I am currently trying to train an image classification model using Pytorch densenet121 with 4 labels (A, B, C, D). I have 224000 images and each image is labeled in the form of [1, 0, 0, 1] (Label A and D are present in the image). I have replaced the last dense layer of densenet121. The model is trained using Adam optimizer, LR of 0.0001 (with the decay of a factor of 10 per epoch), and trained for 4 epochs. I will try more epochs after I am confident that the class imbalanced issue is resolved.
The estimated number of positive classes is [19000, 65000, 38000, 105000] respectively. When I trained the model without class balancing and weights (with BCELoss), I have a very low recall for label A and C (in fact the True Positive TP and False Positive FP is less than 20)
I have tried 3 approaches to deal with the class imbalance after an extensive search on Google and Stackoverflow.
Approach 1: Class weights
I have tried to implement class weights by using the ratio of negative samples to positive samples.
y = train_df[CLASSES];
pos_weight = (y==0).sum()/(y==1).sum()
pos_weight = torch.Tensor(pos_weight)
if torch.cuda.is_available():
pos_weight = pos_weight.cuda()
criterion = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
The resultant class weights are [10.79, 2.45, 4.90, 1.13]. I am getting the opposite effect; having too many positive predictions which result in low precision.
Approach 2: Changing logic for class weights
I have also tried to get class weights by getting the proportion of the positive samples in the dataset and getting the inverse. The resultant class weights are [11.95, 3.49, 5.97, 2.16]. I am still getting too many positive predictions.
class_dist = y.apply(pd.Series.value_counts)
class_dist_norm = class_dist.loc[1.0]/class_dist.loc[1.0].sum()
pos_weight = 1/class_dist_norm
Approach 3: Focal Loss
I have also tried Focal Loss with the following implementation (but still getting too many positive predictions). I have used the class weights for the alpha parameter. This is referenced from https://gist.github.com/f1recracker/0f564fd48f15a58f4b92b3eb3879149b but I made some modifications to suit my use case better.
class FocalLoss(nn.CrossEntropyLoss):
''' Focal loss for classification tasks on imbalanced datasets '''
def __init__(self, alpha=None, gamma=1.5, ignore_index=-100, reduction='mean', epsilon=1e-6):
super().__init__(weight=alpha, ignore_index=ignore_index, reduction='mean')
self.reduction = reduction
self.gamma = gamma
self.epsilon = epsilon
self.alpha = alpha
def forward(self, input_, target):
# cross_entropy = super().forward(input_, target)
# Temporarily mask out ignore index to '0' for valid gather-indices input.
# This won't contribute final loss as the cross_entropy contribution
# for these would be zero.
target = target * (target != self.ignore_index).long()
# p_t = p if target = 1, p_t = (1-p) if target = 0, where p is the probability of predicting target = 1
p_t = input_ * target + (1 - input_) * (1 - target)
# Loss = -(alpha)( 1 - p_t)^gamma log(p_t), where -log(p_t) is cross entropy => loss = (alpha)(1-p_t)^gamma * cross_entropy (Epsilon added to prevent error with log(0) when class probability is 0)
if self.alpha != None:
loss = -1 * self.alpha * torch.pow(1 - p_t, self.gamma) * torch.log(p_t + self.epsilon)
else:
loss = -1 * torch.pow(1 - p_t, self.gamma) * torch.log(p_t + self.epsilon)
if self.reduction == 'mean':
return torch.mean(loss)
elif self.reduction == 'sum':
return torch.sum(loss)
else:
return loss
One thing to note is that the loss using stagnant after the first epoch, but the metrics varied between epochs.
I have considered undersampling and oversampling but I am unsure of how to proceed due to the fact that each image can have more than 1 label. One possible method is to oversample images with only 1 label by replicating them. But I am concerned that the model will only generalize on images with 1 label but perform poorly on images with multiple labels.
Therefore I would like to ask if there are methods that I should try, or did I make any mistakes in my approaches.
Any advice will be greatly appreciated.
