user_q
stringlengths
3
20
question
stringlengths
20
31k
user_a
stringclasses
5 values
answer
stringlengths
18
302
Sherzod_Bek
Hi,I did binary classification(dog&cat) output is between 0 and 1. It works perfectly with two classes.But when I test totally different image(human, nature, etc) output is closer to 1(class1) or closer to 0(class2).I think it should show around 0.5. because image doesn’t look like any class.could you explain reason...
KFrank
Hi Sherzod! Here’s my speculative way of looking at this: You train your classifier to predict one of two classes. When it gets it right, it gets rewarded for predicting that class with high certainty, so there is some bias towards making high-certainty predictions. Because you don’t train w…
Charamba
Hi Guys,I would like to remove zero values of a tensor and “join” the non-zero values in each row of a tensor in format [B, C, H, W]. A naive way would to doout_x = x[x!=0], this approach is bad because would destruct the Tensor dimensions.For resume, I would like to transform an input tensor like this:in_x = torch.ten...
KFrank
Hi Luiz! At the cost of n log (n) time complexity, you can use argsort()* and then use gather() to index back into your input tensor: >>> import torch >>> torch.__version__ '1.9.0' >>> in_x = torch.tensor([[[[0., 0., 2., 20., 250., 0., 0., 0., 0., 0., 250., ... …
LuukJ
Hi everyone,I am currently working on a setup where the output of my network is first modified by my own module before computing the loss and backpropagating this, as follows:class NeuralNet(nn.Module): def __init__(self): super(NeuralNet, self).__init__() self.Layer1 = Layer() self...
KFrank
Hi Luuk! Yes, but … You won’t want to bypass your Own_module backpropagation entirely. Suppose Own_module were simply return -input, that is, it just flipped the sign of (the gradient of) the loss function? Then you would be training your model to maximize the loss function, your predictions…
gkrisp9
Hello, I am new to PyTorch and I want to make a classifier for 3D DICOM MRIs. I want to use the pretrained resnet18 from monai library but I am confused with the input dimensions of the tensor. The shape of the images in my dataloader is [2,160,256,256] where 2 is the batch_size, 160 is the number of dicom images for e...
KFrank
Hi gkrisp! It would almost certainly be worse, if I understand correctly that you are working with 3d images. The point is that your images have substantive spatial structure. That is, just as a pixel with x = 17 is right next to x = 18, but far away from x = 148, a slice in your 3d “z-stack”…
Federico_Ottomano
Hello,I’m wondering if there is a pretty straightforward way (possibly this is trivial) to implement a neural network in PyTorch with a variable number of input features.Many thanks,Federico
KFrank
Hi Frederico! Focusing on your word “standard,” no, a fully-connected network will not be able to accept a variable number of input features. Let’s say you have an input batch of shape [nBatch, nFeatures] and the first network layer is Linear (in_features, out_features). If nFeatures != in_fea…
codeflux
Hi,I am implementing an A2C algorithm, the loss function for the critic is simply the advantage function and that works.For the actor I use this:loss = -distribution.log_prob(sample)*advantage.detach() loss.requires_grad = True loss.backwards()where advantage is calculated using the critic and sample is from sampling f...
KFrank
Hi Codeflux! I believe that MultivariateNormal and Normal have the same behavior in this regard. In general, you can’t differentiate or back propagate through calling .sample() on a Distribution. This is true for Normal, as well as MultivariateNormal. In contrast, you typically can backpropa…
HmmmmML
Hi! I have a very simple exampleModulebelow, where I have acircular buffer, and attempt to learn a_delayparameter (i.e. distance between a read and write pointer) that transforms an input sin wave into a phase-shifted and zero-padded output. The goal is to emulate adelay line(as in DSP), learning an unknown delay param...
KFrank
Hi Hmmm! Although _delay is a Parameter and potentially trainable, at a technical level, as soon as you call .item() on it you “break the computation graph,” as the result of .item() is no longer a pytorch tensor and therefore no longer participates in autograd. At a more conceptual level, you…
mmitjans
Hi,I’m trying to build a convolutional 2-D layer for 3-channel images which applies a different convolution per channel. This brought me to investigate thegroupsparameter innn.Conv2d. If I’m not mistaken, to do this I should simply create aConv2dlayer in a manner similar than:conv_layer = Conv2d(3,3,(1,5),groups=3)For ...
KFrank
Hi Marc! Your understanding of groups does seem to be correct. However, in your test you’ve overlooked Conv2d’s bias. You can either turn bias off (bias = False) or copy bias over from conv1 to conv2, along with weight: >>> import torch >>> import torch.nn as nn >>> >>> I = torch.randn(10,3,4…
111414
I have two real-valued matrices A and B with the shapes of (m, n) and (n, p), respectively.I know there are fast APIs (e.g. mm, matmul, einsum) in PyTorch to achieve matrix multiplication, but I want to knowif there exists an efficient way to change the dimension-reduction function from inner product to the operation I...
KFrank
Hi Bin! I think, in general, if your g() is built from a broadcastable operation and a reduction operation, you can accomplish your goal with pytorch broadcasting. In your specific case, you can use unsqueeze() in the appropriate locations to replace the nested loops with broadcasting, and t…
my3bikaht
Documentation mentions that it is possible to pass per class probabilities as a target.The target that this criterion expects should contain either:…Probabilities for each class;…Target: … If containing class probabilities, same shape as the input.which also comes with example:>>> # Example of target with class probabi...
KFrank
Hi Sergey! Support for “soft,” probabilistic targets for CrossEntropyLoss is new as of (I believe) the current stable version, 1.10.0. Your best bet will be to upgrade to the current stable release. Best. K. Frank
droid
I have a vanilla implementation of UNet, which I want to use for multiclass segmentation (where each pixel can belong to many classes). I am interested in advice on which loss function to select in this application.This long threadsuggests usingCrossEntropyLossat first, before recommendingBCELoss. Is one of these metho...
KFrank
Hi Droid! Let me distinguish between a (single-label) multi-class problem and a multi-label, multi-class problem. In both cases you have multiple classes (one of which might be a “background” or catch-all “other” class). In the single-label case, each pixel (or more generally, each item to b…
blu_head
Hi everyone,I’ve been trying to optimize images from a specific class and I created the tensor training_data_new myself (serves as a training dataloader). However, it still gives me the error “can’t optimize a non-leaf Tensor” on the line where I create optimizer2. I have tried casting it to float and just multiplying ...
KFrank
Hi Blu! The seemingly innocuous image = image * 1.0 creates a new image (that is then assigned to the python reference variable image) that, being the result of a tensor computation, is no longer a leaf variable. Consider: >>> import torch >>> torch.__version__ '1.9.0' >>> image = torch.zeros…
snowball
I’m slightly confused as to what param.grad is calculating mathematically?Here, it says “Computes and returns the sum of gradients of outputs with respect to the inputs.”https://pytorch.org/docs/stable/autograd.htmlCan someone help me interpret this?Suppose I’m looking at the weight of one neuron in a fully connected h...
KFrank
Hi Snowball! Yes. Not exactly. Let w be a Parameter (or for than matter, just a Tensor that has requires_grad = True, but is not wrapped in a Parameter), and let L be a scalar (that is, a tensor with a single element) that has been calculated from w (and a bunch of other Parameters and non-P…
hankdikeman
I have a custom loss function defined and I hit a wall debugging it. It is designed to return loss that is scaled according to the output value:import torch from torch import nn import torch.nn.functional as F class ConditionalMeanRelativeLoss(nn.Module): def __init__(self): super(ConditionalMeanRelativeLo...
