I intend to implement an LSTM with 2 layers and 256 cells in each layer. I am trying to understand the PyTorch LSTM framework for the same. The variables in torch.nn.LSTM that I can edit are input_size, hidden_size, num_layers, bias, batch_first, dropout and bidirectional.
However, how do I have multiple cells in a single layer?
These cells will be automatically unrolled based on your sequence size in the input. Please check out this code:
# One cell RNN input_dim (4) -> output_dim (2). sequence: 5, batch 3
# 3 batches 'hello', 'eolll', 'lleel'
# rank = (3, 5, 4)
inputs = Variable(torch.Tensor([[h, e, l, l, o],
[e, o, l, l, l],
[l, l, e, e, l]]))
print("input size", inputs.size()) # input size torch.Size([3, 5, 4])
# Propagate input through RNN
# Input: (batch, seq_len, input_size) when batch_first=True
# B x S x I
out, hidden = cell(inputs, hidden)
print("out size", out.size()) # out size torch.Size([3, 5, 2])
You can find more examples at https://github.com/hunkim/PyTorchZeroToAll/.
Related
i would like to ask a question regarding the nn.BatchNorm1d in PyTorch.
I have one main tensor, which has shape [B, 3, N]. Then, i have two additional tensors which have shape [B, 3, V1] and [B, 3, V2]. I will concatenate the main tensor with the two tensors separately, to construct new tensors [B, 3, N+V1] and [B, 3, N+V2].
I pass my tensors to a plain MLP (consists of conv1d and batchnorm1d). Ideally, i want to predict something "point-wise", like no matter what the number of dimension 2, it has some consistent prediction only given the value. However, the batchnorm1d will have different results given input [B, 3, N+V1] and [B, 3, N+V2], while i am only focusing on first N points in 2nd dimension.
import torch
import torch.nn as nn
# nn.BatchNorm1d
B=2
dim=64
N=40000
V1=1000
v2=2000
torch.manual_seed(0)
x = torch.rand(B, dim, N) # here imgs are flattened from 28x28
v1 = torch.rand(B, dim, V1)
v2 = torch.rand(B, dim, v2)
layer = nn.BatchNorm1d(dim) # batch norm is done on channels
out2 = layer(torch.cat((x, v1), dim=2))
out3 = layer(torch.cat((x, v2), dim=2))
torch.equal(out2[:, :, :N], out3[:, :, :N])
Is there any possible way to have consistent prediction of first N points?
Is this more along the lines of what you're looking for? Normalizing just across the channels?
out2 = torch.cat((x, v1), dim=2) / torch.linalg.norm(torch.cat((x, v1), dim=2), dim=1, keepdim=True)
out3 = torch.cat((x, v2), dim=2) / torch.linalg.norm(torch.cat((x, v2), dim=2), dim=1, keepdim=True)
torch.equal(out2[:, :, :N], out3[:, :, :N])
# True
I think if you want to do something like this within pytorch nn libraries you'll need to transpose your channels and feature dimensions that way you can use LayerNorm1d or InstanceNorm. See here for a nice visual example of the different normalization techniques
Update answer:
In case you want to use an nn module specifically. InstanceNorm or GroupNorm could also get you the response. However the number of channels now differs between the two so you'll need two distinct layers.
layer1 = nn.GroupNorm(V1+N, V1+N)
layer2 = nn.GroupNorm(V2+N, V2+N)
out2 = layer1(torch.cat((x, v1), dim=2).transpose(1,2))
out3 = layer2(torch.cat((x, v2), dim=2).transpose(1,2))
torch.equal(out2[:, :N, :], out3[:, :N, :])
True
I am trying to implement a custom loss function in a Pytorch Autoencoder.
The loss function tries to maximize the cosine similarity between a given output tensor U (a vector) and 100 random vectors J where both U and J have the same dimension of [300]. This is repeated for each batch.
Suppose we have 30 items per batch, then the output tensor is
train_Y.shape = [30,300]
Random_vectors.shape = [30,100,300]
I can implement the loss function in two ways:
All_Y =[]
for Y,z_r in zip(train_y, random_vectors):
Y_cosine_list =[]
for z in z_r:
cosi = torch.dot(Y,z) / (torch.norm(Y)*torch.norm(z))
Y_cosine_list.append(cosi)
All_Y.append(Y_cosine_list)
All_Y = torch.tensor(All_Y).to(device)
train_loss = torch.sum(torch.abs(All_Y))/dim_0
train_loss = torch.tensor(train_loss.data, requires_grad = True)
or
train_Y = torch.zeros([dim_0, 100])
for i, (Y,z_r) in enumerate(zip(train_Y, random_vectors)):
for j,z in enumerate(z_r):
train_Y[i,j] = cos(Y,z)
train_Y = train_Y.to(device)
train_loss = torch.sum(torch.abs(train_Y))/dim_0
The second one is more elegant and to the point. However it is giving a "Cuda illegal memory access error". I have checked that the memory is not exceeded in either case. Is there anything wrong with the second implementation?