Thank you!
Related
I am trying to train a machine learning model where the loss function is binary cross entropy, because of gpu limitations i can only do batch size of 4 and i'm having lot of spikes in the loss graph. So I'm thinking to back-propagate after some predefined batch size(>4). So it's like i'll do 10 iterations of batch size 4 store the losses, after 10th iteration add the losses and back-propagate. will it be similar to batch size of 40.
TL;DR
f(a+b) = f(a)+f(b) is it true for binary cross entropy?
f(a+b) = f(a) + f(b) doesn't seem to be what you're after. This would imply that BCELoss is additive which it clearly isn't. I think what you really care about is if for some index i
# false
f(x, y) == f(x[:i], y[:i]) + f([i:], y[i:])
is true?
The short answer is no, because you're missing some scale factors. What you probably want is the following identity
# true
f(x, y) == (i / b) * f(x[:i], y[:i]) + (1.0 - i / b) * f(x[i:], y[i:])
where b is the total batch size.
This identity is used as motivation for the gradient accumulation method (see below). Also, this identity applies to any objective function which returns an average loss across each batch element, not just BCE.
Caveat/Pitfall: Keep in mind that batch norm will not behave exactly the same when using this approach since it updates its internal statistics based on batch size during the forward pass.
We can actually do a little better memory-wise than just computing the loss as a sum followed by backpropagation. Instead we can compute the gradient of each component in the equivalent sum individually and allow the gradients to accumulate. To better explain I'll give some examples of equivalent operations
Consider the following model
import torch
import torch.nn as nn
import torch.nn.functional as F
class MyModel(nn.Module):
def __init__(self):
super().__init__()
num_outputs = 5
# assume input shape is 10x10
self.conv_layer = nn.Conv2d(3, 10, 3, 1, 1)
self.fc_layer = nn.Linear(10*5*5, num_outputs)
def forward(self, x):
x = self.conv_layer(x)
x = F.max_pool2d(x, 2, 2, 0, 1, False, False)
x = F.relu(x)
x = self.fc_layer(x.flatten(start_dim=1))
x = torch.sigmoid(x) # or omit this and use BCEWithLogitsLoss instead of BCELoss
return x
# to ensure same results for this example
torch.manual_seed(0)
model = MyModel()
# the examples will work as long as the objective averages across batch elements
objective = nn.BCELoss()
# doesn't matter what type of optimizer
optimizer = torch.optim.SGD(model.parameters(), lr=0.001)
and lets say our data and targets for a single batch are
torch.manual_seed(1) # to ensure same results for this example
batch_size = 32
input_data = torch.randn((batch_size, 3, 10, 10))
targets = torch.randint(0, 1, (batch_size, 20)).float()
Full batch
The body of our training loop for an entire batch may look something like this
# entire batch
output = model(input_data)
loss = objective(output, targets)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_value = loss.item()
print("Loss value: ", loss_value)
print("Model checksum: ", sum([p.sum().item() for p in model.parameters()]))
Weighted sum of loss on sub-batches
We could have computed this using the sum of multiple loss functions using
# This is simpler if the sub-batch size is a factor of batch_size
sub_batch_size = 4
assert (batch_size % sub_batch_size == 0)
# for this to work properly the batch_size must be divisible by sub_batch_size
num_sub_batches = batch_size // sub_batch_size
loss = 0
for sub_batch_idx in range(num_sub_batches):
start_idx = sub_batch_size * sub_batch_idx
end_idx = start_idx + sub_batch_size
sub_input = input_data[start_idx:end_idx]
sub_targets = targets[start_idx:end_idx]
sub_output = model(sub_input)
# add loss component for sub_batch
loss = loss + objective(sub_output, sub_targets) / num_sub_batches
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_value = loss.item()
print("Loss value: ", loss_value)
print("Model checksum: ", sum([p.sum().item() for p in model.parameters()]))
Gradient accumulation
The problem with the previous approach is that in order to apply back-propagation, pytorch needs to store intermediate results of layers in memory for every sub-batch. This ends up requiring a relatively large amount of memory and you may still run into memory consumption issues.