KFrank
Hi Henry! It looks like your issue is due to a troublesome bug in the innards of autograd – not specific to torch.where(), but in lower-level infrastructure. However, in your use case, you can work around it by clamping the denominator of your potential divide-by-zero away from zero. Here is …
enemis
Hi,I havent found exactly this problem before, and keen if anyone has any ideas. I have a 2d matrixxcontaining integer values. I want to count how many of each value per row and return a new 2d matrixnum_of_valuethat has the same number of rows, but the columns represent different values. So in num_of_value[0,1] is the...
KFrank
Hi Simen! Please try torch.nn.functional.one_hot (x).sum (dim = 1). Best. K. Frank
zvant
I always thought 32-bits floats should be sufficient for most ML calculations. I tested the actual precision of a simple matrix multiplication operation on NumPy, PyTorch CPU, and PyTorch CUDA. Here is my code:import numpy as np import torch np.random.seed(0) M = np.random.randn(1000, 1000).astype(np.float64) for T_n...
KFrank
Hi Zvant! Depending on your GPU, nvidia might be switching you over by default to the misleading (dishonestly?) named “tf32” floating-point arithmetic. (tf32 is essentially half-precision floating-point.) You can try turning tf32 off with: torch.backends.cuda.matmul.allow_tf32 = False See t…
Qiyao_Wei
Hi all,Even though torch.symeig() seems to give accurate results most of the time, I ran a test case today and noticed that the eigenvalues and eigenvectors don’t make sense. My Pytorch version is 1.8.0, see following minimal reproducible example—import torcha = torch.tensor([[1,3,4],[2,2,4],[8,7,6]]).float()d,q = torc...
KFrank
Hi Qiyao! Your matrix, a, is not symmetric, but torch.symeig() expects a symmetric matrix to be passed in. Use torch.eig() for a general, square, non-symmetric matrix. (In practice, I believe that symeig() just uses the upper triangle of the matrix you pass in.) Best. K. Frank
danishnazir
Hi I have a normalized data which is from [-1,1] with mean 0 and standard deviation 1. I want to map it to [0,1] in order to give to the network. Can anyone tell me here how does this remapping work and will we get any benefit of normalization even after changing the range to [0,1]?
KFrank
Hi Danish! Note that taking what you’ve written literally this can only be true if half of your data is -1 and half is 1. You probably mean something a little different. mapped_data = (data + 1) / 2 If you are feeding your data to an already-trained network that was trained with data in th…
afsaneh_ebrahimi
I’m trying to set one element of weight to 1 and then hold it until the end (prevent it from updating in the next epochs). I know I can set requires_grad but I just want it for one element?
KFrank
Hi Afsaneh! The short answer is to set the specific element in question back to 1 after calling opt.step(). As you’ve recognized, requires_grad applies to an entire tensor, not separately to individual elements. It is also true that an Optimizer applies to entire tensors, and not to individua…
mpalaourg
Hello everyone,I have a -nooby- question about gradient computation. Let’s say I have an input matrixXwith shape of[T, T]. I want to process this matrix in a block diagonal way, where these blocks will have a shape of[B, B]with (B << T).For the first implementation and to check if my logic make sense, I went with a sim...
KFrank
Hi George! I believe that your discrepancy is due to (accumulated) round-off error. An easy way to test this is to repeat your computation using double precision (by using torch.double tensors). If the discrepancy is due to round-off error it should drop dramatically, say to about 1e-13. If …
ZimoNitrome
I have a network with a linear head followed by a sigmoid activation function. Most of my output neurons should be in a range 0-1 and thus sigmoid makes sense, but some of the neurons should output normalized (standardized) colors.So for the last three neurons I dotorch.logit(x)as they can be in the range [-1.3, 1.4]. ...
KFrank
Hi Zimo! Instead of applying sigmoid() to all of your neurons and then applying logit() to your color neurons to undo the sigmoid(), apply sigmoid() only to your non-color neurons so that the color neurons never get transformed. (Applying sigmoid() to a large-enough (non-inf) value will map it …
ZimoNitrome
I am trying to train a model to output the correct angle. I am currently restricting my outputs to [0, 2pi]:2 * torch.pi * torch.sigmoid(logits)My question is: What if the prediction is something like 1.9pi, but the target is 0.1pi. The loss should in theory be something like 0.2pi but in my current situation it will b...
KFrank
Hi Zimo! Rather than outputting a single value that is your angle, I think it makes more geometric sense to output two values, x and y, on the ray that defines your angle. Then regulate your x and y by pushing them towards (but not constraining them to lie on) the unit circle with a term like…
gemsanyou
I know that we can’t call backward on vector Tensor without reducing it to a scalar. However, I also need the gradient to be a vector, with each column is the gradient of each column of the Loss with respect to the parameters, i.e.,import torch x = torch.tensor([4.0], requires_grad=True) L1 = torch.sin(x) * torch.cos(x...
KFrank
Hi Kadek! Does autograd’sjacobian()do what you want? Best. K. Frank
BramVanroy
Doing multi-label binary classification with BCEWithLogitsLoss, but I get a RunTimeError “RuntimeError: result type Float can’t be cast to the desired output type Long”. Running 1.8.1 on Windows but the problem also occurred on 1.4. Am I doing something wrong with respect to multi-labels, perhaps?from torch import Flo...
KFrank
Hi Bram! Two comments: First, as you’ve seen, BCEWithLogitsLoss requires its target to be a float tensor, not long (or a double tensor, if the input is double). And yes, converting to float (labels.float()) is the correct solution. Second, as to why: Unlike pytorch’s CrossEntropyLoss, BCE…
ljmzlh
How to implement this (essentially calculating average of vectors)batch_size, d= 8, 128 n, N= 100, 10 feature=torch.rand(batch_size, n, d) belong=torch.randint(0, N, (batch_size,n)) output=torch.zeros(batch_size, N, d) for i in range(batch_size): for j in range(N): output[i][j]=feature[i][belong[i]==j].me...
KFrank
Hi ljmzlh! The core issue is that feature[i][belong[i]==j] has a different shape for different values of i and j. Therefore if you try to build a “tensor” whose “slices” are given by the above expression, you will end up with a “ragged tensor” (that is, a tensor whose slices have differing sh…
snau
I am trying to learn a parameter which is one element of a bigger matrix. The loss function does not directly use the learnable parameter, but a matrix having this parameter.Simplified code snippet is given below to reproduce the errorimport torch as t import torch from matplotlib import pyplot as plt param = t.tensor...
KFrank
Hi Rishi! The problem is that you have to rebuild the “computation graph” before you call .backward() for a second time. Both indexing into a and setting that “value” to param count as part of the computation graph. Consider: >>> import torch >>> print (torch.__version__) 1.9.0 >>> >>> param…
hadaev8
For example, I have models net0 and net1 working like thisout0 = net0(inp0) out1 = net0(net1(net0(inp1))) loss = criterion(out0, out1, target)I want to reverse the gradient sign from the output of model net1 but calculate net0(inp0) as usual.In simple case, I would doout0 = net0(inp0) out1 = net0(net1(inp1)) loss = cri...
