python_code
stringlengths
0
83.2k
# first 3 functions taken from: # http://www.johnvinyard.com/blog/?p=268 import numpy as np from numpy.lib.stride_tricks import as_strided as ast # from .arrays import normalizeMat def norm_shape(shape): ''' Normalize numpy array shapes so they're always expressed as a tuple, even for one-dimensional shapes. Parameters shape - an int, or a tuple of ints Returns a shape tuple ''' try: i = int(shape) return (i,) except TypeError: # shape was not a number pass try: t = tuple(shape) return t except TypeError: # shape was not iterable pass raise TypeError('shape must be an int, or a tuple of ints') def sliding_window(a, ws, ss=None, flatten=True): ''' Return a sliding window over a in any number of dimensions Parameters: a - an n-dimensional numpy array ws - an int (a is 1D) or tuple (a is 2D or greater) representing the size of each dimension of the window ss - an int (a is 1D) or tuple (a is 2D or greater) representing the amount to slide the window in each dimension. If not specified, it defaults to ws. flatten - if True, all slices are flattened, otherwise, there is an extra dimension for each dimension of the input. Returns an array containing each n-dimensional window from a ''' if None is ss: # ss was not provided. the windows will not overlap in any direction. ss = ws ws = norm_shape(ws) ss = norm_shape(ss) # convert ws, ss, and a.shape to numpy arrays so that we can do math in every # dimension at once. ws = np.array(ws) ss = np.array(ss) shape = np.array(a.shape) # ensure that ws, ss, and a.shape all have the same number of dimensions ls = [len(shape), len(ws), len(ss)] if 1 != len(set(ls)): raise ValueError( 'a.shape, ws and ss must all have the same length. They were %s' % str(ls)) # ensure that ws is smaller than a in every dimension if np.any(ws > shape): raise ValueError( 'ws cannot be larger than a in any dimension.' 'a.shape was %s and ws was %s' % (str(a.shape), str(ws))) # how many slices will there be in each dimension? newshape = norm_shape(((shape - ws) // ss) + 1) # the shape of the strided array will be the number of slices in each dimension # plus the shape of the window (tuple addition) newshape += norm_shape(ws) # the strides tuple will be the array's strides multiplied by step size, plus # the array's strides (tuple addition) newstrides = norm_shape(np.array(a.strides) * ss) + a.strides strided = ast(a, shape=newshape, strides=newstrides) if not flatten: return strided # Collapse strided so that it has one more dimension than the window. I.e., # the new array is a flat list of slices. meat = len(ws) if ws.shape else 0 firstdim = (np.product(newshape[:-meat]),) if ws.shape else () dim = firstdim + (newshape[-meat:]) return strided.reshape(dim) def sliding_windows_of_elements(a, ss, ws=None, flatten=False): return [sliding_window(row, ss, ws, flatten) for row in a] def sliding_windows_of_rows(a, ss, ws=None, flatten=True): windowsForRows = sliding_windows_of_elements(a, ss, ws, flatten) return np.vstack(windowsForRows) def _compute_from_seq(allSubseqs, n): seqLens = np.array(map(lambda subseqs: subseqs.shape[0], allSubseqs)) startIdxs = np.r_[0, np.cumsum(seqLens)[:-1]] endIdxs = np.r_[startIdxs[1:], n] fromSeq = np.zeros(n) for i in range(len(startIdxs)): startIdx, endIdx = startIdxs[i], endIdxs[i] fromSeq[startIdx:endIdx] = i return fromSeq # def flattened_subseqs_of_length(seqs, m, norm=None, return_from_seq=False): # # TODO should have flags for returning X and allSubseqs, not just fromSeq # # each element of seqs is assumed to be a 1D or 2D array # origM = m # step = 1 # origDims = len(seqs[0].shape) # if origDims > 1: # sampleDimensions = np.prod(seqs[0].shape[1:]) # num cols in mat # m *= sampleDimensions # TODO don't enforce stepping in only one direction # step *= sampleDimensions # for i, seq in enumerate(seqs): # seqs[i] = seq.flatten() # allSubseqs = sliding_windows_of_elements(seqs, m, step) # X = np.asarray(allSubseqs, dtype=np.float).reshape((-1, m)) # -1 = compute it # Xnorm = normalizeMat(X, origM, how=norm) # if not return_from_seq: # return Xnorm, X, allSubseqs # fromSeq = _compute_from_seq(allSubseqs, Xnorm.shape[0]) # return Xnorm, X, allSubseqs, fromSeq # simple function for common case def sliding_window_1D(x, windowLen, step=1): return sliding_window(x, windowLen, step) class InputTooSmallException(Exception): pass def extract_conv2d_windows( X, filt_shape, strides=(1, 1), flatten_spatial_dims=False, flatten_examples_dim=False, padding='valid'): # TODO support NCHW format orig_X_ndim = X.ndim if X.ndim == 3: X = X[np.newaxis, ...] assert X.ndim == 4 assert len(filt_shape) == 2 assert len(strides) in (2, 4) filt_shape = int(filt_shape[0]), int(filt_shape[1]) if filt_shape[0] > X.shape[1]: # TODO rm after debug raise InputTooSmallException( "filt_shape[0] ({}) > X.shape[1] ({})".format( filt_shape[0], X.shape[1])) if filt_shape[1] > X.shape[2]: raise InputTooSmallException( "filt_shape[1] ({}) > X.shape[2] ({})".format( filt_shape[0], X.shape[2])) padding = padding.lower() assert padding in ('same', 'valid') pad_nrows = filt_shape[0] - 1 pad_ncols = filt_shape[1] - 1 if padding == 'same' and (pad_nrows > 0 or pad_ncols > 0): padded = np.zeros((X.shape[0], X.shape[1] + pad_nrows, X.shape[1] + pad_ncols, X.shape[3])) # NOTE: this should mirror the padding used by scipy and tensorflow; # however, since their exact behavior is only vaguely documented, it # may diverge from their behavior at any time. See the source code for # scipy.signal.convolve2d or https://stackoverflow.com/a/38111069 row_start = int(pad_nrows) // 2 row_end = row_start + X.shape[1] col_start = int(pad_ncols) // 2 col_end = col_start + X.shape[2] # print("padding to shape:", padded.shape) # print("padding: data row start, end:", row_start, row_end) # print("padding: data col start, end:", col_start, col_end) padded[:, row_start:row_end, col_start:col_end, :] = X X = padded filt_shape = (1, filt_shape[0], filt_shape[1], X.shape[3]) if len(strides) == 2: strides = (1, strides[0], strides[1], X.shape[3]) windows = sliding_window(X, filt_shape, strides, flatten=False) # strip out dims 3 and 4, since these are always 1; dim 3 is filter # position across channels (only one position, since doing 2D conv), # and dim 4 is all filter data across examples (not actually # convolving across examples); e.g., for first 200 examples from # MNIST with a 5x5 filter, goes from shape: # (200, 24, 24, 1, 1, 5, 5, 1) # to shape: # (200, 24, 24, 5, 5, 1) windows = windows.reshape(windows.shape[:3] + windows.shape[5:]) if flatten_spatial_dims: # nexamples x npositions x filt_size windows = windows.reshape(X.shape[0], -1, np.prod(filt_shape)) if flatten_examples_dim: windows = windows.reshape(-1, *windows.shape[2:]) if orig_X_ndim == 3: windows = windows.reshape(windows.shape[1:]) return windows if __name__ == '__main__': A = np.arange(24).reshape((6, 4)) print(A) ws = 3 ss = 1 print(sliding_windows_of_rows(A, ws, ss))
#!#!