I am having an issue with processing continuous data using RNN. So far I've used MEL spectograms as inputs for my Listen, Attend and Spell architecture but I've decided to play around with that input by interpolating spectogram bins data rather than using MEL. Here is an example how one could achieve this: https://dspguru.com/dsp/howtos/how-to-interpolate-fft-peak/
The idea is simple, information gets lost in the process of creating MEL spectograms, some components get shifted etc. and my goal was to try and see if some sort of preprocessing could help deep network in the process of learning but I can't express this newly formed input to match RNN. MEL spectogram is 2D data with final number of bins, let's say 39. This new input can have arbitrary number of components for each time step. Every of these components is represented by a tuple of frequency and amplitude while both of them are continual variables. Frequency range for each time step can vary so I can't adopt a final number of them, first index for one time step could hold frequency value of, let's say, 5Hz while in the next time step it could very well be few hundreds of Hz. Could convolutional network be used as a tool before RNN? Is there any other solution? Thanks!
Related
I'm training a CNN (one using a series of ConvTranspose2D in pytorch) that uses input data from JSON to constitute an image. Unlike natural language, the input data can be in any order, as it contains info about various sprites in a scene.
In my first attempts to train the model, I didn't change the order of the input data (meaning, on each epoch, each sprite was represented in the same place in the input data). The model learned for about 10 epochs, but then there started to be divergence between the training loss (which continued to go down) and the test loss. So classic overfitting.
I tried to solve this by doing a form of data augmentation where the output data (in this case an image) stayed the same but I shuffled the order of the input data. As I have around 400 sprites, the maximum shuffling is 400!, so theoretically this can vastly expand the amount of training data. For example, instead of 100k JSON documents corresponding to 100K images, by shuffling the order of sprites in the input data, you have 400!*100000 training data points. In practice of course this amount of data is impractical, so I went with around 2m data points for an initial test. The issue I ran into here was that the model was not learning at all - after getting to a certain loss very quickly (after the first few mini-batches), it didn't learn at all for around 4 epochs. So classic underfitting.
Like Goldilocks, I'd like to find "just right" between the initial overfitting and subsequent underfitting. I'm wondering other strategies I could try out. One idea I had was letting the model train on a predetermined order of sprites (the overfitting case) and then, once overfitting starts (ie two straight epochs with divergence between the test and training loss) shuffling the data. I can also play with changing the model, although it can only be so big because of constraints with the hardware and the fact that inference needs to happen in under 20ms.
Are there any papers or techniques that are recommended in this scenario where data augmentation can lead to vastly more data points but results in a model ceasing to learn? Thanks in advance for any tips!
I've designed a variational autoencoder (VAE) that clusters sequential time series data.
To evaluate the performance of VAE on labeled data, First, I run KMeans on the raw data and compare the generated labels with the true labels using Adjusted Mutual Info Score (AMI). Then, after the model is trained, I pass validation data to it, run KMeans on latent vectors, and compare the generated labels with the true labels of validation data using AMI. Finally, I compare the two AMI scores with each other to see if KMeans has better performance on the latent vectors than the raw data.
My question is this: How can we evaluate the performance of VAE when the data is unlabeled?
I know we can run KMeans on the raw data and generate labels for it, but in this case, since we consider the generated labels as true labels, how can we compare the performance of KMeans on the raw data with KMeans on the latent vectors?
Note: The model is totally unsupervised. Labels (if exist) are not used in the training process. They're used only for evaluation.
In unsupervised learning you evaluate the performance of a model by either using labelled data or visual analysis. In your case you do not have labelled data, so you would need to do analysis. One way to do this is by looking at the predictions. If you know how the raw data should be labelled, you can qualitatively evaluate the accuracy. Another method is, since you are using KMeans, is to visualize the clusters. If the clusters are spread apart in groups, that is usually a good sign. However, if they are closer together and overlapping, the labelling of vectors in the respective areas may be less accurate. Alternatively, there may be some sort of a metric that you can use to evaluate the clusters or come up with your own.
I'm currently working on a project regarding a dataset that contains smartphone usage data from roundabout 200 users over a period of 4 months. For each user, I have a dataframe consisting of app-log events (Name of the App, Time, Location etc.). My goal is to predict the dwell time for the next app a user is going to open. I don't want to build one model for each user, but instead, I'm trying to build a model for all combined users. Now I'm struggling with finding an architecture that is suitable for this project.