KFrank
Hi Had! As an alternative to using a hook, you could write a customFunctionwhose forward() simply passes through the tensor(s) unchanged, but whose backward() flips the sign of the gradient(s). You would then insert it at the desired place in your network, e.g.: out1 = net0 (GradientReversal…
Aesteban
Given a 2d matrix of size (2000x1000) i need to compute the outer product of each row with itself. Finally all the outer products must be averaged. What is the fastest/most efficient way of doing so?This is by far the biggest bottle neck in my program. Ive come up with 2 solutions. The 2. of which, against all intuitio...
KFrank
Hi Aesteban! The calculation you are asking for can be performed with a single matrix multiplication: >>> import torch >>> torch.__version__ '1.9.0' >>> _ = torch.manual_seed (2021) >>> def avg_matrix_outer_products_v2(a): ... x_dim = a.shape[0] ... ourter_products = torch.outer(a[0], …
mrityu
Hey.Generally, when we fine-tune a classifier by keeping a pre-trained model as a feature-extractor only, we set therequires_grad = Falsefor the pre-trained block and only train the newly added FC layer.For eg., see the code snippet below:-# Setting up the model # Note that the parameters of imported models are set to ...
KFrank
Hi Mrityunjay! Yes, you are correct. When you call optimizer_ft.step(), the optimizer will only update the weights that were specified in that optimizer. Even, if other weights – not in the optimizer – have had their gradients computed, those other weights will not be updated. Yes. Just to…
James_Lee
Hi there. Is there a way for me to calculate the BCE loss for different areas of a batch with different weights? Seemed that the *weight(Tensor,optional) – a manual rescaling weight if provided it’s repeated to match input tensor shape fortorch.nn.functional.binary_cross_entropy_with_logits — PyTorch 1.9.1 documentatio...
KFrank
Hi James! If I understand your question correctly, you can just use weights of 0.0 and 1.0 to do this. Let’s say that the input to your model is a batch of images of shape [nBatch, height, width], and that the output of your model (which will be the input to BCEWithLogitsLoss) and the target y…
UmarMubeen
here is my code and errorResidual blockdef Residual_block(in_channel,out_channel):res_layer = [nn.Conv2d(in_channel, out_channel, kernel_size=3, padding=1, stride=1 ),nn.BatchNorm2d(out_channel), nn.ReLU(inplace= True), nn.Conv2d(out_channel, out_channel, kernel_size=3, padding=1, stri...
KFrank
Hi Umar! The problem is that nn.ReLU(out) constructs a ReLU object so that you then pass that ReLU object to conv_3. You want either: out = nn.ReLU() (out) # or out = nn.functional.relu (out) (I’m a little surprised that nn.ReLU(out) itself doesn’t throw an error because in most cases your ou…
Haziq
I am training a model for classification where each ground truth has some uncertainty and is thus a vector of probability scores e.g.[0.1, 0, 0.7, 0, 0.2, 0, 0]instead of[0,0,1,0,0,0,0]. The cross-entropy loss in PyTorch however, accepts only an integer target so I was hoping if someone could recommend a solution or an...
KFrank
Hi Haziq! (As an aside, your title for this thread, “Cross entropy with logit targets” should probably be “Cross entropy with probability targets,” since you ask about “probability scores” and use probabilities in the example you give.) Your ground-truth target probabilities are what are somet…
Usman_Mahmood
Hello, I have a tensor of shape N * X * Y where N is the number of subjects. I want to average the upper and lower triangle for each matrix i in N. For example if subject 1 has the matrix:tensor([[1, 2, 3],[4, 5, 6],[7, 8, 9]])Then the output should be:tensor([[1, 3, 5],[3, 5, 7],[5, 7, 9]])So, adding upper and lower ...
KFrank
Hi Usman! Probably the most direct approach is to average the matrix with its transpose. Here is an illustration for the 3-dimensional-tensor use case you mention: >>> torch.__version__ '1.7.1' >>> t = torch.tensor ([[[1, 2, 3], [4, 5, 6], [7, 8, 9]], [[10, 20, 30], [40, 50, 60], [70, 80, 90]]…
Alex07
I have a tensor likea = torch.Tensor([3,2,7,9,2,9, 9])However, I need the values to start from 0 and go ton-1wherenis the number of unique elements ina. One possible way to compute this seems to be:r = torch.zeros_like(a) for i,el in enumerate(torch.unique(a)): r[a == el] = igivingtensor([1., 0., 2., 3., 0., 3., 3....
KFrank
Hi Alex! There’s no way you can avoid unique() (or its equivalent). That’s an inherent bit of processing required by your problem. You can avoid the loop – depending on the details of your use case, such as having the values in a being non-negative integers that are not ridiculously large – b…
James_Lee
Hi there. I found this statement (BCELoss — PyTorch 1.6.0 documentation. ) in PyTorch doc. Does this mean that I could use thetorch.nn.BCELossfor soft labels without transforming the soft labels into hard labels with a threshold when calling BCE loss? Thanks.
KFrank
Hi James! Neither BCEWithLogitsLoss (which you should use), nor BCELoss (which you should not use) has a built-in thresholding feature. But it’s easy enough to threshold target (the labels) before you pass it in: my_threshold = 0.65 thresholded_target = (target > my_threshold).float() loss = t…
jhp
Hi I have a quick question about kl divergence loss in PytorchIs it okay to use sigmoid instead of softmax for input? Most of the case I noticed that softmax probability distribution is used, but I wonder does that make sense to use kl divergence loss for multi-label target.I want to use KL divergence loss by giving mo...
KFrank
Hi Jhp! Not really. KLDivLoss doesn’t care what any of the dimesions – including the batch dimension – are. It simply performs an element-wise computation and takes the mean (but see itsdocumentationfor how its reduction = 'batchmean' and deprecated reduce and size_average constructor ar…
Shar
Hi,I’m trying to use the function:loss = criterion(outputs, labels)where the inputs are:labels=tensor([[2]])outputs=tensor([[274.3095, 842.6060, -52.4284]], grad_fn=<AddmmBackward>)but I get the error:multi-target not supported at /Users/distiller/project/conda/conda-bld/pytorch_1570710797334/work/aten/src/THNN/generic...
KFrank
Hi Shar! Your labels has one too many dimensions. It appears that you are using a cross–entropy-type loss criterion. Pytorch’s CrossEntropyLoss (and related loss criteria) expect your outputs to have shape [nBatch, nClass] and your labels to have shape [nBatch] (with no nClass dimension). I…
melike
Hi all, I have a tensor with integers in the range [0, N-1] and I need to create N masks, one for each integer value, i.e. in the 0’s mask all values of 0 should be True and the rest False, and so on. So far I could come up with a for loop on the range, but that drastically increases the runtime. I’d be grateful if som...
KFrank
Hi Melike! Try: >>> import torch >>> torch.__version__ '1.9.0' >>> _ = torch.manual_seed (2021) >>> N = 10 >>> t = torch.randint (N, (2, 3, 5)) >>> i = torch.ones ([N] + list (t.shape)).cumsum (0) - 1 >>> masks = t == i >>> t tensor([[[4, 9, 3, 2, 5], [6, 2, 0, 5, 0], [1, 9, 7, …
azer
Hi all,I am training resnet18 for image classification and I got good results but accuracy of the validation dataset is higher than the training dataset.| train loss: 0.0059 | train acc 0.9621 | val loss: 0.007485 | val_acc: 0.9771For the train dataset I used the weighted cross entropy as loss function since my data...
