Darpan #14
Darpan #14
Conversation
| import pickle as p | ||
| import numpy as np | ||
| import scipy.io.wavfile | ||
| from scipy.fftpack import dct |
There was a problem hiding this comment.
Move this down to the from imports and correct the spacing.
| Data is divided by an 80-20 split into train and test | ||
| ''' | ||
| def __init__(self, path=None): | ||
| self.data_dir = '/home/darpan/sem4/free-spoken-digit-dataset/recordings/' |
There was a problem hiding this comment.
This is a hard-coded path, and will fail on general machines. Take a look at the MNIST class for a default way of doing this.
| class SpokenMNIST: | ||
| ''' | ||
| Data is divided by an 80-20 split into train and test | ||
| ''' |
There was a problem hiding this comment.
More descriptive doc-string + punctuaction.
| labels.append(label) | ||
| return audios, torch.Tensor(labels) | ||
|
|
||
| def pre_process(self, file): |
There was a problem hiding this comment.
Maybe preprocess instead of pre_process?
| filter_banks = 20 * np.log10(filter_banks) # dB | ||
|
|
||
| return filter_banks, label | ||
|
|
There was a problem hiding this comment.
In general, this method seems really messy / hard-coded. Perhaps some of the parameters should be arguments to the function (e.g., NFFT, frame_size, frame_stride, etc.).
| spike train from that timestep onwards | ||
| Inputs must be non-negative. Spike inter-arrival times are inversely proportional to | ||
| input magnitude, so data must be scaled according to desired spike frequency. | ||
There was a problem hiding this comment.
But, it's not a "mixture" per se, is it? I feel like the function header is misleading.
|
|
||
| # Yield Bernoulli-distributed spike trains. | ||
| return s | ||
|
|
There was a problem hiding this comment.
I think we can simply add functionality to get_bernoulli in which the time dimension is implicit; e.g., we could pass time=None, and based on this, discard the explicit time dimension.
| Most important functions to use: | ||
| confidence_weighting() | ||
| ngram() | ||
There was a problem hiding this comment.
Dunno what this string is for. Could you remove it?
| ''' | ||
|
|
||
| def get_fire_order(example): |
There was a problem hiding this comment.
Unclear what this function is doing without a doc string. I suggest get_firing_order.
| fire_order.append(n_id) | ||
| return fire_order | ||
|
|
||
| def normalize_probability(v): |
There was a problem hiding this comment.
Is it necessary to put this in a separate function? In general, you should avoid proliferation of functions inside of library modules, since they can be imported by end-users.
| Top-k | ||
| Precision | ||
| Recall | ||
| ''' |
There was a problem hiding this comment.
Re-write doc string to conform to style elsewhere in package.
| if not tuple(seq) in ngrams: | ||
| ngrams[tuple(seq)] = np.zeros(10) | ||
| ngrams[tuple(seq)][true_label] += 1 | ||
|
|
There was a problem hiding this comment.
Consider folding both of these functions into the ngram function.
| from nodes import AdaptiveLIFNodes, LIFNodes, Input | ||
| from analysis.plotting import plot_input, plot_spikes, plot_weights | ||
|
|
||
| from evaluation import * |
| images, labels = MNIST(path='../data').get_train() | ||
| images /= (255 * min_isi) # Normalize and enforce minimum expected inter-spike interval. | ||
| images = images.view(images.size(0), -1) # Flatten images to one dimension. | ||
| labels = [int(lbl) for lbl in labels] |
There was a problem hiding this comment.
Why do this? Having the labels lazily returned from a generator is a good thing. Same with images, which should be returned as a generator as in the get_poisson and get_bernoulli functions.
| print('Begin training.\n') | ||
| start = default_timer() | ||
|
|
||
| train_spikes = [] |
There was a problem hiding this comment.
Consider just appending spike_record to disk every update_interval. This way, we won't have to store n_samples * time * n_neurons spikes.
djsaunde
left a comment
djsaunde
left a comment
There was a problem hiding this comment.
See inline code comments.
Update contains: