I have a fully working seq2seq attention model with beam search and it does give improved results. But it takes > 1min for inferencing (batch-size 1024) with k=5 (k is my hypotheses) because none of it is parallelised. Everything happens 1 sample at a time.
Task (simplified)
Goal is sentence translation, 15 words Lang A to 15 words Lang B.
Encoder is a RNN that takes in 15 word sentence and encodes a representation of it, gives out a [timestep, 512] matrix along with final hidden state.
Decoder is another RNN, takes encoder hidden state as initial state, uses [timestep, 512] matrix for attention and outputs translated words[batches] one timestep at a time. Naturally, there is some form of parallelization till this point.
During inference stage, beam search is used. At each timestep of the decoder, rather than taking the predicted word with highest prob, I take k best words. And provide k words as input to the next timestep so that it can predict the next word in the sentence (rest of the algorithm is given below). Algorithm makes decoding less greedy anticipating results with higher total probability in succeeding timesteps.
for each element in the test-set
calculate initial k (decoder-encoder step)
for range(timesteps-1)
for each prev k
get hidden state
obtain its best k
save hidden state
find new k from k*k possible ones
##update hypotheses based on new found k
for element in k
copy hidden state
change hypotheses if necessary
append new k to hypotheses
There are 6 tensors and 2 lists to keep track and handle state changes. Is there any room for speedup or parallelisation here? perhaps each k can go through enncode-decode simultaneously? Any help is much appreciated.
Related
Would this be a valid implementation of a cross entropy loss that takes the ordinal structure of the GT y into consideration? y_hat is the prediction from a neural network.
ce_loss = F.cross_entropy(y_hat, y, reduction="none")
distance_weight = torch.abs(y_hat.argmax(1) - y) + 1
ordinal_ce_loss = torch.mean(distance_weight * ce_loss)
I'll attempt to answer this question by first fully defining the task, since the question is a bit sparse on details.
I have a set of ordinal classes (e.g. first, second, third, fourth,
etc.) and I would like to predict the class of each data example from
among this set. I would like to define an entropy-based loss-function
for this problem. I would like this loss function to weight the loss
between a predicted class torch.argmax(y_hat) and the true class y
according to the ordinal distance between the two classes. Does the
given loss expression accomplish this?
Short answer: sure, it is "valid". You've roughly implemented L1-norm ordinal class weighting. I'd question whether this is truly the correct weighting strategy for this problem.
For instance, consider that for a true label n, the bin n response is weighted by 1, but the bin n+1 and n-1 responses are weighted by 2. This means that a lot more emphasis will be placed on NOT predicting false positives than on correctly predicting true positives, which may imbue your model with some strange bias.
It also means that examples on the edge will result in a larger total sum of weights, meaning that you'll be weighting examples where the true label is say "first" or "last" more highly than the intermediate classes. (Say you have 5 classes: 1,2,3,4,5. A true label of 1 will require distance_weight of [1,2,3,4,5], the sum of which is 15. A true label of 3 will require distance_weight of [3,2,1,2,3], the sum of which is 11.
In general, classification problems and entropy-based losses are underpinned by the assumption that no set of classes or categories is any more or less related than any other set of classes. In essence, the input data is embedded into an orthogonal feature space where each class represents one vector in the basis. This is quite plainly a bad assumption in your case, meaning that this embedding space is probably not particularly elegant: thus, you have to correct for it with sort of a hack-y weight fix. And in general, this assumption of class non-correlation is probably not true in a great many classification problems (consider e.g. the classic ImageNet classification problem, wherein the class pairs [bus,car], and [bus,zebra] are treated as equally dissimilar. But this is probably a digression into the inherent lack of usefulness of strict ontological structuring of information which is outside the scope of this answer...)
Long Answer: I'd highly suggest moving into a space where the ordinal value you care about is instead expressed in a continuous space. (In the first, second, third example, you might for instance output a continuous value over the range [1,max_place]. This allows you to benefit from loss functions that already capture well the notion that predictions closer in an ordered space are better than predictions farther away in an ordered space (e.g. MSE, Smooth-L1, etc.)