KFrank
Hi Azeai! There are a number of possible reasons for this: (First, a quick note: One would normally expect your model to perform at least a little better on the training data.) The character of your training and validation datasets could be different. For example, maybe your validation image…
Marek_Balaz
Hello, I would like to multiply tensor t1 = torch.tensor([1, 2, 3, 4, 5]) and t2 = torch.tensor([1, 2, 1]) in order to get result similar like from 1D kernel application: [1 * 1 + 2 * 2 + 3 * 1, 2 * 1 + 3 * 2 + 4 * 1, 3 * 1 + 4 * 2 + 5 * 1]. Can you give me advice, how to compute this case without any for loop?Thank yo...
KFrank
Hi Marek! Useunfold()to generate the correct “slices” of t1: >>> import torch >>> torch.__version__ '1.9.0' >>> t1 = torch.tensor([1, 2, 3, 4, 5]) >>> t2 = torch.tensor([1, 2, 1]) >>> t1.unfold (0, 3, 1) @ t2 tensor([ 8, 12, 16]) You could also useconv1d()if you unsqueeze() t1 and t2 to add …
ivo-1
Hello,I am using the ResNet architecture as a feature extractor for an OCR task. While debugging I noticed that the output values of ResNet are very similar for any input image out of my dataset. So I decided to look at the output values when an untrained ResNet is given random input.resnet = torchvision.models.resnet1...
KFrank
Hi Ivo! I can reproduce what you see. I don’t see anything wrong with this result (although I don’t have an expectation one way or the other whether it should work out this way). My intuition – that could be wrong – suggests that pumping a random input through a randomly-initialized Res…
talipturkmen
Hello,Firstly I’m very sorry with this newbie question. I couldn’t find an efficient answer. I want to create a mask with zeros and ones for the n by n matrix. For value each i-th row I want to set the columns between i-a and i+a to 1 and rest of them to 0.As an example of the valuesn = 4 (4x4 matrix)a = 1I want to get...
KFrank
Hi Talip! I think thattorch.tril()and/ortorch.triu()will be the key to what you want. Here is one specific approach: >>> torch.__version__ '1.6.0' >>> n = 4 >>> a = 1 >>> torch.tril (torch.ones ((n, n)), a) * torch.triu (torch.ones ((n, n)), -a) tensor([[1., 1., 0., 0.], [1., 1., 1.,…
omer_as
Hey all,I have a tensor t with shape (b,c,n,m) where b is the batch size, c is the number of channels, n is the sequence length (number of tokens) and m a number of parallel representations of the data (similar to the different heads in the transformer).I want to perform a 1D conv over the channels and sequence length,...
KFrank
Hi Omer! You can .reshape() your input tensor and use the groups constructor argument of Conv1d: >>> import torch >>> torch.__version__ '1.9.0' >>> _ = torch.manual_seed (2021) >>> b = 2 >>> c = 3 >>> n = 5 >>> m = 2 >>> t = torch.rand ([b, c, n, m]) >>> u = t.permute (0, 3, 1, 2).reshape (b, m …
Aditya_Prakash
In my neural network (RNN), I am defining the loss function such that the output of the neural network is used to find the index (binary) and then the index is used to extract the required element from an array which in turn will be used to calculate MSELoss.However, the program givesparameter().grad = Noneerror which ...
KFrank
Hi Aditya! I haven’t looked at your code or error messages in any detail. But, yes, your “graph is breaking.” code, as a function of output, is not (usefully) differentiable, so, regardless of what you subsequently do with code, you won’t be able to back propagate through it. I have no idea …
Ada
My network threw an error during backprop because layer shapes did not match, and I wondered how/why the network was able to do the forward pass without throwing an error. After investigating, I’ve discovered some weird behavior.If the model is on the CPU, all is well:model = nn.Linear(100,1) x = torch.randn(1,200) mo...
KFrank
Hi Ada! This appears to be a known bug that has recently been fixed. See this thread: Best. K. Frank
xianqian
If we have a tensor like this[[True, False, False,False], [False,False,False,False], [True,True,False,False], [[True,True,True,True]]The check should returnTrue. But in case there is something like this[[True, False, True,True], <--- wrong [False,False,False,False], [True,True,False,False], [[True,True,True,True]]shoul...
KFrank
Hi Xianqian! cumprod() – for boolean values – will test for “decreasing values”: >>> import torch >>> torch.__version__ '1.9.0' >>> >>> b1 = torch.tensor ([ ... [True, False, False, False], ... [False, False, False, False], ... [True, True, False, False], ... [True, True, Tru…
miturian
I apologize if this is adequately explained in the documention. If so, please just give me a link…I have only foundDoes PyTorch average or sum gradients over a minibatch?autogradFor example when you retrieve the gradients like so: loss = F.nll_loss(output, target) loss.backward() for key, value in model.named_paramete...
KFrank
Hi Miturian! I think you’re trying to say something slightly different, so let me rephrase it: x is a batchSize-2 “input” to the “model”; w is the (scalar) weight that defines this very simple “model”. Yes. Yes, but I think it’s more helpful to think of this as a “tensor” chain rule, rathe…
andreys42
HelloIm wonder why NLLLoss doesn’t change when using [2,2] weights in example below?There is really incomplete documentation related to details of calculations.log_prob = torch.tensor([[0.1, 0.5], [0.8, 0.8], [0.8, 0.1]]) target = torch.tensor([0, 0, 1]) weight = torc...
KFrank
Hi Андрей! You have two mistakes in your calculation: First, NLLLoss does not take the log() of its input, but you do in your manual calculation. Second, when you manually calculate the weighted average, you should be dividing by 3 * 2 (not (3 * 4)). Thus: >>> import math >>> import torch >>>…
matthewleigh
Hey everyone,I have a minor issue and I am not sure if it is a bug or I am simply not understanding something. But I noticed that in my project I was getting some strange results for certain configurations and error messages not popping up when expected.To boil it down, I am allowed to multiply matrices of incompatible...
KFrank
Hi@ptrblck! I can confirm that the issue has been fixed in the current nightly (that is, the cuda version also throws the shape-mismatch error): >>> torch.__version__ '1.10.0.dev20210830' >>> torch.nn.Linear (3, 3).cuda() (torch.randn (5, 10).cuda()) Traceback (most recent call last): File "<s…
thomas
Hello everyone!I have an input of this shape:(num_samples, num_symbols, num_features, num_observations)I would like to feed through this input through a neural network.The neural network as(num_features, num_observations)shape input and(num_outputs)outputs, giving me(num_samples, num_symbols, num_outputs)when I apply a...
KFrank
Hi Thomas! I do not know of any functionality built into pytorch similar to your apply_along_axis(). And even if there were, it would still impose a performance penalty, as it would still be breaking up what might have been a single, larger tensor operation into many smaller, axis-wise tensor…
Sitaraman
Hi,Here is my code, on training the logits dont seem to change at all. What could be the reason?from dgl.nn import GraphConv, SAGEConvclass GCN(nn.Module):definit(self, in_feats, h_feats, num_classes):super(GCN, self).init()#self.conv1 = GraphConv(in_feats, h_feats, allow_zero_in_degree=True)#self.conv2 = GraphConv(h_...