/bin/env/python from __future__ import print_function import numpy as np import torch import torch.nn.functional as F import torch.optim as optim from .utils import kmeans from joblib import Memory _memory = Memory('.', verbose=0) def _to_np(A): return A.cpu().detach().numpy() def _class_balanced_sampling(X, labels, k): np.random.seed(123) N, D = X.shape # intialize centroids by sampling from each class in proportion to its # relative frequency uniq_lbls, counts = np.unique(labels, return_counts=True) sort_idxs = np.argsort(counts) uniq_lbls = uniq_lbls[sort_idxs] counts = counts[sort_idxs] remaining_counts = np.cumsum(counts[::-1])[::-1] nremaining_samples = k # C = np.empty((k, D), dtype=np.float32) C = [] C_labels = [] # affinities = np.zeros((k, nclasses), dtype=np.float32) for i, lbl in enumerate(uniq_lbls): count = counts[i] target_frac = count / remaining_counts[i] target_nsamples = int(nremaining_samples * target_frac + .999) target_nsamples = max(1, target_nsamples) target_nsamples = min(target_nsamples, count) nremaining_samples -= target_nsamples lbl_idxs = np.where(labels == lbl)[0] # print("lbl, count, num lbl idxs: ", lbl, count, len(lbl_idxs)) assert len(lbl_idxs) == count use_idxs = np.random.choice(count, size=target_nsamples, replace=False) keep_idxs = lbl_idxs[use_idxs] C.append(X[keep_idxs]) C_labels.append(np.full(target_nsamples, lbl, dtype=np.int32)) # if len(C).shape[0] < k: C = np.vstack(C).astype(np.float32) # print("k, C shape", k, C.shape) assert C.shape == (k, D) C_labels = np.hstack(C_labels) assert C_labels.shape == (k,) return C, C_labels def neighbor_compression(X, labels, k, niters=1000, rel_tol=.0001, verbose=1): N, D = X.shape # one-hot encode labels nclasses = len(np.unique(labels)) # Y = np.zeros((N, nclasses), dtype=np.float32) # for i in range(N): # Y[i, labels[i]] = 1 # intialize centroids # C, _ = kmeans(X, k) C, C_labels = _class_balanced_sampling(X, labels, k) # convert to torch tensors for optimization # Y = torch.from_numpy(Y) C = torch.tensor(C.T, requires_grad=True) # not from_numpy to allow grad X = torch.from_numpy(X) # having trained class affinities doesn't really seem to help # Z = torch.randn(k, nclasses, requires_grad=True) # print("uniq labels: ", np.unique(labels)) # print("uniq C_labels: ", np.unique(C_labels)) # one-hot encode labels affinities = torch.zeros((k, nclasses), dtype=torch.float32, requires_grad=True) for kk in range(k): affinities[kk, C_labels[kk]] = 1 Z = affinities.clone().detach().requires_grad_(True) labels = torch.from_numpy(labels) loss_fn = torch.nn.CrossEntropyLoss() # opt = optim.SGD([C], lr=.1, momentum=.9) # opt = optim.SGD([C, affinities], lr=.1, momentum=.9) opt = optim.SGD([C, Z], lr=.1, momentum=.9) # X_norms_sq = (X * X).sum(dim=1).view(-1, 1) prev_loss = np.inf for t in range(niters): temperature = np.log2(t + 2) # +2 so that it starts at 1 at t=0 # # compute distances to all centroids # # prods = torch.mm(X, C) # prods = X @ C # # norms_sq = torch.sqrt(torch.sum(C * C)) # # dists_sq = prods - norms_sq # # C_norms_sq = torch.sqrt(torch.sum(C * C, dim=0)) # # C_norms_sq = torch.sum(C * C, dim=0) # # dists_sq = -2 * prods # # dists_sq += X_norms_sq # # dists_sq += C_norms_sq # # neg_dists_sq = -dists_sq # neg_dists_sq = prods # # # update soft labels for each centroid # # similarities = F.softmax(neg_dists_sq, dim=0) # N x C; sim to each sample # # class_affinities = similarities.transpose(0, 1) @ Y # C x nclasses # # class_affinities = F.softmax(class_affinities * temperature, dim=1) # # update class assignments for inputs # # centroid_similarities = F.softmax(neg_dists_sq * temperature, dim=1) # N x C # centroid_similarities = F.softmax(neg_dists_sq, dim=1) # N x C # # centroid_similarities = torch.exp(neg_dists_sq / np.sqrt(D)) # # logits = centroid_similarities @ class_affinities # logits = centroid_similarities @ Z # way simpler version similarities = F.softmax(X @ C, dim=1) # N x C # logits = similarities @ Z affinities = F.softmax(Z * temperature, dim=1) # affinities = F.softmax(affinities * temperature, dim=1) logits = similarities @ affinities # update params and print how we're doing loss = loss_fn(logits, labels) loss.backward() opt.step() opt.zero_grad() loss_pyfloat = loss.item() change = prev_loss - loss_pyfloat thresh = rel_tol * min(loss_pyfloat, prev_loss) if np.abs(change) < thresh: if verbose > 0: _, labels_hat = torch.max(logits, dim=1) acc = torch.mean((labels == labels_hat).type(torch.float)) print("converged after {} iters with acc {:.3f}, loss: {:.4f}" "".format(t + 1, acc.item(), loss_pyfloat)) break # converged prev_loss = loss_pyfloat if (verbose > 1) and ((t + 1) % 10 == 0): _, labels_hat = torch.max(logits, dim=1) acc = torch.mean((labels == labels_hat).type(torch.float)).item() print("acc: ", acc) print("{:.3f}".format(loss.item())) # convert to python float # return _to_np(C).T, _to_np(class_affinities) centroid_labels = np.argmax(_to_np(Z), axis=1) return _to_np(C).T, centroid_labels # or at least, ProtoNN without the L0 constraints; also with simultaneous # updates to all param tensors instead of alternating # def protonn(X, labels, k, niters=10000, verbose=1, gamma=1): def protonn(X, labels, k, d=-1, niters=1000, verbose=1, gamma=-1): N, D = X.shape if gamma < 1: gamma = 1. / np.sqrt(D) # makes it struggle less / not make NaNs # gamma = 1. / D if d < 1: d = D labels = torch.from_numpy(labels) # # one-hot encode labels nclasses = len(np.unique(labels)) # Y = np.zeros((N, nclasses), dtype=np.float32) # for i in range(N): # Y[i, labels[i]] = 1 # intialize centroids C, _ = kmeans(X, k) W = np.random.randn(D, d).astype(np.float32) # C = C @ W # W = np.eye(D).astype(np.float32)[:, :d] # better than randn init # convert to torch tensors for optimization # Y = torch.from_numpy(Y) C = torch.tensor(C.T, requires_grad=True) # not from_numpy to allow grad X = torch.from_numpy(X) W = torch.tensor(W, requires_grad=True) # not from_numpy to allow grad # gamma = torch.tensor(np.array(gamma, dtype=np.float32), requires_grad=True) # labels = torch.from_numpy(labels) # print("W", W[:10]) # return None, None, None Z = torch.randn(k, nclasses, requires_grad=True) loss_fn = torch.nn.CrossEntropyLoss() # opt = optim.SGD([C, Z], lr=.1, momentum=.9) opt = optim.SGD([C, W, Z], lr=.1, momentum=.9) # opt = optim.SGD([C, W, Z, gamma], lr=.1, momentum=.9) nbatches = 1 batch_sz = int(np.ceil(N / nbatches)) # batch_sz = 1024 # nbatches = int(np.ceil(N / batch_sz)) # for t in range(1): for t in range(niters): perm = np.random.permutation(N) for b in range(nbatches): start_idx = b * batch_sz end_idx = min(start_idx + batch_sz, N) perm_idxs = perm[start_idx:end_idx] X_batch = X[perm_idxs] labels_batch = labels[perm_idxs] # temperature = np.log2(t + 2) # +2 so that it starts at 1 at t=0 # compute distances to all centroids # embeddings = X @ W # embeddings = X_batch @ W embeddings = X_batch embed_norms_sq = (embeddings * embeddings).sum(dim=1, keepdim=True) # prods = torch.mm(embeddings, C) prods = embeddings @ C C_norms_sq = torch.sum(C * C, dim=0) dists_sq = -2 * prods dists_sq += embed_norms_sq dists_sq += C_norms_sq neg_dists_sq = -dists_sq # print("gamma: ", gamma) # use_gamma = torch.clamp(gamma, max=1.) # use_gamma = torch.clamp(gamma, 0, 1) # use_gamma = F.sigmoid(gamma) # gamma = torch.min((1, gamma)) # gamma = torch.max((0, gamma)) assert np.min(_to_np(dists_sq)) >= 0 assert np.max(_to_np(neg_dists_sq)) <= 0 similarities = torch.exp(gamma * neg_dists_sq) # N x C # similarities = torch.exp(use_gamma * neg_dists_sq) # N x C logits = similarities @ Z # print("logits shape: ", logits.shape) # print("logits shape: ", logits.shape) # logits_np = _to_np(logits) # print("dists_sq shape", dists_sq.shape) # print("dists_sq", dists_sq[:10]) # print("C_norms_sq", C_norms_sq) # print("embed_norms_sq", embed_norms_sq[:10]) # print("similarities", similarities[:10]) # print("logits", logits[:10]) # update soft labels for each centroid # similarities = F.softmax(neg_dists_sq, dim=0) # N x C; sim to each sample # class_affinities = similarities.transpose(0, 1) @ Y # C x nclasses # class_affinities = F.softmax(class_affinities * temperature, dim=1) # # update class assignments for inputs # centroid_similarities = F.softmax(neg_dists_sq * temperature, dim=1) # N x C # logits = centroid_similarities @ affinities # update params and print how we're doing # loss = loss_fn(logits, labels) loss = loss_fn(logits, labels_batch) # loss += .01 * (gamma * gamma).sum() loss.backward() opt.step() opt.zero_grad() # if (verbose > 0) and (t % 10 == 0): # if (verbose > 0) and ((t + 1) % 10 == 0): if (verbose > 0) and ((t + 1) % 10 == 0) and b == 0: _, labels_hat = torch.max(logits, dim=1) acc = torch.mean((labels[perm_idxs] == labels_hat).type(torch.float)) print("acc: ", acc) print("{:.3f}".format(loss.item())) # convert to python float # print("gamma: ", gamma.item()) return _to_np(C).T, _to_np(W), _to_np(Z) @_memory.cache def stochastic_neighbor_compression(X, labels, k, niters=1000, gamma=-1, rel_tol=.0001, verbose=1): N, D = X.shape nclasses = len(np.unique(labels)) if gamma < 1: gamma = 1 # gamma = 1. / np.sqrt(D) # makes it struggle less / not make NaNs # gamma = 1. / D # labels = torch.from_numpy(labels) # C = np.random.randn(k, D).astype(np.float32) C, C_labels = _class_balanced_sampling(X, labels, k) # one-hot encode labels affinities = torch.zeros((k, nclasses), dtype=torch.float32) for kk in range(k): affinities[kk, C_labels[kk]] = 1 # so that there's actual gradient flow affinities += torch.randn(k, nclasses) * .1 # W = np.random.randn(D, D).astype(np.float32) # C = C @ W # W = np.eye(D).astype(np.float32) # better than randn init # convert to torch tensors for optimization # Y = torch.from_numpy(Y) C = torch.tensor(C.T, requires_grad=True) # not from_numpy to allow grad X = torch.from_numpy(X) labels = torch.from_numpy(labels) gamma = torch.tensor(np.array(gamma, dtype=np.float32)) # affinities = torch.from_numpy(affinities) # print("labels shape: ", labels.shape) # print("uniq labels: ", uniq_lbls) # print("uniq label counts: ", counts) # labels = labels.reshape(-1, 1) # print("labels shape: ", labels.shape) # W = torch.tensor(W, requires_grad=True) # not from_numpy to allow grad # print("W", W[:10]) # return None, None, None # Z = torch.randn(k, nclasses, requires_grad=True) loss_fn = torch.nn.CrossEntropyLoss() opt = optim.SGD([C], lr=.1, momentum=.9) # opt = optim.SGD([C, Z], lr=.1, momentum=.9) # opt = optim.SGD([C, gamma], lr=.1, momentum=.9) nbatches = 1 batch_sz = int(np.ceil(N / nbatches)) # batch_sz = 1024 # nbatches = int(np.ceil(N / batch_sz)) # for t in range(50): prev_loss = np.inf converged = False t = 0 while t < niters and not converged: perm = np.random.permutation(N) for b in range(nbatches): if nbatches > 1: start_idx = b * batch_sz end_idx = min(start_idx + batch_sz, N) perm_idxs = perm[start_idx:end_idx] X_batch = X[perm_idxs] labels_batch = labels[perm_idxs] else: X_batch = X labels_batch = labels # temperature = np.log2(t + 2) # +2 so that it starts at 1 at t=0 # compute distances to all centroids # embeddings = X @ W # embeddings = X_batch @ W embeddings = X_batch embed_norms_sq = (embeddings * embeddings).sum(dim=1, keepdim=True) # prods = torch.mm(embeddings, C) prods = embeddings @ C C_norms_sq = torch.sum(C * C, dim=0) dists_sq = -2 * prods dists_sq += embed_norms_sq dists_sq += C_norms_sq neg_dists_sq = -dists_sq # print("min dist sq: ", torch.min(dists_sq).item()) minval_dist_sq = torch.min(dists_sq).item() if minval_dist_sq < -.01: print("min dist sq: ", minval_dist_sq) print("min C_norms_sq", torch.min(C_norms_sq).item()) print("min X_norms_sq", torch.min(embed_norms_sq).item()) print("dists_sq: ", dists_sq[:10]) assert minval_dist_sq >= -.01 # assert np.min(_to_np(dists_sq)) >= -1e-3 # assert np.max(_to_np(neg_dists_sq)) <= 1e-3 similarities = torch.exp(gamma * neg_dists_sq) # N x C logits = similarities @ affinities # logits = similarities @ Z # print("logits shape: ", logits.shape) # print("logits shape: ", logits.shape) # print("dists_sq shape", dists_sq.shape) # print("dists_sq", dists_sq[:10]) # print("C_norms_sq", C_norms_sq) # print("embed_norms_sq", embed_norms_sq[:10]) # print("similarities", similarities[:10]) # print("logits", logits[:10]) # update params and print how we're doing loss = loss_fn(logits, labels_batch) # loss += gamma * gamma loss.backward() opt.step() opt.zero_grad() loss_pyfloat = loss.item() change = prev_loss - loss_pyfloat thresh = rel_tol * min(loss_pyfloat, prev_loss) if np.abs(change) < thresh: if verbose > 0: _, labels_hat = torch.max(logits, dim=1) labels_true = labels[perm_idxs] if nbatches > 1 else labels acc = torch.mean( (labels_true == labels_hat).type(torch.float)) print("converged after {} iters with acc {:.3f}, loss: {:.4f}" # noqa "".format(t + 1, acc.item(), loss_pyfloat)) converged = True # converged break prev_loss = loss_pyfloat # if (verbose > 0) and ((t + 1) % 10 == 0): # if (verbose > 0) and ((t + 1) % 10 == 0) and b == 0: if (verbose > 1) and (t % 10 == 0) and b == 0: _, labels_hat = torch.max(logits, dim=1) labels_true = labels[perm_idxs] if nbatches > 1 else labels acc = torch.mean( (labels_true == labels_hat).type(torch.float)) print("acc: {:.3f}".format(acc.item())) print("{:.3f}".format(loss.item())) # convert to python float # print("gamma: ", gamma.item()) t += 1 return _to_np(C).T, C_labels def linear_regression_log_loss( X, Y, lamda=1, max_niters=1000, rel_tol=.0001, verbose=2): N, D = X.shape N, M = Y.shape X = X.astype(np.float32) Y = Y.astype(np.float32) # initialize W to OLS solution XtX = X.T @ X XtX += np.eye(D) * np.std(X) XtY = X.T @ Y W = np.linalg.solve(XtX, XtY).astype(np.float32) # W += np.random.randn(*W.shape) X = torch.from_numpy(X) Y = torch.from_numpy(Y) W = torch.tensor(W, requires_grad=True) # W = torch.randn(D, M, requires_grad=True) # W += torch.randn(D, M, requires_grad=False) opt = optim.SGD([W], lr=.1, momentum=.9) # now optimize using pytorch prev_loss = np.inf for t in range(max_niters): Y_hat = X @ W diffs = Y - Y_hat # errs = torch.floor(torch.abs(diffs)) # loss = torch.abs(diffs) # TODO rm # loss = diffs * diffs loss = torch.log2(1 + torch.abs(diffs)) # loss = torch.log2(1e-10 + torch.abs(diffs)) # loss *= (loss > 0).type(torch.float32) loss = torch.mean(loss) # loss = torch.max(loss, 0) loss.backward() opt.step() opt.zero_grad() loss_pyfloat = loss.item() change = prev_loss - loss_pyfloat thresh = rel_tol * min(loss_pyfloat, prev_loss) if np.abs(change) < thresh: if verbose > 0: print("converged after {} iters with loss: {:.4f}".format( t + 1, loss_pyfloat)) break # converged prev_loss = loss_pyfloat if (verbose > 1) and ((t + 1) % 10 == 0): print("loss: {:.4f}".format(loss_pyfloat)) return _to_np(W) def main(): # N, D = 10000, 20 N, D = 1000, 20 # niters = 1000 niters = 10000 X = np.random.randn(N, D).astype(np.float32) # ------------------------ linear regression with weird loss # M = 20 # Y = np.random.randn(N, M).astype(np.float32) # linear_regression_log_loss(X, Y) # ------------------------ neighbor compression K = 16 nclasses = 5 # labels = torch.randint(nclasses, size=(N,)) # labels = _to_np(torch.randint(nclasses, size=(N,))) labels = np.random.randint(nclasses, size=(N,)) # C, W, Z = protonn(X, labels, K, niters=niters) # significantly worse C, centroid_labels = stochastic_neighbor_compression(X, labels, K, niters=niters) # C, centroid_labels = neighbor_compression(X, labels, K, niters=niters) print("centroid_labels:", centroid_labels) print("C type, shape", type(C), C.shape) print("done") if __name__ == '__main__': main()
#!/usr/bin/env python import os import numpy as np import pandas as pd # TODO this file is hideous (but necessarily so for deadline purposes...) # # Also, this file is tightly coupled to figs.py; it basically has a func # for each figure func that spits out data in exactly the required form MCQ_RESULTS_DIR = '../results/timing/' MATMUL_RESULTS_DIR = '../results/matmul/' def get_mcq_path(D, nbytes): fname = 'mcq_D={}_M={}.txt'.format(D, nbytes) return os.path.join(MCQ_RESULTS_DIR, fname) class McqResults(object): def __init__(self, path=None, D=None, nbytes=None): if path is None: path = get_mcq_path(D=D, nbytes=nbytes) self.path = path with open(self.path, 'r') as f: self.lines = f.readlines() self.stats = {line.split(':')[0].strip(): line.split(':')[1].strip() for line in self.lines if ':' in line} self.bolt_nbytes = int(self.stats['bolt M']) self.pq_nbytes = int(self.stats['pq M']) self.bolt_D = int(self.stats['bolt subvect_len']) * self.bolt_nbytes * 2 self.pq_D = int(self.stats['pq subvect_len']) * self.pq_nbytes assert self.bolt_nbytes == self.pq_nbytes assert self.bolt_D == self.pq_D self.nbytes = self.bolt_nbytes self.D = self.bolt_D # check that file was named properly expected_path = get_mcq_path(D=self.D, nbytes=self.nbytes) if expected_path != path: print("expected path, path = ", expected_path, path) assert expected_path == path def __str__(self): # for debugging s = "" sorted_keys = sorted(self.stats.keys()) for k in sorted_keys: v = self.stats[k] s += "'{}': '{}'\n".format(k, v) return s def _extract_thruput(profile_str): result_strs = profile_str.split(':')[-1] rep_strs = result_strs.strip(' ,').split(',') thruput_parens = [s.strip(' ').split(' ')[1] for s in rep_strs] return np.array([int(s.strip('()s/')) for s in thruput_parens]) def _extract_times(profile_str): result_strs = profile_str.split(':')[-1] rep_strs = result_strs.strip(' ,').split(',') time_strs = [s.strip(' ').split(' ')[0] for s in rep_strs] return np.array([float(s) for s in time_strs]) def popcount_results_256(): LENGTH = 256 popcnt_times = {} popcnt_times[8] = '2.456 (1302931596/s), 2.344 (1365187713/s), 2.125 (1505882352/s), 2.829 (1131141746/s), 2.148 (1489757914/s), 2.167 (1476695892/s), 2.327 (1375161151/s), 2.145 (1491841491/s), 2.12 (1509433962/s), 2.112 (1515151515/s)' popcnt_times[16] = '4.368 (732600732/s), 4.121 (776510555/s), 3.926 (815078960/s), 4.105 (779537149/s), 4.176 (766283524/s), 4.119 (776887594/s), 4.464 (716845878/s), 4.153 (770527329/s), 4.364 (733272227/s), 4.198 (762267746/s)' popcnt_times[32] = '7.612 (420388859/s), 7.347 (435551925/s), 7.694 (415908500/s), 9.122 (350800263/s), 7.343 (435789186/s), 9.344 (342465753/s), 8.148 (392734413/s), 9.046 (353747512/s), 8.455 (378474275/s), 7.685 (416395575/s)' bolt_times = {} bolt_times[8] = '0.461 (2169197396/s), 0.456 (2192982456/s), 0.539 (1855287569/s), 0.53 (1886792452/s), 0.456 (2192982456/s), 0.452 (2212389380/s), 0.442 (2262443438/s), 0.438 (2283105022/s), 0.434 (2304147465/s), 0.547 (1828153564/s)' bolt_times[16] = '0.894 (1118568232/s), 1.08 (925925925/s), 0.88 (1136363636/s), 0.877 (1140250855/s), 0.881 (1135073779/s), 0.847 (1180637544/s), 1.011 (989119683/s), 0.866 (1154734411/s), 0.984 (1016260162/s), 0.838 (1193317422/s)' bolt_times[32] = '2.047 (488519785/s), 1.726 (579374275/s), 1.924 (519750519/s), 2.085 (479616306/s), 2.076 (481695568/s), 1.748 (572082379/s), 1.757 (569151963/s), 2.064 (484496124/s), 1.742 (574052812/s), 1.725 (579710144/s)' out_dicts = [] algos = ['Bolt', 'Binary Embedding'] dicts = [bolt_times, popcnt_times] for algo, d in zip(algos, dicts): for nbytes, s in list(d.items()): thruputs = _extract_thruput(s) out_dicts += [{'algo': algo, 'nbytes': nbytes, 'length': LENGTH, 'trial': i, 'y': t} for i, t in enumerate(thruputs)] return pd.DataFrame.from_records(out_dicts) def encode_results(): dicts = [] for D in [64, 128, 256, 512, 1024]: for nbytes in [8, 16, 32]: res = McqResults(D=D, nbytes=nbytes) abbrevs = ['bolt', 'pq', 'opq'] names = ['Bolt', 'PQ', 'OPQ'] for abbrev, name in zip(abbrevs, names): # results for encoding data key = abbrev + ' encode (10x5)' thruputs = _extract_thruput(res.stats[key]) dicts += [{'task': 'encode_x', 'D': D, 'nbytes': nbytes, 'algo': name, 'trial': i, 'y': t} for i, t in enumerate(thruputs)] # results for encoding query if abbrev == 'bolt': key = abbrev + ' encode lut (10x5)' else: key = abbrev + ' encode lut float dist (10x5)' thruputs = _extract_thruput(res.stats[key]) dicts += [{'task': 'encode_q', 'D': D, 'nbytes': nbytes, 'algo': name, 'trial': i, 'y': t} for i, t in enumerate(thruputs)] return pd.DataFrame.from_records(dicts) def matmul_results(which='square'): if which == 'square': SIZES = [64, 128, 256, 512, 1024, 4096, 8192] data_fname = 'square_matmul_results.txt' elif which == 'tall': SIZES = [32, 64, 128, 256, 512, 1024] data_fname = 'tall_matmul_results.txt' with open(MATMUL_RESULTS_DIR + data_fname) as f: lines = f.readlines() stats = {line.split(':')[0].strip(): line.split(':')[1].strip() for line in lines if ':' in line} dicts = [] # add in results from bolt for nbytes in [8, 16, 32]: prefix = 'bolt<{}>'.format(nbytes) algo = 'Bolt {}B'.format(nbytes) for sz in SIZES: for enc in (0, 1): # don't vs do encode X at start key = '{} encode={} matmul {} (10x5)'.format(prefix, enc, sz) times = _extract_times(stats[key]) dicts += [{'algo': algo, 'size': sz, 'enc': enc, 'nbytes': nbytes, 'trial': i, 'y': t} for i, t in enumerate(times)] # also add in "encode" version of bolt if enc: enc_algo_name = algo + ' + Encode' dicts += [{'algo': enc_algo_name, 'size': sz, 'enc': enc, 'nbytes': nbytes, 'trial': i, 'y': t} for i, t in enumerate(times)] # add in matmul results for sz in SIZES: key = 'matmul {} (10x5)'.format(sz) times = _extract_times(stats[key]) dicts += [{'algo': 'Floats', 'size': sz, 'enc': -1, 'trial': i, 'y': t} for i, t in enumerate(times)] return pd.DataFrame.from_records(dicts) def encode_data_results_256(): LENGTH = 256 pq_times = {} pq_times[8] = 'pq encode (10x5): 6.696 (149342/s), 6.688 (149521/s), 6.639 (150625/s), 6.648 (150421/s), 6.711 (149009/s), 6.67 (149925/s), 6.634 (150738/s), 6.684 (149611/s), 6.663 (150082/s), 6.67 (149925/s),' pq_times[16] = 'pq encode (10x5): 7.181 (139256/s), 7.194 (139004/s), 7.179 (139295/s), 7.146 (139938/s), 7.123 (140390/s), 7.123 (140390/s), 7.162 (139625/s), 7.148 (139899/s), 7.116 (140528/s), 7.193 (139024/s),' pq_times[32] = 'pq encode (10x5): 8.089 (123624/s), 8.175 (122324/s), 8.117 (123198/s), 8.096 (123517/s), 8.48 (117924/s), 8.071 (123900/s), 8.126 (123061/s), 8.123 (123107/s), 8.069 (123931/s), 8.21 (121802/s),' opq_times = {} opq_times[8] = 'opq encode (10x5): 8.441 (118469/s), 8.385 (119260/s), 8.368 (119502/s), 8.39 (119189/s), 8.355 (119688/s), 8.388 (119217/s), 8.383 (119289/s), 8.412 (118877/s), 8.401 (119033/s), 8.391 (119175/s),' opq_times[16] = 'opq encode (10x5): 8.88 (112612/s), 8.786 (113817/s), 8.874 (112688/s), 8.834 (113199/s), 8.874 (112688/s), 8.902 (112334/s), 8.899 (112372/s), 8.925 (112044/s), 8.867 (112777/s), 8.907 (112271/s),' opq_times[32] = 'opq encode (10x5): 9.761 (102448/s), 9.718 (102901/s), 9.717 (102912/s), 9.726 (102817/s), 9.908 (100928/s), 9.796 (102082/s), 10.164 (98386/s), 9.792 (102124/s), 9.735 (102722/s), 9.729 (102785/s),' bolt_times = {} bolt_times[8] = 'bolt encode (10x5): 3.43 (2915451/s), 3.586 (2788622/s), 3.421 (2923121/s), 3.408 (2934272/s), 3.409 (2933411/s), 3.406 (2935995/s), 3.407 (2935133/s), 3.412 (2930832/s), 3.411 (2931691/s), 3.409 (2933411/s),' bolt_times[16] = 'bolt encode (10x5): 3.93 (2544529/s), 3.687 (2712232/s), 3.826 (2613695/s), 4.007 (2495632/s), 3.705 (2699055/s), 3.976 (2515090/s), 3.709 (2696144/s), 3.681 (2716653/s), 3.693 (2707825/s), 3.802 (2630194/s),' bolt_times[32] = 'bolt encode (10x5): 5.039 (1984520/s), 4.591 (2178174/s), 5.081 (1968116/s), 4.697 (2129018/s), 4.591 (2178174/s), 4.763 (2099517/s), 4.832 (2069536/s), 4.805 (2081165/s), 4.961 (2015722/s), 4.665 (2143622/s),' out_dicts = [] algos = ['Bolt', 'PQ', 'OPQ'] dicts = [bolt_times, pq_times, opq_times] for algo, d in zip(algos, dicts): for