The first implementation is inelegant and I am not sure that it makes sense from a neural net optimization perspective. But it does not give errors and am able to complete training for all the epochs.
Ps: I have tried encapsulating this code block in a loss_fn method but I get the same illegal memory access error.
I have tried everything that I could find for the illegal memory access error - changing GPUs, removing a torch.stack block etc. But I can't seem to get rid of the problem.
Here is a vectorized way to do it
class CosineLoss(nn.Module):
def __init__(self, ):
super().__init__()
pass
def forward(self, x, y):
"""
Args:
x (torch.tensor): [batchsize, N, M] - tensor.
y (torch.tensor): [batchsize, M] - tensor.
Returns:
torch.tensor: scalar mean cosine loss
"""
# dot product along dimension 'm' i.e multiply and sum along 'm'.
dotp = torch.einsum("bm, bnm -> bn", y, x)
# L2 norm along dimension 'm' and multiply by broadcasting
length = torch.norm(y, dim=-1)[:, None]*torch.norm(x, dim=-1)
# cosine = dotproduct of unit vectors
cos = dotp/length
return cos.mean()
def test():
b, n, m = 30, 100, 300
train_Y = torch.randn(b, m, device='cuda')
random_vectors = torch.randn(b, n, m, requires_grad=True, device='cuda')
print(f'{random_vectors.grad = }')
cosineloss = CosineLoss()
loss = cosineloss(random_vectors, train_Y)
print(f'{loss = }')
loss.backward()
print(f'{random_vectors.grad.shape = }')
References:
einsum
broadcasting
julia language deep learning framework,
This is a quick start for Knet.jl,
https://denizyuret.github.io/Knet.jl/latest/tutorial/#Tutorial
ENV ["COLUMNS"] = 72
using Knet, MLDatasets, IterTools
struct Conv; w; b; f; end
(c :: Conv) (x) = c.f. (pool (conv4 (c.w, x). + C.b))
Conv (w1, w2, cx, cy, f = relu) = Conv (param (w1, w2, cx, cy), param0 (1,1, cy, 1), f);
The complex type Conv has three fields, w, b, and f.
The Conv type c (x) function broadcasts the next function with the f function.
The inner product of the w matrix and the x matrix is ​​calculated with conv4 (c.w, x), and the addition with c.b is performed with. +.
I don't know what the pool is looking for in that matrix.
This (pool (conv4 ...)) is passed through the relu activation function.
At the last Conv (w1, w2, cx, cy, f = relu) = Conv (param (w1, w2, cx, cy), param0 (1,1, cy, 1), f);
I don't know what I'm trying to do.
This is the situation of understanding.
What are you trying to do, especially in the pool?
Why are there two params on the 5th line?
I do not know.
Actually, the layer does a convolution followed by max pooling:
Pooling layers reduce the dimensions of the data by combining the outputs of neuron clusters at one layer into a single neuron in the next layer. Local pooling combines small clusters, typically 2 x 2. Global pooling acts on all the neurons of the convolutional layer. There are two common types of pooling: max and average. Max pooling uses the maximum value of each cluster of neurons at the prior layer, while average pooling instead uses the average value. (source: https://en.wikipedia.org/wiki/Convolutional_neural_network#Pooling_layers)
There are two params on the 5th line, because a convolutional layers has two trainable parameters: the kernel weights w and the bias b. The function param (and param0) initialize them with the correct size and mark them as trainable parameters that will be updated during the optimization.
To learn neural networks, I found these examples: linear regression and a simple feed-forward network (multilayer perceptron) quite useful.
I would like to understand how mIoU is calculated for multi-class classification. The formula for each class is
IoU formula
and then the average is done over the classes to get the mIoU. However, I don't understand what happens for the classes that are not represented. The formula becomes a division by 0, so I ignore them and the average is only computed for the classes represented.
The problem is that when a prediction is wrong, the accuracy is really lowered. It adds another class to make the average. For instance : in semantic segmentation the ground-truth of an image is made of 4 classes (0,1,2,3) and 6 classes are represented over the dataset. The prediction is also made of 4 classes (0,1,4,5) but all the items classified in 2 and 3 (in the ground-truth) are classified in 4 and 5 (in the prediction). In this case should we calculate the mIoU over 6 classes ? Even if 4 classes are totally wrong and there respective IoU is 0 ? So the problem is that if just one pixel is predicted in a class that is not in the ground_truth, we have to divide by a higher denominator and it lows a lot the score.
Is it the correct way to compute the mIoU for multi-class (and the semantic segmentation) ?
Instead of calculating the miou of each image and then calculate the "mean" miou over all the images, I calculate the miou as one big image. If a class is not in the image and is not predicited, I set there respective iou equal to 1.