KFrank
Hi Vilayannur! I haven’t looked at your code in any detail, however: “breaks the computation graph” because Variable(loss, requires_grad = True) creates a new tensor (that also happens to be called loss), and (by design) gradients backpropagated to the new loss don’t get further backpropagate…
julliet
I have a binary classification problem. Right now I’m using several linear layers with ReLU activation.I’m using BCEWithLogitsLoss() for Loss, so I’m not implementing any Softmax on the layers.Predictions of the model look something like this:-0.2443, 6.6122, 25.0909, ..., -62.7383, 0.3066, 61.6255So I can’t rea...
KFrank
Hi Julia! In terms of comparing your predictions to the “zeroes and ones” of your label, BCEWithLogitsLoss does precisely this (without converting your predictions into 1s and 0s). BCEWithLogitsLoss takes predictions that are raw-score logits (such as those produced by your final Linear layer …
abhisek
Trying to implement Tversky metric as follows:import torch def tversky_score( predictions, targets, alpha=0.5, beta=0.5, eps=1e-16, encode_target=True ): """ Ref A: https://arxiv.org/abs/1706.05721 alpha = beta = 0.5 : Dice coefficient alpha = beta = 1 : T...
KFrank
Hi Abhisek! Could you post a trimmed-down, runnable, fully-self-contained script that reproduces the error? (Please have the script print out the version of pytorch you’re using, torch.__version__.) Please provide hard-coded (or random) sample data (with size and dimensionality as small as po…
Chris_XU
Hi, I would like to understand the cross entropy loss for multi-dimensional input by implementing it by myself. Currently I am able to get the close result by iterating usingnp.ndindex,K = 10 X = torch.randn(32, K, 20, 30) t = torch.randint(K, (32, 20, 30)).long() w = torch.randn(K).abs() CrossEntropyLoss(weight=w, re...
KFrank
Hi Chris! You can get a tensor of weights corresponding to the class labels in t by indexing into w, and you can get a tensor of (negative) log-probabilities by calling .take_along_dim() with t as the argument: >>> import torch >>> torch.__version__ '1.9.0' >>> _ = torch.manual_seed (2021) >>> …
kabu1204
docs:torch.nn.KLDivLosswhich I think should be:图片3550×452 46.2 KB
KFrank
Hi Kabu! No, the documentation for KLDivLoss is correct. The equation you posted an image of contains the term: y_n (log y_n - x_n) You are expecting log x_n rather than just x_n. But KLDivLoss expects x_n to be already a log-probability. Quoting from the documentation you linked to: As w…
EricPengShuai
Is it impossible to install version 0.3.0 or 0.3.1 of pytorch on Windows now?
KFrank
Hi Shuai! I was able to install and run pytorch 0.3.0 on windows 10. (My motivation was to run pytorch on an old laptop gpu, and 0.3.0 was the newest version of pytorch I could find that was pre-built with support for my old gpu.) See this post from@ptrblckwhere he gives a link to some “leg…
ArshadIram
I am trying to optimize over the value z and netG is trained GAN model.Here is the code. I am getting an error. I want to min the loss between the generated images x and the z. something like this:maxEpochs = 100 z = nn.Parameter(torch.rand(64,100,1,1), requires_grad=True).to(device) opt = torch.optim.Adam([z]).to(devi...
KFrank
Hi Iram! Looking at an object’s __dict__ property is often a good way to probe all of its properties. So start with: >>> import torch >>> torch.__version__ '1.9.0' >>> opt = torch.optim.Adam ([torch.randn (3, requires_grad = True)]) >>> opt.__dict__ {'defaults': {'lr': 0.001, 'betas': (0.9, 0.9…
ph14
How does the functiontorch.randngenerate variates? What specific method does it use, and where can I find a reference (if there is one)? Also, does the method change based on CPU/GPU usage?
KFrank
Hi Piers! As near as I can tell, randn() uses theBox-Mullermethod, both on the cpu and gpu. I am not aware that the algorithm used is documented anywhere but in the code. It’s also quite opaque to me how any given pytorch python call gets dispatched down to the code that actually does the …
MrPositron
Hi all!I am stuck at one problem. Let say that that I have tensors A and B. Now, I want to update elements in A at locations given in tensor B.For example, let say there are two tensors A and B with size 3 x 6, and 3 x 2.A = [ [1, 2, 3, 4, 5, 6], [2, 3, 4, 5, 6, 1], [3, 4, 5, 6, 1, 2], ] B = [ [0, 1], ...
KFrank
Hi Nauryzbay! You may usescatter_add_()(with pytorch tensors) for this: >>> import torch >>> torch.__version__ '1.9.0' >>> A = torch.tensor ([ ... [1, 2, 3, 4, 5, 6], ... [2, 3, 4, 5, 6, 1], ... [3, 4, 5, 6, 1, 2], ... ]) >>> B = torch.tensor ([ ... [0, 1], ... [0, 3], ... …
PhysicsIsFun
Greetings,I am confused by PyTorchs usage of the momentum SGD method.https://pytorch.org/docs/stable/generated/torch.optim.SGD.html#torch.optim.SGDThe formulas state thatv_{t+1} = \mu * v_{t} + g_{t+1} p_{t+1} = p_{t} - lr * v_{t+1}The documentation remarks that this differs from the original formula used by Sutskeve...
KFrank
Hi Physics! The two formulations are equivalent. Let’s call the “velocity” in the first, pytorch, formulation vPytorch_{t} and in your second proposed version vPhysics_{t}. The two formulations only differ in a redefinition of v, namely vPhysics_{t} = -vPytorch_{t}, that drops out of the fin…
John_Doe
Hello,I am trying to implement a BCEWithLogitsLoss version with a different weight for some pixels. The goal is to have good borders in a segmentation task.The outputs of my net are logits.My weight map have the same size as my target. (For example [batch_size, height, width]).My implementation currently looks like thi...
KFrank
Hi John! I would recommend that you use BCEWithLogitsLoss’s weight argument: mean_weighted_loss = torch.nn.BCEWithLogitsLoss (weight = pix_map) (pred, target) As a further comment: You can certainly write your own version of BCEWithLogitsLoss, but if you do, you should, for reasons of nu…
Patrickens
Dear All,Im working on a simulation algorithm where the linear algebra is handled bypytorch. One step in the algorithm is to do a 1d convolution of two vectors. This needs to happen many times and so it needs to be fast. I decided to try to speed things further by allowing batch processing of input. This means that I s...
KFrank
Hi Tomek! Package your nbatch dimension as “channels” (in_features) and then use conv1d()'s groups feature so that you apply each weight vector in your batch of weight vectors separately to the corresponding “channel” of your input. Here is your code with the “groups” version added on at the e…
wasabi
I have a tensor X of shape (a, b, c) and a matrix of permutation (not a permutation matrix) P of shape (a,b), where each row of P is an output of torch.randperm(). I want to shuffle X in the following way:for i in range(a): Y[i] = X[i][P[i]] return Ywhat’s the best way to achieve this? Thanks.