To alleviate this problem, instead of computing a single loss and performing back-propagation once, we could perform gradient accumulation. This gives equivalent results of the previous version. The difference here is that we instead perform a backward pass on each component of
the loss, only stepping the optimizer once all of them have been backpropagated. This way the computation graph is cleared after each sub-batch which will help with memory usage. Note that this works because .backward() actually accumulates (adds) the newly computed gradients to the existing .grad member of each model parameter. This is why optimizer.zero_grad() must be called only once, before the loop, and not during or after.
# This is simpler if the sub-batch size is a factor of batch_size
sub_batch_size = 4
assert (batch_size % sub_batch_size == 0)
# for this to work properly the batch_size must be divisible by sub_batch_size
num_sub_batches = batch_size // sub_batch_size
# Important! zero the gradients before the loop
optimizer.zero_grad()
loss_value = 0.0
for sub_batch_idx in range(num_sub_batches):
start_idx = sub_batch_size * sub_batch_idx
end_idx = start_idx + sub_batch_size
sub_input = input_data[start_idx:end_idx]
sub_targets = targets[start_idx:end_idx]
sub_output = model(sub_input)
# compute loss component for sub_batch
sub_loss = objective(sub_output, sub_targets) / num_sub_batches
# accumulate gradients
sub_loss.backward()
loss_value += sub_loss.item()
optimizer.step()
print("Loss value: ", loss_value)
print("Model checksum: ", sum([p.sum().item() for p in model.parameters()]))
I think 10 iterations of batch size 4 is same as one iteration of batch size 40, only here the time taken will be more. Across different training examples losses are added before backprop. But that doesn't make the function linear. BCELoss has a log component, and hence it is not a linear function. However what you said is correct. It will be similar to batch size 40.
I am trying to do a multiclass classification problem (containing 3 labels) with softmax regression.
This is my first rough implementation with gradient descent and back propagation (without using regularization and any advanced optimization algorithm) containing only 1 layer.
Also when learning-rate is big (>0.003) cost becomes NaN, on decreasing learning-rate the cost function works fine.
Can anyone explain what I'm doing wrong??
# X is (13,177) dimensional
# y is (3,177) dimensional with label 0/1
m = X.shape[1] # 177
W = np.random.randn(3,X.shape[0])*0.01 # (3,13)
b = 0
cost = 0
alpha = 0.0001 # seems too small to me but for bigger values cost becomes NaN
for i in range(100):
Z = np.dot(W,X) + b
t = np.exp(Z)
add = np.sum(t,axis=0)
A = t/add
loss = -np.multiply(y,np.log(A))
cost += np.sum(loss)/m
print('cost after iteration',i+1,'is',cost)
dZ = A-y
dW = np.dot(dZ,X.T)/m
db = np.sum(dZ)/m
W = W - alpha*dW
b = b - alpha*db
This is what I get :
cost after iteration 1 is 6.661713420377916
cost after iteration 2 is 23.58974203186562
cost after iteration 3 is 52.75811642877174
.............................................................
...............*upto 100 iterations*.................
.............................................................
cost after iteration 99 is 1413.555298639879
cost after iteration 100 is 1429.6533630169406
Well after some time i figured it out.
First of all the cost was increasing due to this :
cost += np.sum(loss)/m
Here plus sign is not needed as it will add all the previous cost computed on every epoch which is not what we want. This implementation is generally required during mini-batch gradient descent for computing cost over each epoch.
Secondly the learning rate is too big for this problem that's why cost was overshooting the minimum value and becoming NaN.
I looked in my code and find out that my features were of very different range (one was from -1 to 1 and other was -5000 to 5000) which was limiting my algorithm to use greater values for learning rate.
So I applied feature scaling :
var = np.var(X, axis=1)
X = X/var
Now learning rate can be much bigger (<=0.001).