The records are not evenly spaced in time, and the length of each dataframe differs. I want to utilize the temporal dependencies while simultaneously learn from multiple users at once, thus my input would be multiple parallel sequences of app usage durations with additional features and my output again multiple parallel sequences containing the dwell-time for the next app, but as the sequences are not evenly spaced in time nor have the same length it seems not suitable. I just wanted to get some ideas on how to structure the data properly and what you think would be a suitable approach. I would really appreciate some ideas or reading recommendations.
The following is how I understand the point of parameter sharing in RNNs:
In regular feed-forward neural networks, every input unit is assigned an individual parameter, which means that the number of input units (features) corresponds to the number of parameters to learn. In processing e.g. image data, the number of input units is the same over all training examples (usually constant pixel size * pixel size * rgb frames).
However, sequential input data like sentences can come in highly varying lengths, which means that the number of parameters will not be the same depending on which example sentence is processed. That is why parameter sharing is necessary for efficiently processing sequential data: it makes sure that the model always has the same input size regardless of the sequence length, as it is specified in terms of transition from one state to another. It is thus possible to use the same transition function with the same weights (input to hidden weights, hidden to output weights, hidden to hidden weights) at every time step. The big advantage is that it allows generalization to sequence lengths that did not appear in the training set.
My questions are:
Is my understanding of RNNs, as summarized above, correct?
In the actual code example in Keras I looked at for LSTMs, they padded the sentences to equal lengths before all. By doing so, doesn't this wash away the whole purpose of parameter sharing in RNNs?
Parameter Sharing
Being able to efficiently process sequences of varying length is not the only advantage of parameter sharing. As you said, you can achieve that with padding. The main purpose of parameter sharing is a reduction of the parameters that the model has to learn. This is the whole purpose of using a RNN.
If you would learn a different network for each time step and feed the output of the first model to the second etc. you would end up with a regular feed-forward network. For a number of 20 time steps, you would have 20 models to learn. In Convolutional Nets, parameters are shared by the Convolutional Filters because when we can assume that there are similar interesting patterns in different regions of the picture (for example a simple edge). This drastically reduces the number of parameters we have to learn. Analogously, in sequence learning we can often assume that there are similar patterns at different time steps. Compare 'Yesterday I ate an apple' and 'I ate an apple yesterday'. These two sentences mean the same, but the 'I ate an apple' part occurs on different time steps. By sharing parameters, you only have to learn what that part means once. Otherwise, you'd have to learn it for every time step, where it could occur in your model.
There is a drawback to sharing the parameters. Because our model applies the same transformation to the input at every time step, it now has to learn a transformation that makes sense for all time steps. So, it has to remember, what word came in which time step, i.e. 'chocolate milk' should not lead to the same hidden and memory state as 'milk chocolate'. But this drawback is small compared to using a large feed-forward network.
Padding
As for padding the sequences: the main purpose is not directly to let the model predict sequences of varying length. Like you said, this can be done by using parameter sharing. Padding is used for efficient training - specifically to keep the computational graph during training low. Without padding, we have two options for training:
We unroll the model for each training sample. So, when we have a sequence of length 7, we unroll the model to 7 time steps, feed the sequence, do back-propagation through the 7 time steps and update the parameters. This seems intuitive in theory. But in practice, this is inefficient, because TensorFlow's computational graphs don't allow recurrency, they are feedforward.
The other option is to create the computational graphs before starting training. We let them share the same weights and create one computational graph for every sequence length in our training data. But when our dataset has 30 different sequence lengths this means 30 different graphs during training, so for large models, this is not feasible.
This is why we need padding. We pad all sequences to the same length and then only need to construct one computational graph before starting training. When you have both very short and very long sequence lengths (5 and 100 for example), you can use bucketing and padding. This means, you pad the sequences to different bucket lengths, for example [5, 20, 50, 100]. Then, you create a computational graph for each bucket. The advantage of this is, that you don't have to pad a sequence of length 5 to 100, as you would waste a lot of time on "learning" the 95 padding tokens in there.
I would like to use one of Caffe's reference model i.e. bvlc_reference_caffenet. I found that my target class i.e. person is one of the classes included in the ILSVRC dataset that has been trained for the model. As my goal is to classify whether a test image contains a person or not, I may achieve this by the following:
Use inference directly with 1000 number of output. This doesn't
require training/learning.
Change the network topology a little bit with the final FC layer's number of output (num_output) is set to 2 (instead of 1000). Retrain it as a binary classification problem.
My concern is about computational effort at deployment/prediction phase (testing). The latter looks more expensive computationally than the former. This is because during prediction phase it needs to compute those 1000 output possibilities to find the one with the highest score. What I'm not sure is that, it could be the case that there's a heuristic (which I'm not aware of) that simplifies the computation.
Can somebody please help cross check my understanding on this.