KFrank
Hi Wasabi! Use gather(): >>> import torch >>> torch.__version__ '1.9.0' >>> _ = torch.manual_seed (2021) >>> a = 2 >>> b = 3 >>> c = 5 >>> X = torch.arange (a * b * c).reshape (a, b, c) >>> X tensor([[[ 0, 1, 2, 3, 4], [ 5, 6, 7, 8, 9], [10, 11, 12, 13, 14]], [[…
Magjo
Hi,I have difficulties to understand why this code does not throw errormodel = torch.nn.Linear(512,2) opt = torch.optim.SGD(model.parameters(), lr=0.1) inputs = torch.rand((32,512)) outputs = torch.rand((32,2)) opt.zero_grad() pred = model(inputs) pred[:,1] = torch.sigmoid(pred[:,1]) loss = torch.nn.functional.mse_l...
KFrank
Hi Magjo! I’m guessing … (But it’s the weekend, so let me guess away!) At issue is not only whether you are using an inplace operation, but also whether the tensor modified by the inplace operation is used directly in the backward() computation. The derivative of relu() depends on the sign o…
Yongjie_Shi
I have two networks, for example, netA and netB. However, I want to use different loss functions to update different networks. For example, I want to use loss1 to update netA, and use loss2 to update netB.I have defined different optimizers for netA and netB. Each optimizer optimizes only the parameters of the correspo...
KFrank
Hi Yongjie! Your process it mostly correct. You need to use retain_graph = True in your first backward() call: loss1.backward (retain_graph = True) No, one forward pass is enough. (You do, of course, have to calculate both loss1 and loss2.) But you do need two backward() passes. (If you w…
joec
Hi, I am seeing an issue on the backward pass when using torch.linalg.eigh on a hermitian matrix with repeated eigenvalues. I was wondering if there is any way to obtain the eigenvector associated with the minimum eigenvalue without the gradients in the backward pass going to nan. I am performing this calculation as a ...
KFrank
Hi Joe! The short answer: No, there is not. The issue is that in the presence of repeated (degenerate) eigenvalues, “the eigenvector associated with” a degenerate eigenvalue is not well defined. An eigendecomposition algorithm (such as torch.linalg.eigh()) will give you a list of eigenvalues…
Samuel_Bachorik
Hi this is my code for solving quadratic equation. This is working fine but problem is when I try to add one more unknown.x = torch.nn.Parameter(torch.zeros(1), requires_grad=True) def model(x): b=3 a=2 c=1 global y y = a*x ** 2 + b*x + c return y optimizer = torch.optim.Adam([x], lr=0.1) for...
KFrank
Hi Samuel! Consider the following: >>> import torch >>> torch.__version__ '1.9.0' >>> >>> # your scalar x = z[0] and y = z[1] >>> z = torch.nn.Parameter (torch.tensor ([3.0, 7.0])) # just to make it more interesting than 0.0 >>> equationCoefficients = torch.tensor ([[1.0, 1.0], [1.0, -1.0]]) >>>…
tianle-BigRice
I needed to optimize my own loss using the Optimizer, but I ran into this problemTraceback (most recent call last): File "train.py", line 138, in <module> fire.Fire() File "D:\Anaconda3\lib\site-packages\fire\core.py", line 141, in Fire component_trace = _Fire(component, args, parsed_flag_args, context, nam...
KFrank
Hi Tianle! What do you think this error message might be telling you? If you were debugging somebody else’s code, what issues might this error message suggest you take a look at? This entry in your params list is a dictionary that contains a Parameter. Does this make sense? But this entry i…
lavish619
I am using a pretrained resnet101 and I want to change the dilation rates and stride of some conv layers.If I initialize the layers again, that will change the weights of that layer, but incase of stride or dilation rate change only, the weights should not get changed because the kernel size is same.So how can I change...
KFrank
Hi Lavish! If it were me, I would probably do: for i in range(23): saveWeight = self.resnet.layer3[i].conv2 self.resnet.layer3[i].conv2 = nn.Conv2d (256, 256, kernel_size=3, stride=1, …
Alef93xx
Hey guys!I’m trying to train a neural network for multiclass semantic segmentation (FCN) using the CrossEntropyLoss function, but the following error occurs:Preds: torch.Size([1, 21, 160, 240])Target: torch.Size([1, 3, 160, 240])The size of tensor a (21) must match the size of tensor b (3) at non-singleton dimension 1A...
KFrank
Hi Alisson! Let me make a number of comments before outlining how one might convert RGB colors to integer class labels, First, just to use a clear-cut word, let me call your ground-truth, annotated (labeled) images “masks.” So your dataset consists of a bunch of images that will be input to y…
_joker
Hi All,I have a very simple use case where I would like to predict a floating number (y), given four floating numbers (x1, x2, x3, x4) that denotes the property of a process. I would like to have some pointers or insights on different supervised and unsupervised approaches to solve this problem and which one among the ...
KFrank
Hi Joker! First, especially because your xs differ in scale by several orders of magnitude, you should most likely normalize your input data to be or order one. This makes it easier for your training to get started, and your model doesn’t have to “learn” to deal with these very different scal…
Bahman_Rouhani
I want to apply a mask to my model’s output and then use the masked output to calculate a loss and update my model.I don’t want the autograd to consider the masking operation when calculating the gradients, i.e. I want the autograd to treat my model as if it had outputed the masked version of my input. In other words, ...
KFrank
Hi Bahman! Your code should do what you want. Autograd tracks computations with pytorch tensors and doesn’t care whether those computations are “free standing” (such as your mask operation) or part of your model or part of your loss – they’re all treated the same. If mask has 0s in it, then w…
Samue1
I have a tensorTwith dimensions(batch_size, size_n, size_n)T = [[[0.9527, 0.8821], [0.8442, 0.6147]], [[0.0672, 0.5737], [0.1963, 0.4532]], [[0.0992, 0.3838], [0.4169, 0.0925]]]and want to extract the diagonal of each matrix in that batch to getdiag_T = [[0.9527, 0.6147], [0.0...
KFrank
Hi Samuel!torch.diagonal()does what you want: >>> import torch >>> torch.__version__ '1.9.0' >>> T = torch.tensor ( ... [[[0.9527, 0.8821], ... [0.8442, 0.6147]], ... [[0.0672, 0.5737], ... [0.1963, 0.4532]], ... [[0.0992, 0.3838], ... [0.4169, 0.0925]]] ... ) >>> torch.…
SebGruber1996
Hi!When I define the following functions ‘n_deriv_’ (calling ‘jac’ recursively) and ‘n_deriv__’ (calling ‘grad’ recursively in a for loop), they work without issues:import torch from torch.autograd.functional import jacobian as jac from torch.autograd.functional import hessian as hes from torch.autograd import grad de...
KFrank
Hi Sebastian! Your error is caused by python scoping rules and is not specific to pytorch nor torch.autograd.functional.jacobian. Please see this python FAQ:Why do lambdas defined in a loop with different values all return the same result?Here is a tweaked version of your code that shows how…
avitase
Hi:)I want to calculate the inverse FFT of a truncated input vector,Python 3.8.10 (default, May 19 2021, 18:05:58) [GCC 7.3.0] :: Anaconda, Inc. on linux Type "help", "copyright", "credits" or "license" for more information. >>> import torch >>> x = torch.rand(5) >>> fx = torch.fft.fft(x) >>> fx tensor([3.2079+0.0000j,...