Let's consider one more time the case of the [first,second,third,etc.] ordinal class example, and say that we are trying to predict the places of a set of runners in a race. Consider two races, one in which the first place runner wins by 30% relative to the second place runner, and the second in which the first place runner wins by only 1%. This nuance is entirely discarded by the ordinal discrete classification. In essence, the selection of an ordinal set of classes truncates the amount of information conveyed in the prediction, which means not only that the final prediction is less useful, but also that the loss function encodes this strange truncation and binarization, which is then reflected (perhaps harmfully) in the learned model. This problem could likely be much more elegantly solved by regressing the finishing position, or perhaps instead by regressing the finishing time, of each athlete, and then performing the final ordinal classification into places OUTSIDE of the network training.
In conclusion, you might expect a well-trained ordinal classifier to produce essentially a normal distribution of responses across the class bins, with the distribution peak on the true value: a binned discretization of a space that almost certainly could, and likely should, be treated as a continuous space.
After reading the research paper on batchnorm and its various descriptions in forums, I am still not clear how the basic computations are performed. The core of my questions is: a vector is normalized with respect to the set to which it belongs; we can thus normalize vectors input to layer 1 using the batch selected from the training set. Each input vector to the next layer needs to be normalized with respect to the set to which it belongs, but how do we get hold of that set?
More precisely, let
N = batch size;
Bi = set of vectors Xij (j = 1..n) whose normalized value will be
input to layer i of the network;
BN = Batch normalization function;
BN(Xij, Bi) = Normalized version of j-th vector, Xij, with respect to the set Bi.
BN(X1j, B1), j = 1..n, can be calculated because we know B1. They are inputs to layer 1.
We need BN(X2j, B2), j = 1..n, to input to layer 2, but we do not have B2 readily available. My questions is how to get B2, B3, etc.
We could process each BN(X1j, B1), j = 1..n, by layer 1 and remember the outputs as X2j (that collection will be B2). Then calculate BN(x2j, B2) for each j by normalizing with respect to B2 and input them to layer 2, etc. So the forward pass would consists of many such steps. For simplicity, I have ignored the scale and shift step, as that is not relevant to my question.
Being new to this topic, I would appreciate expert opinion on it.
Batch Norm first subtracts the mean to center the output around zero. Then Batch Normalization devides by the standart derivation to scale the output to unit variance.
This means that each vector is normalized with respect to the batch it belongs to. Assume you run your x through your first layer, lets call its output y1. Batch Normalization is then applied to y1 in the form of (y1 - y1.mean()) / y.std(). The 'set' you're talking about is just the vector y1, the output of the layer.
All this is of course ignoring versions where a bias and a derivation.
I want to predict the trajectory of a ball falling. That trajectory is parabolic. I know that LSTM may be too much for this (i.e. a simpler method could suffice).
I thought that we can do this with 2 LSTM layers and a Dense layer at the end.
The end result that I want is to give the model 3 heights h0,h1,h2 and let it predict h3. Then, I want to give it h1, h2, and the h3 it outputted previously to predict h4, and so on, until I can predict the whole trajectory.
Firstly, what would the input shape be for the first LSTM layer ?
Would it be input_shape = (3,1) ?
Secondly, would the LSTM be able to predict a parabolic path ?
I am getting almost a flat line, not a parabola, and I want to rule out the possibility that I am misunderstanding how to feed and shape input.
Thank you
The input shape is in the form (samples, timeSteps, features).
Your only feature is "height", so features = 1.
And since you're going to input sequences with different lengths, you can use timeSteps = None.
So, your input_shape could be (None, 1).
Since we're going to use a stateful=True layer below, we can use batch_input_shape=(1,None,1). Choose the amount of "samples" you want.
Your model can predict the trajectory indeed, but maybe it will need more than one layer. (The exact answer about how many layers and cells depend on knowing how the match inside LSTM works).
Training:
Now, first you need to train your network (only then it will be able to start predicting good things).
For training, suppose you have a sequence of [h1,h2,h3,h4,h5,h6...], true values in the correct sequence. (I suggest you have actually many sequences (samples), so your model learns better).
For this sequence, you want an output predicting the next step, then your target would be [h2,h3,h4,h5,h6,h7...]
So, suppose you have a data array with shape (manySequences, steps, 1), you make:
x_train = data[:,:-1,:]
y_train = data[:,1:,:]
Now, your layers should be using return_sequences=True. (Every input step produces an output step). And you train the model with this data.
A this point, whether you're using stateful=True or stateful=False is not very relevant. (But if true, you always need model.reset_state() before every single epoch and sequence)
Predicting:
For predicting, you can use stateful=True in the model. This means that when you input h1, it will produce h2. And when you input h2 it will remember the "current speed" (the state of the model) to predict the correct h3.