I've been implementing VAE and IWAE models on the caltech silhouettes dataset and am having an issue where the VAE outperforms IWAE by a modest margin (test LL ~120 for VAE, ~133 for IWAE!). I don't believe this should be the case, according to both theory and experiments produced here.
I'm hoping someone can find some issue in how I'm implementing that's causing this to be the case.
The network I'm using to approximate q and p is the same as that detailed in the appendix of the paper above. The calculation part of the model is below:
data_k_vec = data.repeat_interleave(K,0) # Generate K samples (in my case K=50 is producing this behavior)
mu, log_std = model.encode(data_k_vec)
z = model.reparameterize(mu, log_std) # z = mu + torch.exp(log_std)*epsilon (epsilon ~ N(0,1))
decoded = model.decode(z) # this is the sigmoid output of the model
log_prior_z = torch.sum(-0.5 * z ** 2, 1)-.5*z.shape[1]*T.log(torch.tensor(2*np.pi))
log_q_z = compute_log_probability_gaussian(z, mu, log_std) # Definitions below
log_p_x = compute_log_probability_bernoulli(decoded,data_k_vec)
if model_type == 'iwae':
log_w_matrix = (log_prior_z + log_p_x - log_q_z).view(-1, K)
elif model_type =='vae':
log_w_matrix = (log_prior_z + log_p_x - log_q_z).view(-1, 1)*1/K
log_w_minus_max = log_w_matrix - torch.max(log_w_matrix, 1, keepdim=True)[0]
ws_matrix = torch.exp(log_w_minus_max)
ws_norm = ws_matrix / torch.sum(ws_matrix, 1, keepdim=True)
ws_sum_per_datapoint = torch.sum(log_w_matrix * ws_norm, 1)
loss = -torch.sum(ws_sum_per_datapoint) # value of loss that gets returned to training function. loss.backward() will get called on this value
Here are the likelihood functions. I had to fuss with the bernoulli LL in order to not get nan during training
def compute_log_probability_gaussian(obs, mu, logstd, axis=1):
return torch.sum(-0.5 * ((obs-mu) / torch.exp(logstd)) ** 2 - logstd, axis)-.5*obs.shape[1]*T.log(torch.tensor(2*np.pi))
def compute_log_probability_bernoulli(theta, obs, axis=1): # Add 1e-18 to avoid nan appearances in training
return torch.sum(obs*torch.log(theta+1e-18) + (1-obs)*torch.log(1-theta+1e-18), axis)
In this code there's a "shortcut" being used in that the row-wise importance weights are being calculated in the model_type=='iwae' case for the K=50 samples in each row, while in the model_type=='vae' case the importance weights are being calculated for the single value left in each row, so that it just ends up calculating a weight of 1. Maybe this is the issue?
Any and all help is huge - I thought that addressing the nan issue would permanently get me out of the weeds but now I have this new problem.
EDIT:
Should add that the training scheme is the same as that in the paper linked above. That is, for each of i=0....7 rounds train for 2**i epochs with a learning rate of 1e-4 * 10**(-i/7)
The K-sample importance weighted ELBO is
$$ \textrm{IW-ELBO}(x,K) = \log \sum_{k=1}^K \frac{p(x \vert z_k) p(z_k)}{q(z_k;x)}$$
For the IWAE there are K samples originating from each datapoint x, so you want to have the same latent statistics mu_z, Sigma_z obtained through the amortized inference network, but sample multiple z K times for each x.
So its computationally wasteful to compute the forward pass for data_k_vec = data.repeat_interleave(K,0), you should compute the forward pass once for each original datapoint, then repeat the statistics output by the inference network for sampling:
mu = torch.repeat_interleave(mu,K,0)
log_std = torch.repeat_interleave(log_std,K,0)
Then sample z_k. And now repeat your datapoints data_k_vec = data.repeat_interleave(K,0), and use the resulting tensor to efficiently evaluate the conditional p(x |z_k) for each importance sample z_k.