nbytes, s in list(d.items()): thruputs = _extract_thruput(s) out_dicts += [{'algo': algo, 'nbytes': nbytes, 'length': LENGTH, 'trial': i, 'y': t} for i, t in enumerate(thruputs)] return pd.DataFrame.from_records(out_dicts) def encode_lut_results(): pq_times = {} pq_times[8] = 'pq encode lut float dist (10x5): 64.986 (153879/s), 65.014 (153813/s), 65.155 (153480/s), 64.808 (154301/s), 66.593 (150165/s), 67.68 (147754/s), 69.399 (144094/s), 66.702 (149920/s), 66.234 (150979/s), 66.286 (150861/s),' pq_times[16] = 'pq encode lut float dist (10x5): 67.893 (147290/s), 67.484 (148183/s), 69.608 (143661/s), 68.083 (146879/s), 70.958 (140928/s), 69.423 (144044/s), 72.129 (138640/s), 74.984 (133361/s), 70.837 (141169/s), 74.967 (133392/s),' pq_times[32] = 'pq encode lut float dist (10x5): 78.809 (126889/s), 79.34 (126039/s), 78.565 (127283/s), 79.171 (126308/s), 78.372 (127596/s), 78.689 (127082/s), 78.094 (128050/s), 80.031 (124951/s), 93.367 (107104/s), 81.896 (122106/s),' opq_times = {} opq_times[8] = 'opq encode lut float dist (10x5): 155.68 (64234/s), 159.49 (62698/s), 160.64 (62249/s), 158.21 (63205/s), 159.37 (62747/s), 159.29 (62778/s), 160.81 (62186/s), 158.5 (63090/s), 155.22 (64423/s), 158.98 (62901/s),' opq_times[16] = 'opq encode lut float dist (10x5): 170.42 (58677/s), 168.41 (59380/s), 169.12 (59129/s), 171.53 (58298/s), 167.32 (59766/s), 168.96 (59185/s), 170.43 (58676/s), 170.7 (58581/s), 169.86 (58870/s), 160.43 (62333/s),' opq_times[32] = 'opq encode lut float dist (10x5): 170.86 (58527/s), 175.79 (56885/s), 169.86 (58870/s), 180.3 (55464/s), 172.46 (57983/s), 171.66 (58254/s), 167.23 (59799/s), 168.19 (59457/s), 164.47 (60801/s), 168.31 (59413/s),' bolt_times = {} bolt_times[8] = 'bolt encode lut (10x5): 2.907 (3439972/s), 2.911 (3435245/s), 2.902 (3445899/s), 2.899 (3449465/s), 2.907 (3439972/s), 2.908 (3438789/s), 2.908 (3438789/s), 2.906 (3441156/s), 2.906 (3441156/s), 2.908 (3438789/s),' bolt_times[16] = 'bolt encode lut (10x5): 2.957 (3381805/s), 2.953 (3386386/s), 2.957 (3381805/s), 2.943 (3397893/s), 2.949 (3390979/s), 2.95 (3389830/s), 2.946 (3394433/s), 3.103 (3222687/s), 2.944 (3396739/s), 3.029 (3301419/s),' bolt_times[32] = 'bolt encode lut (10x5): 2.511 (3982477/s), 2.51 (3984063/s), 2.587 (3865481/s), 2.508 (3987240/s), 2.847 (3512469/s), 2.508 (3987240/s), 2.508 (3987240/s), 2.769 (3611412/s), 2.729 (3664345/s), 2.556 (3912363/s),' out_dicts = [] algos = ['Bolt', 'PQ', 'OPQ'] dicts = [bolt_times, pq_times, opq_times] for algo, d in zip(algos, dicts): for nbytes, s in list(d.items()): thruputs = _extract_thruput(s) out_dicts += [{'algo': algo, 'nbytes': nbytes, 'y': t} for t in thruputs] return pd.DataFrame.from_records(out_dicts) def query_speed_results(): # NOTE: all thruputs in this function (except matmul ones) need be # multiplied by 100,000 because we're reporting distances/sec, not time # to query 100k points bolt_times = {} bolt_times[8] = '4.385 (22805/s), 4.385 (22805/s), 4.408 (22686/s), 4.385 (22805/s), 5.117 (19542/s), 4.378 (22841/s), 4.392 (22768/s), 4.393 (22763/s), 4.381 (22825/s), 4.383 (22815/s)' bolt_times[16] = '8.268 (12094/s), 9.807 (10196/s), 8.389 (11920/s), 8.681 (11519/s), 8.711 (11479/s), 8.293 (12058/s), 9.797 (10207/s), 8.32 (12019/s), 9.767 (10238/s), 9.499 (10527/s)' bolt_times[32] = '19.385 (5158/s), 17.215 (5808/s), 18.612 (5372/s), 18.117 (5519/s), 17.323 (5772/s), 18.436 (5424/s), 18.979 (5268/s), 16.274 (6144/s), 19.696 (5077/s), 17.026 (5873/s)' popcnt_times = {} popcnt_times[8] = '2.456 (1302931596/s), 2.344 (1365187713/s), 2.125 (1505882352/s), 2.829 (1131141746/s), 2.148 (1489757914/s), 2.167 (1476695892/s), 2.327 (1375161151/s), 2.145 (1491841491/s), 2.12 (1509433962/s), 2.112 (1515151515/s)' popcnt_times[16] = '4.368 (732600732/s), 4.121 (776510555/s), 3.926 (815078960/s), 4.105 (779537149/s), 4.176 (766283524/s), 4.119 (776887594/s), 4.464 (716845878/s), 4.153 (770527329/s), 4.364 (733272227/s), 4.198 (762267746/s)' popcnt_times[32] = '7.612 (420388859/s), 7.347 (435551925/s), 7.694 (415908500/s), 9.122 (350800263/s), 7.343 (435789186/s), 9.344 (342465753/s), 8.148 (392734413/s), 9.046 (353747512/s), 8.455 (378474275/s), 7.685 (416395575/s)' pq_times = {} pq_times[8] = '36.499 (2739/s), 35.729 (2798/s), 36.521 (2738/s), 37.924 (2636/s), 37.079 (2696/s), 36.444 (2743/s), 36.115 (2768/s), 36.955 (2705/s), 35.913 (2784/s), 40.354 (2478/s)' pq_times[16] = '79.482 (1258/s), 82.546 (1211/s), 84.992 (1176/s), 84.996 (1176/s), 86.218 (1159/s), 84.495 (1183/s), 90.637 (1103/s), 82.164 (1217/s), 85.954 (1163/s), 82.255 (1215/s)' pq_times[32] = '214.85 (465/s), 217.41 (459/s), 212.49 (470/s), 210.75 (474/s), 211.12 (473/s), 212.54 (470/s), 209.91 (476/s), 219.95 (454/s), 212.97 (469/s), 213.44 (468/s)' opq_times = {} opq_times[8] = '38.653 (2587/s), 36.958 (2705/s), 37.684 (2653/s), 35.902 (2785/s), 38.032 (2629/s), 39.511 (2530/s), 42.321 (2362/s), 38.94 (2568/s), 39.224 (2549/s), 39.06 (2560/s)' opq_times[16] = '82.636 (1210/s), 82.401 (1213/s), 88.424 (1130/s), 86.649 (1154/s), 83.329 (1200/s), 82.719 (1208/s), 82.281 (1215/s), 80.581 (1240/s), 80.777 (1237/s), 81.107 (1232/s)' opq_times[32] = '221.61 (451/s), 230.01 (434/s), 241.68 (413/s), 222.39 (449/s), 215.13 (464/s), 215.49 (464/s), 212.27 (471/s), 213.95 (467/s), 213.96 (467/s), 217.79 (459/s)' # 1, 16 -> rowmajor times; 64, 256, 1024 -> colmajor times; (ie, use times from best layout) matmul1_times = '12.063 (8289811/s), 11.231 (8903926/s), 10.283 (9724788/s), 10.864 (9204712/s), 10.492 (9531071/s), 10.877 (9193711/s), 10.79 (9267840/s), 10.85 (9216589/s), 11.041 (9057150/s), 10.647 (9392317/s)' matmul16_times = '21.707 (73708941/s), 21.38 (74836295/s), 21.71 (73698756/s), 21.54 (74280408/s), 21.454 (74578167/s), 21.989 (72763654/s), 22.486 (71155385/s), 22.048 (72568940/s), 23.18 (69025021/s), 21.771 (73492260/s)' matmul64_times = '56.496 (113282356/s), 55.488 (115340253/s), 54.853 (116675478/s), 56.689 (112896681/s), 56.482 (113310435/s), 55.644 (115016893/s), 54.623 (117166761/s), 55.773 (114750865/s), 54.726 (116946241/s), 54.918 (116537383/s)' matmul256_times = '164.72 (155414306/s), 168.41 (152014488/s), 169.93 (150652927/s), 164.99 (155157157/s), 166.66 (153609831/s), 163.04 (157012830/s), 167.45 (152880544/s), 161.06 (158949936/s), 171.13 (149594750/s), 168.49 (151940505/s)' matmul1024_times = '653.63 (156664035/s), 677.26 (151197248/s), 692.88 (147788938/s), 664.79 (154032909/s), 702.61 (145742096/s), 651.74 (157116904/s), 656.4 (156003388/s), 664.69 (154056314/s), 665.34 (153906736/s), 651.88 (157083643/s)' out_dicts = [] algos = ['Bolt', 'PQ', 'OPQ', 'Binary Embedding'] dicts = [bolt_times, pq_times, opq_times, popcnt_times] for algo, d in zip(algos, dicts): for nbytes, s in list(d.items()): thruputs = _extract_thruput(s) * 1e5 if algo == 'Binary Embedding': thruputs /= 1e5 # these are already dists/sec, not qps out_dicts += [{'algo': algo, 'nbytes': nbytes, 'y': t} for t in thruputs] matmul_strs = [matmul1_times, matmul16_times, matmul64_times, matmul256_times, matmul1024_times] batch_sizes = [1, 16, 64, 256, 1024] nbytes_list = [8, 16, 32] # replicate results in each plot for s, sz in zip(matmul_strs, batch_sizes): algo = 'Matmul {}'.format(sz) for nbytes in nbytes_list: thruputs = _extract_thruput(s) out_dicts += [{'algo': algo, 'nbytes': nbytes, 'y': t} for t in thruputs] return pd.DataFrame.from_records(out_dicts) def main(): pass # print _extract_thruput('foo (10x5): 2.456 (1302931596/s), 2.344 (1365187713/s), 2.125 (1505882352/s), 2.829 (1131141746/s), 2.148 (1489757914/s), 2.167 (1476695892/s), 2.327 (1375161151/s), 2.145 (1491841491/s), 2.12 (1509433962/s), 2.112 (1515151515/s)') # print McqResults('../results/tmp.txt') # print McqResults('../results/mcq/mcq_D=256_M=8.txt') # res = query_speed_results() # print res.loc[res['algo'] == 'Matmul 1'] # print res.loc[res['algo'] == 'Matmul 256'] if __name__ == '__main__': main()