(In the training phase, it's not important to have this, because you're inputting the entire sequences at once. So the speed will be understood between steps of the long sequences).
You can se the method reset_states() as set_current_speed_to(0). You will use it whenever the step you're going to input is the first step in a sequence.
Then you can do loops like this:
model.reset_states() #make speed = 0
nextH = someValueWithShape((1,1,1))
predictions = [nextH]
for i in range(steps):
nextH = model.predict(nextH)
predictions.append(nextH)
There is an example here, but using two features. There is a difference that I use two models, one for training, one for predicting, but you can use only one with return_sequences=True and stateful=True (don't forget to reset_states() at the beginning of every epoch in training).
Closed. This question needs to be more focused. It is not currently accepting answers.
Want to improve this question? Update the question so it focuses on one problem only by editing this post.
Closed 2 years ago.
Improve this question
I'm going through this tutorial on RNNs/LSTMs and I'm having quite a hard time understanding stateful LSTMs. My questions are as follows :
1. Training batching size
In the Keras docs on RNNs, I found out that the hidden state of the sample in i-th position within the batch will be fed as input hidden state for the sample in i-th position in the next batch. Does that mean that if we want to pass the hidden state from sample to sample we have to use batches of size 1 and therefore perform online gradient descent? Is there a way to pass the hidden state within a batch of size >1 and perform gradient descent on that batch ?
2. One-Char Mapping Problems
In the tutorial's paragraph 'Stateful LSTM for a One-Char to One-Char Mapping' were given a code that uses batch_size = 1 and stateful = True to learn to predict the next letter of the alphabet given a letter of the alphabet. In the last part of the code (line 53 to the end of the complete code), the model is tested starting with a random letter ('K') and predicts 'B' then given 'B' it predicts 'C', etc. It seems to work well except for 'K'. However, I tried the following tweak to the code (last part too, I kept lines 52 and above):
# demonstrate a random starting point
letter1 = "M"
seed1 = [char_to_int[letter1]]
x = numpy.reshape(seed, (1, len(seed), 1))
x = x / float(len(alphabet))
prediction = model.predict(x, verbose=0)
index = numpy.argmax(prediction)
print(int_to_char[seed1[0]], "->", int_to_char[index])
letter2 = "E"
seed2 = [char_to_int[letter2]]
seed = seed2
print("New start: ", letter1, letter2)
for i in range(0, 5):
x = numpy.reshape(seed, (1, len(seed), 1))
x = x / float(len(alphabet))
prediction = model.predict(x, verbose=0)
index = numpy.argmax(prediction)
print(int_to_char[seed[0]], "->", int_to_char[index])
seed = [index]
model.reset_states()
and these outputs:
M -> B
New start: M E
E -> C
C -> D
D -> E
E -> F
It looks like the LSTM did not learn the alphabet but just the positions of the letters, and that regardless of the first letter we feed in, the LSTM will always predict B since it's the second letter, then C and so on.
Therefore, how does keeping the previous hidden state as initial hidden state for the current hidden state help us with the learning given that during test if we start with the letter 'K' for example, letters A to J will not have been fed in before and the initial hidden state won't be the same as during training ?
3. Training an LSTM on a book for sentence generation
I want to train my LSTM on a whole book to learn how to generate sentences and perhaps learn the authors style too, how can I naturally train my LSTM on that text (input the whole text and let the LSTM figure out the dependencies between the words) instead of having to 'artificially' create batches of sentences from that book myself to train my LSTM on? I believe I should use stateful LSTMs could help but I'm not sure how.
Having a stateful LSTM in Keras means that a Keras variable will be used to store and update the state, and in fact you could check the value of the state vector(s) at any time (that is, until you call reset_states()). A non-stateful model, on the other hand, will use an initial zero state every time it processes a batch, so it is as if you always called reset_states() after train_on_batch, test_on_batch and predict_on_batch. The explanation about the state being reused for the next batch on stateful models is just about that difference with non-stateful; of course the state will always flow within each sequence in the batch and you do not need to have batches of size 1 for that to happen. I see two scenarios where stateful models are useful:
You want to train on split sequences of data because these are very long and it would not be practical to train on their whole length.