KFrank
Hi Avitase! Pytorch’s fft functions, including ifft(), pad by simply appending zeros to the end of the input vector – they don’t insert zeros into the middle. You can get the behavior I believe you want by using rfft() / irfft(). The following illustrates these two points: >>> import torch >>> …
Vignesh_Baskaran
I am trying to understand the discrepancy that happens while performing matrix multiplication in batch.To summarize I am trying to do the following:matmul(matrix, tensor) → Slice the output → sliced outputslice the input tensor → matmul(matrix, sliced_tensor) → sliced outputI expect the results to match exactly but th...
KFrank
Hi Vignesh! Yes, such a discrepancy is to be expected due to floating-point round-off error. In short, when you use floating-point arithmetic to perform operations in different orders that should be mathematically equivalent, you, in general, expect to get (slightly) different results. In p…
Samue1
Given a one-dimensional input tensorS, I want to evaluate the following expression:J_ij = S_i(delta_ij - S_j)wheredelta_ijrepresents the Kronecker delta. The resultJof this expression is a square matrix. I would like to know how I can efficiently evaluate this expression using PyTorch? My current implementation is very...
KFrank
Hi Samuel! No. If you had tried it, you would have discovered that torch.outer() does not accept multidimensional tensors. No, this will throw an error (because you pass a multidimensional tensor to torch.outer()). You can, however, use pytorch’s swiss army knife of tensor multiplication fu…
lkp411
Is there a way to efficiently, in a vectorized manner, compute the combinations along each rows of a 2D tensor individually?For example:a = torch.tensor([[2, 5, 6], [7, 9, 4]]) result = torch.combinations(a, 2, dim=1)results should look liketorch.Tensor([[[2, 5], [2, 6], [...
KFrank
Hi lkp411! Yes, the idea would be to construct the combinations of a 1D vector of indices and then index into your multidimensional tensor, swapping the dimensions around appropriately. Like this: >>> import torch >>> torch.__version__ '1.9.0' >>> a = torch.tensor([[2, 5, 6], [7, 9, 4]]) >>> a…
Samue1
I have a batchx = torch.rand(size=(M, N))and want to create for each of theMinputs a diagonal matrix with dimensionsN x Nsuch that the output has dimensionsM x N x N. How can I do that? If I passxtotorch.diagI get a one-dimensional output.Any idea what I do wrong?
KFrank
Hi Samuel! Try: torch.diag_embed (torch.rand (size = (M, N))) Best. K. Frank
WilliManFR
Hello,I have a problem. I have done a CNN which gives me an output of the form : (BATCH_SIZE, 1, HEIGHT, WIDTH). However, in order to apply the CrossEntropyLoss function, I need an output of the form, (BATCH_SIZE, NB_CLASS, HEIGHT, WIDTH). How can I get such a matrix from my original output? I currently have three clas...
KFrank
Hi William! As an aside, it sounds like your current CNN performs binary (i.e., background-foreground) segmentation. It sounds like your desired use case is multi-class (with NB_CLASS classes) semantic segmentation. First, once you’re down to (BATCH_SIZE, 1, HEIGHT, WIDTH), that is, a singl…
masteryoda
I have atensor with multinomial probabilities. Each vector (in my example below is 1x3) inside the tensor represent multinomial probability distribution for 1 random variable:I want to sample from the multinomial probability distributiontensor of random variables.For example, this is the tensor (MxNxIxJx3) with the mul...
KFrank
Hi Yoda! As I understand your question, you are asking for the categorical distribution (which is a special case of the multinomial distribution). torch.distributions.Categorical take batches of probs (as does torch.distributions.Multinomial), so you can just pass your probs in as-is. Here is…
drbeethoven
Hello,I am trying to create a method that allows me to create a nn of variable number of layers and size. As such, I am using a module list.However, I notice that when I used “nn.Linear” before using a module list, I would have to specify Sigmoid in between layers and soft max at the end. The soft max I can just put at...
KFrank
Hi James! I’m not entirely sure what you’re asking or how you intend to use your ModuleList, but note that a torch.nn.ReLU is a Module so you can include it in your ModuleList, in between, for example, some torch.nn.Linears. Best. K. Frank
metro
Hi there,I’m looking for a loss function which will motivate a network to output on the last layer either near 0 or near 1 values. I’m imagining a function that looks similar to this. The output layer is of type float/double and will be used as a mask.Background:I’m training an unsupervised network to predict masks. My...
KFrank
Hi Metro! I don’t understand your use case, but, in general, you can certainly penalize outputs that are not close to 0 or 1. Let x be a single output. You could then try things like: penalty_0_1 = (x * (1.0 - x))**2 penalty_0_1 = torch.abs (x * (1.0 - x)) penalty_0_1 = ((x * (1.0 - x))**2)**a…
NPC
Hey,i am sorry if i missed a similar topic.I want to skip a weight in a convolution layer. (I know bad description).Lets say i have a kernel like [a, b, c].But i want a == c (or more precisely I want the optim. to threat it as one parameter).Is there a proper way to archive this?I am currently just setting them to (a ...
KFrank
Hi NPC! To me the conceptually cleanest approach would be to build your constrained kernel from a separate two-element Parameter that you actually optimize. Something like: import torch from itertools import chain # initialize a = 1.1, b = 2.2 k = torch.nn.Parameter (torch.tensor ([1.1, 2.2]…
jhuteau
I want an efficient (vectorized) implementation instead of my current for-loop implementation.I have a MxN matrix named Sim corresponding to the similarity scores of M anchors with N documents.I have multiple negative (not matching the anchors) documents as well as multiple positive (matching) documents for each anchor...
KFrank
Hi Jhuteau! Easy peasy: import math import torch print (torch.__version__) torch.manual_seed (20212021) # code with tensor operations def hardest_triplet_margin_lossB (preds, targets, margin, hardest_fraction): # preds: the similarity scores, of shape [num_anchors, num-docum…
phantom90
Hi there,Just curios what will happen if the image on the disk changes when we run DataLoader to traverse the dataset. Will the loaded data be the latest file on the disk or otherwise.Thanks!
KFrank
Hi Phantom! What will happen is whatever your code decides to do. A torch.utils.data.DataLoader is a wrapper that lets you iterate over a Dataset. Per the documentation, aDatasetis an abstract class that you have to implement concretely. In particular, you have to implement its __getitem__…
lv_Poellhuber
So I have this network that has the following architecture:image1960×1464 527 KBThe goal of this architecture is to extract features out of a sequence of 30 frames, without knowing what those features are. The 30 frames are represented as images, but they aren’t real images, which is why we can’t have a classification ...
KFrank
Hi Louis-Vincent! I haven’t looked at your code, but yes, in general the kind of thing you propose doing should work just fine. Make sure you understand the basics of how pytorch uses the require_grad property of a tensor. Normally the weights, etc., of your model, and the tensors that depend…
Shawn_Zhuang
Hi,in pytorch lightning, recently I found that if I use F.Dropout in forward step, even when I set mode to model.val, the dropout still work, then I realize I should replace it with nn.Dropout as module attribute. After that, everything performs normally.Then my concern would be like, if I use F.relu in forward instead...