#!/bin/env/python import functools import numpy as np import pprint import scipy import time from . import amm from . import matmul_datasets as md from . import pyience as pyn from . import compress from . import amm_methods as methods from joblib import Memory _memory = Memory('.', verbose=0) # NUM_TRIALS = 1 NUM_TRIALS = 10 # @_memory.cache def _estimator_for_method_id(method_id, **method_hparams): return methods.METHOD_TO_ESTIMATOR[method_id](**method_hparams) def _hparams_for_method(method_id): if method_id in methods.SKETCH_METHODS: # dvals = [2, 4, 6, 8, 12, 16, 24, 32, 48, 64] # d=1 undef on fd methods # dvals = [1, 2, 4, 8, 16, 32, 64, 128] dvals = [1, 2, 4, 8, 16, 32, 64] # dvals = [1, 2, 4, 8, 16, 32] # dvals = [1, 2, 4, 8] # dvals = [32] # TODO rm after debug # dvals = [16] # TODO rm after debug # dvals = [8] # TODO rm after debug # dvals = [4] # TODO rm after debug # dvals = [3] # TODO rm after debug # dvals = [2] # TODO rm after debug # dvals = [1] # TODO rm after debug if method_id == methods.METHOD_SPARSE_PCA: # first one gets it to not return all zeros on caltech alpha_vals = (1. / 16384, .03125, .0625, .125, .25, .5, 1, 2, 4, 8) # alpha_vals = (.0625, .125, .25, .5, 1, 2, 4, 8) # alpha_vals = (.0625, .125) # alpha_vals = [.0625] # TODO rm # alpha_vals = [.03125] # TODO rm # alpha_vals = [1./1024] # TODO rm # alpha_vals = [1./16384] # TODO rm # alpha_vals = [0] # TODO rm # alpha_vals = (2, 4, 5) # alpha_vals = [.1] # alpha_vals = [1.] # alpha_vals = [10.] # alpha_vals = [20.] # alpha_vals = [50.] return [{'d': d, 'alpha': alpha} for d in dvals for alpha in alpha_vals] return [{'d': dval} for dval in dvals] if method_id in methods.VQ_METHODS: # mvals = [1, 2, 4, 8, 16, 32, 64] mvals = [2, 4, 8, 16, 32, 64] # mvals = [64] # mvals = [1, 2, 4, 8, 16] # mvals = [1, 2, 4, 8] # mvals = [8, 16] # TODO rm after debug # mvals = [8, 16, 64] # TODO rm after debug # mvals = [128] # TODO rm after debug # mvals = [64] # TODO rm after debug # mvals = [32] # TODO rm after debug # mvals = [16] # TODO rm after debug # mvals = [8] # TODO rm after debug # mvals = [4] # TODO rm after debug # mvals = [1] # TODO rm after debug if method_id == methods.METHOD_MITHRAL: lut_work_consts = (2, 4, -1) # lut_work_consts = [-1] # TODO rm params = [] for m in mvals: for const in lut_work_consts: params.append({'ncodebooks': m, 'lut_work_const': const}) return params return [{'ncodebooks': m} for m in mvals] if method_id in [methods.METHOD_EXACT, methods.METHOD_SCALAR_QUANTIZE]: return [{}] raise ValueError(f"Unrecognized method: '{method_id}'") def _ntrials_for_method(method_id, ntasks): # return 1 # TODO rm if ntasks > 1: # no need to avg over trials if avging over multiple tasks return 1 # return NUM_TRIALS if method_id in methods.NONDETERMINISTIC_METHODS else 1 return NUM_TRIALS if method_id in methods.RANDOM_SKETCHING_METHODS else 1 # ================================================================ metrics def _compute_compression_metrics(ar): # if quantize_to_type is not None: # ar = ar.astype(quantize_to_type) # ar -= np.min(ar) # ar /= (np.max(ar) / 65535) # 16 bits # ar -= 32768 # center at 0 # ar = ar.astype(np.int16) # elem_sz = ar.dtype.itemsize # return {'nbytes_raw': ar.nbytes, # 'nbytes_blosc_noshuf': len(_blosc_compress( # ar, elem_sz=elem_sz, shuffle=blosc.NOSHUFFLE)), # 'nbytes_blosc_byteshuf': len(_blosc_compress( # ar, elem_sz=elem_sz, shuffle=blosc.SHUFFLE)), # 'nbytes_blosc_bitshuf': len(_blosc_compress( # ar, elem_sz=elem_sz, shuffle=blosc.BITSHUFFLE)), # 'nbytes_zstd': len(_zstd_compress(ar)), # 'nbits_cost': nbits_cost(ar).sum() // 8, # 'nbits_cost_zigzag': # nbits_cost(zigzag_encode(ar), signed=False).sum() // 8, # 'nbytes_sprintz': compress.sprintz_packed_size(ar) # } return {'nbytes_raw': ar.nbytes, 'nbytes_sprintz': compress.sprintz_packed_size(ar)} def _cossim(Y, Y_hat): ynorm = np.linalg.norm(Y) + 1e-20 yhat_norm = np.linalg.norm(Y_hat) + 1e-20 return ((Y / ynorm) * (Y_hat / yhat_norm)).sum() def _compute_metrics(task, Y_hat, compression_metrics=True, **sink): Y = task.Y_test diffs = Y - Y_hat raw_mse = np.mean(diffs * diffs) normalized_mse = raw_mse / np.var(Y) # Y_meannorm = Y - Y.mean() # Y_hat_meannorm = Y_hat - Y_hat.mean() # ynorm = np.linalg.norm(Y_meannorm) + 1e-20 # yhat_norm = np.linalg.norm(Y_hat_meannorm) + 1e-20 # r = ((Y_meannorm / ynorm) * (Y_hat_meannorm / yhat_norm)).sum() metrics = {'raw_mse': raw_mse, 'normalized_mse': normalized_mse, 'corr': _cossim(Y - Y.mean(), Y_hat - Y_hat.mean()), 'cossim': _cossim(Y, Y_hat), # 'bias': diffs.mean(), 'y_mean': Y.mean(), 'y_std': Y.std(), 'yhat_std': Y_hat.std(), 'yhat_mean': Y_hat.mean()} if compression_metrics: # Y_q = compress.quantize(Y, nbits=8) # Y_hat_q = compress.quantize(Y_hat, nbits=8) # diffs_q = Y_q - Y_hat_q # # diffs_q = compress.zigzag_encode(diffs_q).astype(np.uint8) # assert Y_q.dtype == np.int8 # assert diffs_q.dtype == np.int8 Y_q = compress.quantize(Y, nbits=12) Y_hat_q = compress.quantize(Y_hat, nbits=12) diffs_q = Y_q - Y_hat_q assert Y_q.dtype == np.int16 assert diffs_q.dtype == np.int16 # Y_q = quantize_i16(Y) # # quantize to 16 bits # Y = Y - np.min(Y) # Y /= (np.max(Y) / 65535) # 16 bits # Y -= 32768 # center at 0 # Y = Y.astype(np.int16) # diffs = metrics_raw = _compute_compression_metrics(Y_q) metrics.update({k + '_orig': v for k, v in metrics_raw.items()}) metrics_raw = _compute_compression_metrics(diffs_q) metrics.update({k + '_diffs': v for k, v in metrics_raw.items()}) if task.info: problem = task.info['problem'] metrics['problem'] = problem if problem == 'softmax': lbls = task.info['lbls_test'].astype(np.int32) b = task.info['biases'] logits_amm = Y_hat + b logits_orig = Y + b lbls_amm = np.argmax(logits_amm, axis=1).astype(np.int32) lbls_orig = np.argmax(logits_orig, axis=1).astype(np.int32) # print("Y_hat shape : ", Y_hat.shape) # print("lbls hat shape: ", lbls_amm.shape) # print("lbls amm : ", lbls_amm[:20]) metrics['acc_amm'] = np.mean(lbls_amm == lbls) metrics['acc_orig'] = np.mean(lbls_orig == lbls) elif problem in ('1nn', 'rbf'): lbls = task.info['lbls_test'].astype(np.int32) lbls_centroids = task.info['lbls_centroids'] lbls_hat_1nn = [] rbf_lbls_hat = [] W = task.W_test centroid_norms_sq = (W * W).sum(axis=0) sample_norms_sq = (task.X_test * task.X_test).sum( axis=1, keepdims=True) k = W.shape[1] nclasses = np.max(lbls_centroids) + 1 affinities = np.zeros((k, nclasses), dtype=np.float32) for kk in range(k): affinities[kk, lbls_centroids[kk]] = 1 for prods in [Y_hat, Y]: dists_sq_hat = (-2 * prods) + centroid_norms_sq + sample_norms_sq # 1nn classification centroid_idx = np.argmin(dists_sq_hat, axis=1) lbls_hat_1nn.append(lbls_centroids[centroid_idx]) # rbf kernel classification (bandwidth=1) # gamma = 1. / np.sqrt(W.shape[0]) # gamma = 1. / W.shape[0] gamma = 1 similarities = scipy.special.softmax(-dists_sq_hat * gamma, axis=1) class_probs = similarities @ affinities rbf_lbls_hat.append(np.argmax(class_probs, axis=1)) lbls_amm_1nn, lbls_orig_1nn = lbls_hat_1nn rbf_lbls_amm, rbf_lbls_orig = rbf_lbls_hat metrics['acc_amm_1nn'] = np.mean(lbls_amm_1nn == lbls) metrics['acc_orig_1nn'] = np.mean(lbls_orig_1nn == lbls) metrics['acc_amm_rbf'] = np.mean(rbf_lbls_amm == lbls) metrics['acc_orig_rbf'] = np.mean(rbf_lbls_orig == lbls) if problem == '1nn': lbls_amm, lbls_orig = rbf_lbls_amm, rbf_lbls_orig elif problem == 'rbf': lbls_amm, lbls_orig = rbf_lbls_amm, rbf_lbls_orig orig_acc_key = 'acc-1nn-raw' if orig_acc_key