Note you may also want to use the logsumexp operation when calculating the IW-ELBO for numerical stability. I can't quite figure out what's going on with the log_w_matrix calculation in your post, but this is what I would do:
log_pz = ...
log_qzCx = ....
log_pxCz = ...
log_iw = log_pxCz + log_pz - log_qzCx
log_iw = log_iw.reshape(-1, K)
iwelbo = torch.logsumexp(log_iw, dim=1) - np.log(K)
EDIT: Actually after thinking about it a bit and using the score function identity, you can interpret the IWAE gradient as an importance weighted estimate of the standard single-sample gradient, so the method in the OP for calculation of the importance weights is equivalent (if a bit wasteful), provided you place a stop_gradient operator around the normalized importance weights, which you call w_norm. So I the main problem is the absence of this stop_gradient operator.
I wrote a vanilla autoencoder using only Dense layer.
Below is my code:
iLayer = Input ((784,))
layer1 = Dense(128, activation='relu' ) (iLayer)
layer2 = Dense(64, activation='relu') (layer1)
layer3 = Dense(28, activation ='relu') (layer2)
layer4 = Dense(64, activation='relu') (layer3)
layer5 = Dense(128, activation='relu' ) (layer4)
layer6 = Dense(784, activation='softmax' ) (layer5)
model = Model (iLayer, layer6)
model.compile(loss='binary_crossentropy', optimizer='adam')
(trainX, trainY), (testX, testY) = mnist.load_data()
print ("shape of the trainX", trainX.shape)
trainX = trainX.reshape(trainX.shape[0], trainX.shape[1]* trainX.shape[2])
print ("shape of the trainX", trainX.shape)
model.fit (trainX, trainX, epochs=5, batch_size=100)
Questions:
1) softmax provides probability distribution. Understood. This means, I would have a vector of 784 values with probability between 0 and 1. For example [ 0.02, 0.03..... upto 784 items], summing all 784 elements provides 1.
2) I don't understand how the binary crossentropy works with these values. Binary cross entropy is for two values of output, right?
In the context of autoencoders the input and output of the model is the same. So, if the input values are in the range [0,1] then it is acceptable to use sigmoid as the activation function of last layer. Otherwise, you need to use an appropriate activation function for the last layer (e.g. linear which is the default one).
As for the loss function, it comes back to the values of input data again. If the input data are only between zeros and ones (and not the values between them), then binary_crossentropy is acceptable as the loss function. Otherwise, you need to use other loss functions such as 'mse' (i.e. mean squared error) or 'mae' (i.e. mean absolute error). Note that in the case of input values in range [0,1] you can use binary_crossentropy, as it is usually used (e.g. Keras autoencoder tutorial and this paper). However, don't expect that the loss value becomes zero since binary_crossentropy does not return zero when both prediction and label are not either zero or one (no matter they are equal or not). Here is a video from Hugo Larochelle where he explains the loss functions used in autoencoders (the part about using binary_crossentropy with inputs in range [0,1] starts at 5:30)
Concretely, in your example, you are using the MNIST dataset. So by default the values of MNIST are integers in the range [0, 255]. Usually you need to normalize them first:
trainX = trainX.astype('float32')
trainX /= 255.
Now the values would be in range [0,1]. So sigmoid can be used as the activation function and either of binary_crossentropy or mse as the loss function.
Why binary_crossentropy can be used even when the true label values (i.e. ground-truth) are in the range [0,1]?
Note that we are trying to minimize the loss function in training. So if the loss function we have used reaches its minimum value (which may not be necessarily equal to zero) when prediction is equal to true label, then it is an acceptable choice. Let's verify this is the case for binray cross-entropy which is defined as follows:
bce_loss = -y*log(p) - (1-y)*log(1-p)
where y is the true label and p is the predicted value. Let's consider y as fixed and see what value of p minimizes this function: we need to take the derivative with respect to p (I have assumed the log is the natural logarithm function for simplicity of calculations):
bce_loss_derivative = -y*(1/p) - (1-y)*(-1/(1-p)) = 0 =>
-y/p + (1-y)/(1-p) = 0 =>
-y*(1-p) + (1-y)*p = 0 =>
-y + y*p + p - y*p = 0 =>
p - y = 0 => y = p
As you can see binary cross-entropy have the minimum value when y=p, i.e. when the true label is equal to predicted label and this is exactly what we are looking for.