KFrank
Hi Shawn! F.relu() (which is to say torch.nn.functional.relu()) is a function. nn.ReLU (torch.nn.ReLU) is a class that simply calls F.relu(). These two ways of packaging the function do the same thing, including when calling .backward(). There is no need to instantiate two instances of the n…
Rohit_R
Say I create a placeholder for a batch of 3 images -batch = torch.zeros(3, 3, 256, 256, dtype=torch.uint8)I have my dummy image -image = torch.randint(size = (3,256,256), low=0, high=256)I then do -batch[0] = imageI am unable to understand the following outputs -id(batch[0]) == id(image) out: FalseShould this...
KFrank
Hi Rohit! Batch is a single “holistic” tensor of shape [3, 3, 256, 256]. It is not a “collection” (in the sense of, say, a python list or dictionary) of three 3x256x256 tensors. Pytorch is doing some moderately fancy stuff with python here. This is not simply assigning image to the 0 element…
LearnedLately
I want to sample a tensor of probability distributions with shape (N, C, H, W), where dimension 1 (size C) contains normalized probability distributions with ‘C’ possibilities. Is there a way to efficiently sample all the distributions in the tensor in parallel? I just need to sample each distribution once, so the resu...
KFrank
Hi Learned! Yes, you can use torch.distributions.Categorical, provided you adjust your distributions tensor so that its last dimension is the distribution dimension. Here is an example script: import torch print (torch.__version__) _ = torch.random.manual_seed (2021) N = 2 C = 3 H = 5 W = 7 …
Harimus
Hello, so I think showing the problematic behavior I’ve encountered first is best:Screenshot from 2021-05-19 19-01-111051×386 24 KBSo I’ve been training Policy-gradient (Reinforcement Learning) methods withTanhDistributions. Tanh distributions built-inlog_prob()method is very sensitive to the numeric value close to 1 &...
KFrank
Hi Dan! This is a documentation bug. Quoting from pytorch’s1.8.1 documentation: eps float The smallest representable number such that 1.0 + eps != 1.0. This is sort of on the right track, but really isn’t correct.Numpy’s (1.20) documentationgets it right: eps : fl…
Abhilash
I’m a beginner in Computer Vision, Pytorch and working on the multi class classification problem ( MNIST), how do I adjust the penalty for misclassifying, for example, a 5 as a 6, because a human would do similar mistake as they look pretty close ( for a terrible image), or 1 as 7 in few images. The weight parameter in...
KFrank
Hi Abhilash! The most straightforward approach would be to use cross entropy as your loss function, but to use probabilistic ("soft’) labels, rather than categorical (“hard”) labels. Pytorch does not have a soft-label version of cross-entropy loss built in, but it is easy to implement one. Se…
ZimoNitrome
Given a tensor of values in the range [0, 1], multiplying these values with a scalarpand applying a softmax gives scaled probabilities that sum to 1. Increasingppushes the values to either 0 or 1. See example:values = torch.tensor([0.00, 0.25, 0.50, 0.75, 1.00]) softmax_values = torch.stack([torch.softmax(p*values, 0)...
KFrank
Hi Zimo! There are lots of possibilities. (You could choose between them by imposing further conditions on the behavior of your function.) Here is an illustration of one possible choice: >>> import torch >>> torch.__version__ '1.7.1' >>> _ = torch.manual_seed (2021) >>> def gradual (values, p)…
krish13
This topic seems to be widely discussed but I got a question. The document in the Pytorch indicate following formula:image1374×402 47.4 KBMy question is related to N.Let’s say my target size is [1* 1 * 10 * 10] - B,C,H,WThe output from my network is also [1* 1 * 10 * 10] - B,C,H,WIn this case N = 1. Thus, according to...
KFrank
Hi Krish! As an aside, you should use BCEWithLogitsLoss rather than BCELoss for reasons of numerical stability. To answer your question, the documentation is a little imperfect on this point. BCELoss simply computes the loss for each pair of matching elements of the output and target tensors,…
AerysS
I have a torch tensor of shape(batch_size, N). I want to apply functional softmax with dim 1 to this tensor, but I also want it to ignore zeros in the tensor and only apply it to non-zero values (the non-zeros in the tensor are positive numbers). I think what I am looking for is the sparse softmax.I came up with this c...
KFrank
Hi Aerys! I didn’t look at the github code you linked to, but, in general, if you have a (properly-written) torch.nn.Module or torch.autograd.Function, you can instantiate it “on the fly” and call it, rather than using it as a “layer” in a model. Thus, for example: probs = torch.nn.Softmax (d…
weiguowilliam
I just started to use PyTorch and I plan to build an ensemble of networks. I implemented it with a list. But I got the following error:I’m confused by the error. The network has just 2 fc layers. Could you please help me explain that? Thanks in advance.Traceback (most recent call last): File "test_list.py", line 107,...
KFrank
Hi Wei! Your list_of_loss keeps growing with every iteration of the loops over num_epochs and train_loader. Therefore you keep calling list_of_loss[0].backward() on the same loss over and over again that was created for the first sample in your first epoch. Put another way, the newly appended…
AnhMinhTran
Hi guys, just wondering if it is possible to combine two different model training on different dataset into one?To illustrate my goalInkedannotation_LI3840×2880 489 KBFor model A: I have A.pthFor model B(red colour): I have B.pthIs it possible to combine them into one since the data is too skewed
KFrank
Hi Ahn! You won’t be able to combine the two models together in some naive way. The problem is that while model A has learned to distinguish a “Worm” from a “Brittlestar” and model B has learned to distinguish a “Prawn” from a “Seaspider,” neither model has learned to distinguish a “Worm” fro…
stevethesteve
At some point during computation, my model needs to compute the logarithm of softplus of a parameter.Currently, I implement this via:torch.nn.functional.softplus(theta).log()Due to the log() call I fear that there might be issues with computational stability.Is there a computationally more stable way of computing log-s...
KFrank
Hi Steve! I haven’t worked through this in detail, but a first look suggests that the softplus() fix in the recent nightlies addresses your issue. Running the script you posted on today’s nightly, 1.9.0.dev20210504, seems to show that the issue is gone: 1.9.0.dev20210504 diffA.abs().max() = te…
jpj
I have a fully connected neural network with 78 input features and two hidden layers with 50 nodes each. Output with 7 nodes.self.hidden_layer_1 = torch.nn.Linear(78,50) self.hidden_layer_2 = torch.nn.Linear(50,50) self.output = torch.nn.Linear(50, 7)def forward(self, input): weights = np.load('./weights.npy'...
KFrank
Hi jpj! This is fine. At this point self.hidden_layer_1.weight is a tensor with shape [50, 78]. (The order of the dimensions is correct because of how pytorch multiplies batch tensors with weight tensors.) This is your problem, You are replacing hidden_layer_1.weight, whose shape is consis…
akvilonBrown
Greetings!I’d like to ask for a suggestion from this venerable community on how I can handle a loss for a mulit-task network?The data goes forward in parallel in an Unet-like network with two heads performing semantic segmentation with different targets (not a multi-class configuration, the tasks differ).I was unable t...
KFrank
Hi Iaroslav! Yes, it is perfectly reasonable to add the two losses together, and there is no need to write a custom loss function. So, something like: loss_task_1 = torch.nn.CrossEntropyLoss() (pred_head_1, target_task_1) loss_task_2 = torch.nn.CrossEntropyLoss() (pred_head_2, target_task_2) to…