From scratch :
def miou(gt,pred,nbr_mask):
intersection = np.zeros(nbr_mask) # int = (A and B)
den = np.zeros(nbr_mask) # den = A + B = (A or B) + (A and B)
for i in range(len(gt)):
for j in range(height):
for k in range(width):
if pred[i][j][k]==gt[i][j][k]:
intersection[gt[i][j][k]]+=1
den[pred[i][j][k]] += 1
den[gt[i][j][k]] += 1
mIoU = 0
for i in range(nbr_mask):
if den[i]!=0:
mIoU+=intersection[i]/(den[i]-intersection[i])
else:
mIoU+=1
mIoU=mIoU/nbr_mask
return mIoU
With gt the array of ground truth labels and pred the prediction of theassociated images (have to correspond in the array and be the same size).
Adding to the previous answer, this is a great fast and efficient pytorch GPU implementation of calculating the mIOU and classswise IOU for a batch of size (N, H, W) (both pred mask and labels), taken from the NeurIPS 2021 paper "Few-Shot Segmentation via Cycle-Consistent Transformer", github repo available here.
def intersectionAndUnionGPU(output, target, K, ignore_index=255):
# 'K' classes, output and target sizes are N or N * L or N * H * W, each value in range 0 to K - 1.
assert (output.dim() in [1, 2, 3])
assert output.shape == target.shape
output = output.view(-1)
target = target.view(-1)
output[target == ignore_index] = ignore_index
intersection = output[output == target]
area_intersection = torch.histc(intersection, bins=K, min=0, max=K-1)
area_output = torch.histc(output, bins=K, min=0, max=K-1)
area_target = torch.histc(target, bins=K, min=0, max=K-1)
area_union = area_output + area_target - area_intersection
return area_intersection, area_union, area_target
Example usage:
output = torch.rand(4, 5, 224, 224) # model output; batch size=4; channels=5, H,W=224
preds = F.softmax(output, dim=1).argmax(dim=1) # (4, 224, 224)
labels = torch.randint(0,5, (4, 224, 224))
i, u, _ = intersectionAndUnionGPU(preds, labels, 5) # 5 is num_classes
classwise_IOU = i/u # tensor of size (num_classes)
mIOU = i.sum()/u.sum() # mean IOU, taking (i/u).mean() is wrong
Hope this helps everyone!
(A non-GPU implementation is available as well in the repo!)
I'm training a LSTM model using pytorch with batch size of 256 and NLLLoss() as loss function.
The loss function is having problem with the data shape.
The softmax output from the forward passing has shape of torch.Size([256, 4, 1181]) where 256 is batch size, 4 is sequence length, and 1181 is vocab size.
The target is in the shape of torch.Size([256, 4]) where 256 is batch size and 4 is the output sequence length.
When I was testing earlier with batch size of 1, the model works fine but when I add batch size, it is breaking. I read that NLLLoss() can take class target as input instead of one hot encoded target.
Am I misunderstanding it? Or did I not format the shape of the target correctly?
class LSTM(nn.Module):
def __init__(self, embed_size=100, hidden_size=100, vocab_size=1181, embedding_matrix=...):
super(LSTM, self).__init__()
self.hidden_size = hidden_size
self.word_embeddings = nn.Embedding(vocab_size, embed_size)
self.word_embeddings.load_state_dict({'weight': torch.Tensor(embedding_matrix)})
self.word_embeddings.weight.requires_grad = False
self.lstm = nn.LSTM(embed_size, hidden_size)
self.hidden2out = nn.Linear(hidden_size, vocab_size)
def forward(self, tokens):
batch_size, num_steps = tokens.shape
embeds = self.word_embeddings(tokens)
lstm_out, _ = self.lstm(embeds.view(batch_size, num_steps, -1))
out_space = self.hidden2out(lstm_out.view(batch_size, num_steps, -1))
out_scores = F.log_softmax(out_space, dim=1)
return out_scores
model = LSTM(self.config.embed_size, self.config.hidden_size, self.config.vocab_size, self.embedding_matrix)
loss_function = nn.NLLLoss()
optimizer = optim.Adam(model.parameters(), lr=self.config.lr)
Error:
~/anaconda3/lib/python3.7/site-packages/torch/nn/functional.py in nll_loss(input, target, weight, size_average, ignore_index, reduce, reduction)
1846 if target.size()[1:] != input.size()[2:]:
1847 raise ValueError('Expected target size {}, got {}'.format(
-> 1848 out_size, target.size()))
1849 input = input.contiguous().view(n, c, 1, -1)
1850 target = target.contiguous().view(n, 1, -1)
ValueError: Expected target size (256, 554), got torch.Size([256, 4])
Your input shape to the loss function is (N, d, C) = (256, 4, 1181) and your target shape is (N, d) = (256, 4), however, according to the docs on NLLLoss the input should be (N, C, d) for a target of (N, d).
Supposing x is your network output and y is the target then you can compute loss by transposing the incorrect dimensions of x as follows:
loss = loss_function(x.transpose(1, 2), y)
Alternatively, since NLLLoss is just averaging all the responses anyway, you can reshape x and y to be (N*d, C) and (N*d). This gives the same result without creating temporary copies of your tensors.
loss = loss_function(x.reshape(N*d, C), y.reshape(N*d))