic picture, noise picture and the timestep quantity. Let’s simply use the picture from the MNIST dataset we loaded earlier for the primary enter (#(1)) and a random Gaussian noise of the very same measurement for the second (#(2)). Run the Codeblock 18 beneath to see what these two photographs appear to be.
# Codeblock 18
picture = photographs[0] #(1)
noise = torch.randn_like(picture) #(2)
plt.imshow(picture.squeeze(), cmap='grey')
plt.present()
plt.imshow(noise.squeeze(), cmap='grey')
plt.present()
As we already acquired the picture and the noise prepared, what we have to do afterwards is to cross them to the forward_diffusion() methodology alongside the t. I truly tried to run the Codeblock 19 beneath a number of instances with t = 50, 100, 150, and so forth as much as t = 300. You may see in Determine 16 that the picture turns into much less clear because the parameter will increase. On this case, the picture goes to be utterly crammed by noise when the t is about to 999.
# Codeblock 19
noisy_image_test = noise_scheduler.forward_diffusion(picture.to(DEVICE), noise.to(DEVICE), t=50)
plt.imshow(noisy_image_test[0].squeeze().cpu(), cmap='grey')
plt.present()
Sadly, we can’t take a look at the backward_diffusion() methodology since this course of requires us to have our U-Web mannequin educated. So, let’s simply skip this half for now. I’ll present you ways we will truly use this operate later within the inference section.
Coaching
Because the U-Web mannequin, MNIST dataset, and the noise scheduler are prepared, we will now put together a operate for coaching. Earlier than we try this, I instantiate the mannequin and the noise scheduler in Codeblock 20 beneath.
# Codeblock 20
mannequin = UNet().to(DEVICE)
noise_scheduler = NoiseScheduler()All the coaching process is carried out within the prepare() operate proven in Codeblock 21. Earlier than doing something, we first initialize the optimizer and the loss operate, which on this case we use Adam and MSE, respectively (#(1–2)). What we mainly wish to do right here is to coach the mannequin such that it will likely be in a position to predict the noise contained within the enter picture, which afterward, the expected noise will probably be used as the idea of the denoising course of within the backward diffusion stage. To truly prepare the mannequin, we first have to carry out ahead diffusion utilizing the code at line #(6). This noising course of will probably be finished on the photographs tensor (#(3)) utilizing the random noise generated at line #(4). Subsequent, we take random quantity someplace between 0 and NUM_TIMESTEPS (1000) for the t (#(5)), which is basically finished as a result of we wish our mannequin to see photographs of various noise ranges as an method to enhance generalization. Because the noisy photographs have been generated, we then cross it by way of the U-Web mannequin alongside the chosen t (#(7)). The enter t right here is helpful for the mannequin because it signifies the present noise stage within the picture. Lastly, the loss operate we initialized earlier is accountable to compute the distinction between the precise noise and the expected noise from the unique picture (#(8)). So, the target of this coaching is mainly to make the expected noise as comparable as doable to the noise we generated at line #(4).
# Codeblock 21
def prepare():
optimizer = Adam(mannequin.parameters(), lr=LEARNING_RATE) #(1)
loss_function = nn.MSELoss() #(2)
losses = []
for epoch in vary(NUM_EPOCHS):
print(f'Epoch no {epoch}')
for photographs, _ in tqdm(loader):
optimizer.zero_grad()
photographs = photographs.float().to(DEVICE) #(3)
noise = torch.randn_like(photographs) #(4)
t = torch.randint(0, NUM_TIMESTEPS, (BATCH_SIZE,)) #(5)
noisy_images = noise_scheduler.forward_diffusion(photographs, noise, t).to(DEVICE) #(6)
predicted_noise = mannequin(noisy_images, t) #(7)
loss = loss_function(predicted_noise, noise) #(8)
losses.append(loss.merchandise())
loss.backward()
optimizer.step()
return lossesNow let’s run the above coaching operate utilizing the codeblock beneath. Sit again and chill out whereas ready the coaching completes. In my case, I used Kaggle Pocket book with Nvidia GPU P100 turned on, and it took round 45 minutes to complete.
# Codeblock 22
losses = prepare()If we check out the loss graph, it looks like our mannequin discovered fairly effectively as the worth is mostly reducing over time with a fast drop at early phases and a extra steady (but nonetheless reducing) development within the later phases. So, I feel we will anticipate good outcomes later within the inference section.
# Codeblock 23
plt.plot(losses)
Inference
At this level we already have our mannequin educated, so we will now carry out inference on it. Take a look at the Codeblock 24 beneath to see how I implement the inference() operate.