On prediction time, you want to retrieve the output for each time point in the sequence, not just at the end (either because you want to feed it back into the network or because your application needs it). I personally do that in the models that I export for later integration (which are "copies" of the training model with batch size of 1).
I agree that the example of an RNN for the alphabet does not really seem very useful in practice; it will only work when you start with the letter A. If you want to learn to reproduce the alphabet starting at any letter, you would need to train the network with that kind of examples (subsequences or rotations of the alphabet). But I think a regular feed-forward network could learn to predict the next letter of the alphabet training on pairs like (A, B), (B, C), etc. I think the example is meant for demonstrative purposes more than anything else.
You may have probably already read it, but the popular post The Unreasonable Effectiveness of Recurrent Neural Networks shows some interesting results along the lines of what you want to do (although it does not really dive into implementation specifics). I don't have personal experience training RNN with textual data, but there is a number of approaches you can research. You can build character-based models (like the ones in the post), where your input and receive one character at a time. A more advanced approach is to do some preprocessing on the texts and transform them into sequences of numbers; Keras includes some text preprocessing functions to do that. Having one single number as feature space is probably not going to work all that well, so you could simply turn each word into a vector with one-hot encoding or, more interestingly, have the network learn the best vector representation for each for, which is what they call en embedding. You can go even further with the preprocessing and look into something like NLTK, specially if you want to remove stop words, punctuation and things like that. Finally, if you have sequences of different sizes (e.g. you are using full texts instead of excerpts of a fixed size, which may or may not be important for you) you will need to be a bit more careful and use masking and/or sample weighting. Depending on the exact problem, you can set up the training accordingly. If you want to learn to generate similar text, the "Y" would be the similar to the "X" (one-hot encoded), only shifted by one (or more) positions (in this case you may need to use return_sequences=True and TimeDistributed layers). If you want to determine the autor, your output could be a softmax Dense layer.
Hope that helps.
How should stereo (2 channel) audio data be represented for FFT? Do you
A. Take the average of the two channels and assign it to the real component of a number and leave the imaginary component 0.
B. Assign one channel to the real component and the other channel to the imag component.
Is there a reason to do one or the other? I searched the web but could not find any definite answers on this.
I'm doing some simple spectrum analysis and, not knowing any better, used option A). This gave me an unexpected result, whereas option B) went as expected. Here are some more details:
I have a WAV file of a piano "middle-C". By definition, middle-C is 260Hz, so I would expect the peak frequency to be at 260Hz and smaller peaks at harmonics. I confirmed this by viewing the spectrum via an audio editing software (Sound Forge). But when I took the FFT myself, with option A), the peak was at 520Hz. With option B), the peak was at 260Hz.
Am I missing something? The explanation that I came up with so far is that representing stereo data using a real and imag component implies that the two channels are independent, which, I suppose they're not, and hence the mess-up.
I don't think you're taking the average correctly. :-)
C. Process each channel separately, assigning the amplitude to the real component and leaving the imaginary component as 0.
Option B does not make sense. Option A, which amounts to convert the signal to mono, is OK (if you are interested in a global spectrum).
Your problem (double freq) is surely related to some misunderstanding in the use of your FFT routines.
Once you take the FFT you need to get the Magnitude of the complex frequency spectrum. To get the magnitude you take the absolute of the complex spectrum |X(w)|. If you want to look at the power spectrum you square the magnitude spectrum, |X(w)|^2.
In terms of your frequency shift I think it has to do with you setting the imaginary parts to zero.
If you imagine the complex Frequency spectrum as a series of complex vectors or position vectors in a cartesian space. If you took one discrete frequency bin X(w), there would be one real component representing its direction in the real axis (x -direction), and one imaginary component in the in the imaginary axis (y - direction). There are four important values about this discrete frequency, 1. real value, 2. imaginary value, 3. Magnitude and, 4. phase. If you just take the real value and set imaginary to 0, you are setting Magnitude = real and phase = 0deg or 90deg. You have hence forth modified the resulting spectrum, and applied a bias to every frequency bin. Take a look at the wiki on Magnitude of a vector, also called the Euclidean norm of a vector to brush up on your understanding. Leonbloy was correct, but I hope this was more informative.
Think of the FFT as a way to get information from a single signal. What you are asking is what is the best way to display data from two signals. My answer would be to treat each independently, and display an FFT for each.
If you want a really fast streaming FFT you can read about an algorithm I wrote here: www.depthcharged.us/?p=176