Could someone give a clear explanation of backpropagation for LSTM RNNs?
This is the type structure I am working with. My question is not posed at what is back propagation, I understand it is a reverse order method of calculating the error of the hypothesis and output used for adjusting the weights of neural networks. My question is how LSTM backpropagation is different then regular neural networks.
I am unsure of how to find the initial error of each gates. Do you use the first error (calculated by hypothesis minus output) for each gate? Or do you adjust the error for each gate through some calculation? I am unsure how the cell state plays a role in the backprop of LSTMs if it does at all. I have looked thoroughly for a good source for LSTMs but have yet to find any.
That's a good question. You certainly should take a look at suggested posts for details, but a complete example here would be helpful too.
RNN Backpropagaion
I think it makes sense to talk about an ordinary RNN first (because LSTM diagram is particularly confusing) and understand its backpropagation.
When it comes to backpropagation, the key idea is network unrolling, which is way to transform the recursion in RNN into a feed-forward sequence (like on the picture above). Note that abstract RNN is eternal (can be arbitrarily large), but each particular implementation is limited because the memory is limited. As a result, the unrolled network really is a long feed-forward network, with few complications, e.g. the weights in different layers are shared.
Let's take a look at a classic example, char-rnn by Andrej Karpathy. Here each RNN cell produces two outputs h[t] (the state which is fed into the next cell) and y[t] (the output on this step) by the following formulas, where Wxh, Whh and Why are the shared parameters:
In the code, it's simply three matrices and two bias vectors:
# model parameters
Wxh = np.random.randn(hidden_size, vocab_size)*0.01 # input to hidden
Whh = np.random.randn(hidden_size, hidden_size)*0.01 # hidden to hidden
Why = np.random.randn(vocab_size, hidden_size)*0.01 # hidden to output
bh = np.zeros((hidden_size, 1)) # hidden bias
by = np.zeros((vocab_size, 1)) # output bias
The forward pass is pretty straightforward, this example uses softmax and cross-entropy loss. Note each iteration uses the same W* and h* arrays, but the output and hidden state are different:
# forward pass
for t in xrange(len(inputs)):
xs[t] = np.zeros((vocab_size,1)) # encode in 1-of-k representation
xs[t][inputs[t]] = 1
hs[t] = np.tanh(np.dot(Wxh, xs[t]) + np.dot(Whh, hs[t-1]) + bh) # hidden state
ys[t] = np.dot(Why, hs[t]) + by # unnormalized log probabilities for next chars
ps[t] = np.exp(ys[t]) / np.sum(np.exp(ys[t])) # probabilities for next chars
loss += -np.log(ps[t][targets[t],0]) # softmax (cross-entropy loss)
Now, the backward pass is performed exactly as if it was a feed-forward network, but the gradient of W* and h* arrays accumulates the gradients in all cells:
for t in reversed(xrange(len(inputs))):
dy = np.copy(ps[t])
dy[targets[t]] -= 1
dWhy += np.dot(dy, hs[t].T)
dby += dy
dh = np.dot(Why.T, dy) + dhnext # backprop into h
dhraw = (1 - hs[t] * hs[t]) * dh # backprop through tanh nonlinearity
dbh += dhraw
dWxh += np.dot(dhraw, xs[t].T)
dWhh += np.dot(dhraw, hs[t-1].T)
dhnext = np.dot(Whh.T, dhraw)
Both passes above are done in chunks of size len(inputs), which corresponds to the size of the unrolled RNN. You might want to make it bigger to capture longer dependencies in the input, but you pay for it by storing all outputs and gradients per each cell.