# Codeblock 24
def inference():
denoised_images = [] #(1)
with torch.no_grad(): #(2)
current_prediction = torch.randn((64, NUM_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)).to(DEVICE) #(3)
for i in tqdm(reversed(vary(NUM_TIMESTEPS))): #(4)
predicted_noise = mannequin(current_prediction, torch.as_tensor(i).unsqueeze(0)) #(5)
current_prediction, denoised_image = noise_scheduler.backward_diffusion(current_prediction, predicted_noise, torch.as_tensor(i)) #(6)
if ipercent100 == 0: #(7)
denoised_images.append(denoised_image)
return denoised_imagesOn the line marked with #(1) I initialize an empty checklist which will probably be used to retailer the denoising end result each 100 timesteps (#(7)). This may later enable us to see how the backward diffusion goes. The precise inference course of is encapsulated inside torch.no_grad() (#(2)). Do not forget that in diffusion fashions we generate photographs from a totally random noise, which we assume that these photographs are initially at t = 999. To implement this, we will merely use torch.randn() as proven at line #(3). Right here we initialize a tensor of measurement 64×1×28×28, indicating that we’re about to generate 64 photographs concurrently. Subsequent, we write a for loop that iterates backwards ranging from 999 to 0 (#(4)). Inside this loop, we feed the present picture and the timestep because the enter for the educated U-Web and let it predict the noise (#(5)). The precise backward diffusion is then carried out at line #(6). On the finish of the iteration, we should always get new photographs much like those now we have in our dataset. Now let’s name the inference() operate within the following codeblock.
# Codeblock 25
denoised_images = inference()Because the inference accomplished, we will now see what the ensuing photographs appear to be. The Codeblock 26 beneath is used to show the primary 42 photographs we simply generated.
# Codeblock 26
fig, axes = plt.subplots(ncols=7, nrows=6, figsize=(10, 8))
counter = 0
for i in vary(6):
for j in vary(7):
axes[i,j].imshow(denoised_images[-1][counter].squeeze().detach().cpu().numpy(), cmap='grey') #(1)
axes[i,j].get_xaxis().set_visible(False)
axes[i,j].get_yaxis().set_visible(False)
counter += 1
plt.present()
If we check out the above codeblock, you may see that the indexer of [-1] at line #(1) signifies that we solely show the pictures from the final iteration (which corresponds to timestep 0). That is the rationale that the pictures you see in Determine 18 are all free from noise. I do acknowledge that this may not be the perfect of a end result since not all of the generated photographs are legitimate digit numbers. — However hey, this as an alternative signifies that these photographs usually are not merely duplicates from the unique dataset.
Right here we will additionally visualize the backward diffusion course of utilizing the Codeblock 27 beneath. You may see within the ensuing output in Determine 19 that we initially begin from an entire random noise, which steadily disappears as we transfer to the best.
# Codeblock 27
fig, axes = plt.subplots(ncols=10, figsize=(24, 8))
sample_no = 0
timestep_no = 0
for i in vary(10):
axes[i].imshow(denoised_images[timestep_no][sample_no].squeeze().detach().cpu().numpy(), cmap='grey')
axes[i].get_xaxis().set_visible(False)
axes[i].get_yaxis().set_visible(False)
timestep_no += 1
plt.present()
Ending
There are many instructions you may go from right here. First, you would possibly most likely have to tweak the parameter configurations in Codeblock 2 if you’d like higher outcomes. Second, additionally it is doable to change the U-Web mannequin by implementing consideration layers along with the stack of convolution layers we used within the downsampling and the upsampling phases. This doesn’t assure you to acquire higher outcomes particularly for a easy dataset like this, however it’s undoubtedly value attempting. Third, you can even attempt to use a extra advanced dataset if you wish to problem your self.
On the subject of sensible functions, there are literally a lot of issues you are able to do with diffusion fashions. The only one could be for knowledge augmentation. With diffusion mannequin, we will simply generate new photographs from a selected knowledge distribution. For instance, suppose we’re engaged on a picture classification mission, however the variety of photographs within the lessons are imbalanced. To handle this downside, it’s doable for us to take the pictures from the minority class and feed them right into a diffusion mannequin. By doing so, we will ask the educated diffusion mannequin to generate quite a lot of samples from that class as many as we wish.
And effectively, that’s just about every thing in regards to the principle and the implementation of diffusion mannequin. Thanks for studying, I hope you study one thing new immediately!
You may entry the code used on this mission by way of this hyperlink. Listed here are additionally the hyperlinks to my earlier articles about Autoencoder, Variational Autoencoder (VAE), Neural Fashion Switch (NST), and Transformer.
References
[1] Jascha Sohl-Dickstein et al. Deep Unsupervised Studying utilizing Nonequilibrium Thermodynamics. Arxiv. https://arxiv.org/pdf/1503.03585 [Accessed December 27, 2024].
[2] Jonathan Ho et al. Denoising Diffusion Probabilistic Fashions. Arxiv. https://arxiv.org/pdf/2006.11239 [Accessed December 27, 2024].
[3] Picture created initially by writer.
[4] Olaf Ronneberger et al. U-Web: Convolutional Networks for Biomedical
Picture Segmentation. Arxiv. https://arxiv.org/pdf/1505.04597 [Accessed December 27, 2024].
[5] Yann LeCun et al. The MNIST Database of Handwritten Digits. https://yann.lecun.com/exdb/mnist/ [Accessed December 30, 2024] (Inventive Commons Attribution-Share Alike 3.0 license).
[6] Ashish Vaswani et al. Consideration Is All You Want. Arxiv. https://arxiv.org/pdf/1706.03762 [Accessed September 29, 2024].