What's different in LSTMs
LSTM picture and formulas look intimidating, but once you coded plain vanilla RNN, the implementation of LSTM is pretty much same. For example, here is the backward pass:
# Loop over all cells, like before
d_h_next_t = np.zeros((N, H))
d_c_next_t = np.zeros((N, H))
for t in reversed(xrange(T)):
d_x_t, d_h_prev_t, d_c_prev_t, d_Wx_t, d_Wh_t, d_b_t = lstm_step_backward(d_h_next_t + d_h[:,t,:], d_c_next_t, cache[t])
d_c_next_t = d_c_prev_t
d_h_next_t = d_h_prev_t
d_x[:,t,:] = d_x_t
d_h0 = d_h_prev_t
d_Wx += d_Wx_t
d_Wh += d_Wh_t
d_b += d_b_t
# The step in each cell
# Captures all LSTM complexity in few formulas.
def lstm_step_backward(d_next_h, d_next_c, cache):
"""
Backward pass for a single timestep of an LSTM.
Inputs:
- dnext_h: Gradients of next hidden state, of shape (N, H)
- dnext_c: Gradients of next cell state, of shape (N, H)
- cache: Values from the forward pass
Returns a tuple of:
- dx: Gradient of input data, of shape (N, D)
- dprev_h: Gradient of previous hidden state, of shape (N, H)
- dprev_c: Gradient of previous cell state, of shape (N, H)
- dWx: Gradient of input-to-hidden weights, of shape (D, 4H)
- dWh: Gradient of hidden-to-hidden weights, of shape (H, 4H)
- db: Gradient of biases, of shape (4H,)
"""
x, prev_h, prev_c, Wx, Wh, a, i, f, o, g, next_c, z, next_h = cache
d_z = o * d_next_h
d_o = z * d_next_h
d_next_c += (1 - z * z) * d_z
d_f = d_next_c * prev_c
d_prev_c = d_next_c * f
d_i = d_next_c * g
d_g = d_next_c * i
d_a_g = (1 - g * g) * d_g
d_a_o = o * (1 - o) * d_o
d_a_f = f * (1 - f) * d_f
d_a_i = i * (1 - i) * d_i
d_a = np.concatenate((d_a_i, d_a_f, d_a_o, d_a_g), axis=1)
d_prev_h = d_a.dot(Wh.T)
d_Wh = prev_h.T.dot(d_a)
d_x = d_a.dot(Wx.T)
d_Wx = x.T.dot(d_a)
d_b = np.sum(d_a, axis=0)
return d_x, d_prev_h, d_prev_c, d_Wx, d_Wh, d_b
Summary
Now, back to your questions.
My question is how is LSTM backpropagation different then regular Neural Networks
The are shared weights in different layers, and few more additional variables (states) that you need to pay attention to. Other than this, no difference at all.
Do you use the first error (calculated by hypothesis minus output) for each gate? Or do you adjust the error for each gate through some calculation?
First up, the loss function is not necessarily L2. In the example above it's a cross-entropy loss, so initial error signal gets its gradient:
# remember that ps is the probability distribution from the forward pass
dy = np.copy(ps[t])
dy[targets[t]] -= 1
Note that it's the same error signal as in ordinary feed-forward neural network. If you use L2 loss, the signal indeed equals to ground-truth minus actual output.
In case of LSTM, it's slightly more complicated: d_next_h = d_h_next_t + d_h[:,t,:], where d_h is the upstream gradient the loss function, which means that error signal of each cell gets accumulated. But once again, if you unroll LSTM, you'll see a direct correspondence with the network wiring.
I think your questions could not be answered in a short response. Nico's simple LSTM has a link to a great paper from Lipton et.al., please read this. Also his simple python code sample helps to answer most of your questions.
If you understand Nico's last sentence
ds = self.state.o * top_diff_h + top_diff_s
in detail, please give me a feed back. At the moment I have a final problem with his "Putting all this s and h derivations together".