in task.info: metrics[orig_acc_key] = task.info[orig_acc_key] metrics['acc_amm'] = np.mean(lbls_amm == lbls) metrics['acc_orig'] = np.mean(lbls_orig == lbls) elif problem == 'sobel': assert Y.shape[1] == 2 grad_mags_true = np.sqrt((Y * Y).sum(axis=1)) grad_mags_hat = np.sqrt((Y_hat * Y_hat).sum(axis=1)) diffs = grad_mags_true - grad_mags_hat metrics['grad_mags_nmse'] = ( (diffs * diffs).mean() / grad_mags_true.var()) elif problem.lower().startswith('dog'): # difference of gaussians assert Y.shape[1] == 2 Z = Y[:, 0] - Y[:, 1] Z_hat = Y_hat[:, 0] - Y_hat[:, 1] diffs = Z - Z_hat metrics['dog_nmse'] = (diffs * diffs).mean() / Z.var() return metrics # ================================================================ driver funcs def _eval_amm(task, est, fixedB=True, **metrics_kwargs): est.reset_for_new_task() if fixedB: est.set_B(task.W_test) # print("eval_amm validating task: ", task.name) # task.validate(train=False, test=True) # print(f"task {task.name} matrix hashes:") # pprint.pprint(task._hashes()) # print("task: ", task.name) # print("X_test shape: ", task.X_test.shape) # print("W_test shape: ", task.W_test.shape) t = time.perf_counter() # Y_hat = est.predict(task.X_test.copy(), task.W_test.copy()) Y_hat = est.predict(task.X_test, task.W_test) # Y_hat = task.X_test @ task.W_test # yep, zero error duration_secs = time.perf_counter() - t metrics = _compute_metrics(task, Y_hat, **metrics_kwargs) metrics['secs'] = duration_secs # metrics['nmultiplies'] = est.get_nmuls(task.X_test, task.W_test) metrics.update(est.get_speed_metrics( task.X_test, task.W_test, fixedB=fixedB)) # print("eval_amm re-validating task: ", task.name) # task.validate(train=False, test=True) # print(f"task {task.name} matrix hashes:") # pprint.pprint(task.hashes()) return metrics def _get_all_independent_vars(): independent_vars = set(['task_id', 'method', 'trial']) for method_id in methods.ALL_METHODS: hparams = _hparams_for_method(method_id)[0] est = _estimator_for_method_id(method_id, **hparams) independent_vars = (independent_vars | set(est.get_params().keys())) return independent_vars # @functools.lru_cache(maxsize=None) # @_memory.cache def _fitted_est_for_hparams(method_id, hparams_dict, X_train, W_train, Y_train, **kwargs): est = _estimator_for_method_id(method_id, **hparams_dict) est.fit(X_train, W_train, Y=Y_train, **kwargs) return est # def _main(tasks, methods=['SVD'], saveas=None, ntasks=None, def _main(tasks_func, methods=None, saveas=None, ntasks=None, verbose=1, limit_ntasks=-1, compression_metrics=False, # TODO uncomment below # verbose=3, limit_ntasks=-1, compression_metrics=False, tasks_all_same_shape=False): methods = methods.DEFAULT_METHODS if methods is None else methods if isinstance(methods, str): methods = [methods] if limit_ntasks is None or limit_ntasks < 1: limit_ntasks = np.inf independent_vars = _get_all_independent_vars() for method_id in methods: if verbose > 0: print("running method: ", method_id) ntrials = _ntrials_for_method(method_id=method_id, ntasks=ntasks) # for hparams_dict in _hparams_for_method(method_id)[2:]: # TODO rm for hparams_dict in _hparams_for_method(method_id): if verbose > 3: print("got hparams: ") pprint.pprint(hparams_dict) metrics_dicts = [] try: prev_X_shape, prev_Y_shape = None, None prev_X_std, prev_Y_std = None, None est = None for i, task in enumerate(tasks_func()): if i + 1 > limit_ntasks: raise StopIteration() if verbose > 1: print("-------- running task: {} ({}/{})".format( task.name, i + 1, ntasks)) task.validate_shapes() # fail fast if task is ill-formed can_reuse_est = ( (i != 0) and (est is not None) and (prev_X_shape is not None) and (prev_Y_shape is not None) and (prev_X_std is not None) and (prev_Y_std is not None) and (task.X_train.shape == prev_X_shape) and (task.Y_train.shape == prev_Y_shape) and (task.X_train.std() == prev_X_std) and (task.Y_train.std() == prev_Y_std)) if not can_reuse_est: try: est = _fitted_est_for_hparams( method_id, hparams_dict, task.X_train, task.W_train, task.Y_train) except amm.InvalidParametersException as e: # hparams don't make sense for task (eg, D < d) if verbose > 2: print(f"hparams apparently invalid: {e}") est = None if tasks_all_same_shape: raise StopIteration() else: continue prev_X_shape = task.X_train.shape prev_Y_shape = task.Y_train.shape prev_X_std = task.X_train.std() prev_Y_std = task.Y_train.std() try: # print(f"task {task.name} matrix hashes:") # pprint.pprint(task.hashes()) for trial in range(ntrials): metrics = _eval_amm( task, est, compression_metrics=compression_metrics) metrics['N'] = task.X_test.shape[0] metrics['D'] = task.X_test.shape[1] metrics['M'] = task.W_test.shape[1] metrics['trial'] = trial metrics['method'] = method_id metrics['task_id'] = task.name # metrics.update(hparams_dict) metrics.update(est.get_params()) print("got metrics: ") pprint.pprint(metrics) # pprint.pprint({k: metrics[k] for k in 'method task_id normalized_mse'.split()}) # print("{:.5f}".format(metrics['normalized_mse'])) # TODO uncomment above metrics_dicts.append(metrics) except amm.InvalidParametersException as e: if verbose > 2: print(f"hparams apparently invalid: {e}") if tasks_all_same_shape: raise StopIteration() else: continue except StopIteration: # no more tasks for these hparams pass if len(metrics_dicts): pyn.save_dicts_as_data_frame( metrics_dicts, save_dir='results/amm', name=saveas, dedup_cols=independent_vars) # def main_ecg(methods=None, saveas='ecg', limit_nhours=1): # tasks = md.load_ecg_tasks(limit_nhours=limit_nhours) # return _main(tasks=tasks, methods=methods, saveas=saveas, ntasks=139, # # limit_ntasks=10, compression_metrics=False) # limit_ntasks=5, compression_metrics=True) def main_caltech(methods=methods.USE_METHODS, saveas='caltech', limit_ntasks=-1, limit_ntrain=-1, filt='sobel'): # tasks = md.load_caltech_tasks() # tasks = md.load_caltech_tasks(limit_ntrain=100e3, limit_ntest=10e3) # TODO rm after debug # tasks = md.load_caltech_tasks(limit_ntrain=-1, limit_ntest=10e3) # TODO rm after debug # tasks = md.load_caltech_tasks(limit_ntrain=100e3) # tasks = md.load_caltech_tasks(limit_ntrain=500e3) # tasks = md.load_caltech_tasks(limit_ntrain=1e6) # does great # tasks = md.load_caltech_tasks(limit_ntrain=15e5) # tasks = md.load_caltech_tasks(limit_ntrain=17.5e5) # bad # tasks = md.load_caltech_tasks(limit_ntrain=2e6) # tasks = md.load_caltech_tasks(limit_ntrain=2.5e6) # return _main(tasks=tasks, methods=methods, saveas=saveas, # limit_ntasks = -1 # limit_ntasks = 10 # filt = 'sharpen5x5' # filt = 'gauss5x5' # filt = 'sobel' saveas = '{}_{}'.format(saveas, filt) # saveas = '{}_{}'.format(saveas, filt) # limit_ntrain = -1 # limit_ntrain = 500e3 task_func = functools.partial( md.load_caltech_tasks, filt=filt, limit_ntrain=limit_ntrain) return _main(tasks_func=task_func, methods=methods, saveas=saveas, ntasks=510, limit_ntasks=limit_ntasks, tasks_all_same_shape=True) def main_ucr(methods=methods.USE_METHODS, saveas='ucr', k=128, limit_ntasks=None, problem='rbf'): # limit_ntasks = 10 # limit_ntasks = 13 # tasks = md.load_ucr_tasks(limit_ntasks=limit_ntasks) # k = 128 tasks_func = functools.partial( md.load_ucr_tasks, limit_ntasks=limit_ntasks, k=k, problem=problem) saveas = '{}_k={}_problem={}'.format(saveas, k, problem) return _main(tasks_func=tasks_func, methods=methods, saveas=saveas, ntasks=76, limit_ntasks=limit_ntasks, tasks_all_same_shape=False) def main_cifar10(methods=methods.USE_METHODS, saveas='cifar10'): # tasks = md.load_cifar10_tasks() return _main(tasks_func=md.load_cifar10_tasks, methods=methods, saveas=saveas, ntasks=1) def main_cifar100(methods=methods.USE_METHODS, saveas='cifar100'): # tasks = md.load_cifar100_tasks() return _main(tasks_func=md.load_cifar100_tasks, methods=methods, saveas=saveas, ntasks=1) def main_all(methods=methods.USE_METHODS): main_cifar10(methods=methods) main_cifar100(methods=methods) # main_ecg(methods=methods) main_caltech(methods=methods) def main(): # main_cifar10(methods='ScalarQuantize') # main_cifar100(methods='ScalarQuantize') # main_ucr(methods='ScalarQuantize') main_caltech(methods='ScalarQuantize', filt='sobel') main_caltech(methods='ScalarQuantize', filt='dog5x5') # main_cifar10(methods='MithralPQ') # main_cifar100(methods='Mithral') # main_caltech(methods='Hadamard') # main_cifar10(methods='MithralPQ') # main_cifar100(methods='MithralPQ') # main_ucr(methods='MithralPQ', k=64, limit_ntasks=5, problem='rbf') # main_ucr(methods='Bolt', k=64, limit_ntasks=5, problem='softmax') # rerun mithral stuff with fixed numerical issues # main_cifar10(methods=['Mithral', 'MithralPQ']) # main_cifar100(methods=['Mithral', 'MithralPQ']) # main_ucr(methods=['Mithral', 'MithralPQ'], k=128, problem='rbf') # main_caltech(methods=['Mithral', 'MithralPQ'], filt='sobel') # main_caltech(methods=['Mithral', 'MithralPQ'], filt='dog5x5') # # # # TODO ideally run this too to put in appendix # # # use_methods = list(methods.USE_METHODS) # use_methods.remove(methods.METHOD_SPARSE_PCA) # main_ucr(methods=use_methods, k=128, problem='softmax') # main_caltech('Mithral', filt='sobel', limit_ntrain=1e6, limit_ntasks=10) # lim = 500e3 # lim = 2e6 # lim = -1 # lim = 4e6 # lim = 5e6 # main_caltech('Mithral', filt='sobel', limit_ntrain=lim, limit_ntasks=10) # main_caltech('MithralPQ', filt='sobel', limit_ntrain=lim, limit_ntasks=10) # main_caltech('Mithral', filt='dog5x5', limit_ntrain=lim, limit_ntasks=10) # main_caltech('MithralPQ', filt='dog5x5', limit_ntrain=lim, limit_ntasks=10) # main_caltech('OldMithralPQ', filt='sobel', limit_ntrain=lim, limit_ntasks=10) # main_ucr(methods='MithralPQ', limit_ntasks=5) # main_caltech(methods='Bolt', limit_ntasks=10, limit_ntrain=500e3, filt='dog5x5') # main_caltech(methods='Bolt', limit_ntasks=10, limit_ntrain=500e3, filt='sobel') # main_caltech(methods='SparsePCA') if __name__ == '__main__': np.set_printoptions(formatter={'float': lambda f: "{:.2f}".format(f)}, linewidth=100) main()
#!/bin/env/python import abc import numpy as np # from sklearn.decomposition import PCA, SparsePCA from sklearn import decomposition from sklearn.decomposition import PCA, SparsePCA, MiniBatchSparsePCA from sklearn.utils.extmath import randomized_svd import numba # conda install numba # import ffht # https://github.com/FALCONN-LIB/FFHT; python setup.py install import scipy from joblib import Memory _memory = Memory('.', verbose=1, compress=9) KEY_NMULTIPLIES = 'muls' OSNAP_DEFAULT_S = 4 # OSNAP_DEFAULT_S = 2 # ================================================================ utils def _nmultiplies_matmul(A, B): return A.shape[0] * A.shape[1] * B.shape[1] def _nmultiplies_matmul_with_sizes(N, D, M): return N * D * M def _nmultiplies_svd(N, D): return min(N * N * D, N * D * D) def _nmultiplies_qr(N, D): return min(N * N * D, N * D * D) # ================================================================ types class InvalidParametersException(Exception): pass class ApproxMatmul(abc.ABC): def __init__(*args_unused, **kwargs_unused): pass def fit(self, A, B, Y=None): # Y = A @ B if not specified pass def set_A(self, A): pass def set_B(self, B): pass def reset_for_new_task(self): pass @abc.abstractmethod def __call__(self, A, B): pass def predict(self, A, B): return self(A, B) def get_params(self): return {} # def get_nmuls(self, A, B, fixedA=False, fixedB=False): @abc.abstractmethod def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): pass class ExactMatMul(ApproxMatmul): def __call__(self, A, B): return A @ B def get_speed_metrics(self, A, B, **sink): return {KEY_NMULTIPLIES: _nmultiplies_matmul(A, B)} def _scalar_quantize(A, axis=1, signed=False, nbits=8): unsigned_maxval = float(1 << int(nbits)) - 1 # # TODO rm # # return np.zeros((A.shape[0], 1)), np.ones((A.shape[0], 1)), A # # offsets = np.zeros((A.shape[0], 1)) # offsets = A.min(axis=1, keepdims=True) # # scales = maxval / np.ones((A.shape[0], 1)) # scales = maxval / A.max(axis=1, keepdims=True) # Aq = (A - offsets) * scales # return offsets, scales, Aq # maxval = float(1 << int(nbits)) - 1 mins = A.min(axis=axis, keepdims=True) # A_offset = A - offsets ranges = (A - mins).max(axis=axis, keepdims=True) + 1e-20 scales = unsigned_maxval / ranges # Aq = (A_offset * (maxval / scales)).astype(np.int) # Aq = (A_offset * scales).astype(np.int) if signed: # sign_offset = 1 << (nbits - 1) # 8 bits -> 128 # A_offset -= sign_offset offsets = mins + (ranges * (128. / 255)) minval = -(1 << (nbits - 1)) maxval = -minval - 1 else: offsets = mins minval = 0 maxval = (1 << nbits) - 1 Aq = (A - offsets) * scales # print("min, max A:", Aq.min(), Aq.max()) # looks good Aq = np.clip(Aq, minval, maxval).astype(np.int) return offsets, scales, Aq class QuantizedMatmul(ApproxMatmul): __slots__ = 'nbits a_offsets a_scales b_offsets b_scales A B'.split() def __init__(self, nbits=8): self.nbits = nbits def __call__(self, A, B): assert A.shape[1] == B.shape[0] # dims need to match N, D = A.shape D, M = B.shape if self.A is None: self.set_A(A) if self.B is None: self.set_B(B) # print("QuantizedMatmul") # print("min, max A:", self.A.min(), self.A.max()) # print("min, max A offsets:", self.a_offsets.min(), self.a_offsets.max()) # print("min, max A scales :", self.a_scales.min(), self.a_scales.max()) # print("min, max B:", self.B.min(), self.B.max()) # print("min, max B offsets:", self.b_offsets.min(), self.b_offsets.max()) # print("min, max B scales :", self.b_scales.min(), self.b_scales.max()) # ((A - a_offsets) / a_scales) @ ((B - b_offsets) / b_scales) # noqa # ignoring scales, we have: # (A - a_off) @ (B - b_off) # = A @ B - (a_off @ B) - (A @ b_off) + a_off @ b_off # maxval = (1 << int(self.nbits)) - 1 ret = (self.A @ self.B).astype(np.float32) ret *= 1. / self.a_scales ret *= 1. / self.b_scales A_off = np.tile(self.a_offsets, (1, D)) B_off = np.tile(self.b_offsets, (D, 1)) return ret + (A_off @ B) + (A @ B_off) - (A_off @ B_off) def set_A(self, A): # unsigned quantization; we *could* learn the offsets and scales # on the training set, but since this is a baseline, we're giving it # the advantage of using the "true" offsets/scales self.a_offsets, self.a_scales, self.A = _scalar_quantize( A, axis=1, signed=False, nbits=self.nbits) # mins = A.min(axis=1, keepdims=True) # A_offset = A - mins # scales = A_offset.max(axis=1, keepdims=True) + 1e-20 # self.A = (A_offset * (255. / scales)).astype(np.int) def set_B(self, B): # signed quantization (for maddubs instruction) self.b_offsets, self.b_scales, self.B = _scalar_quantize( B, axis=0, signed=True, nbits=self.nbits) # self.b_offsets, self.b_scales, self.B = _scalar_quantize( # B.T, nbits=self.nbits, signed=True) # # quantize each col, not each row # self.b_offsets = self.b_offsets.ravel() # self.b_scales = self.b_scales.ravel() # self.B = self.B.T def reset_for_new_task(self): self.A = None self.B = None def get_speed_metrics(self, A, B, **sink): # neglect packing, postprocessing, etc return {KEY_NMULTIPLIES: _nmultiplies_matmul(A, B)} class SketchedMatmul(ApproxMatmul, abc.ABC): __slots__ = 'd' def __init__(self, d): self.d = int(d) def get_params(self): return {'d': self.d} def sketch(self, A, B): pass def call(self, A, B): A_hat, B_hat = self.sketch(A, B) assert A_hat.shape[0] == A.shape[0] assert B_hat.shape[1] == B.shape[1] assert A_hat.shape[1] <= self.d # verify sketch size not cheating return A_hat @ B_hat def __call__(self, A, B): assert A.shape[1] == B.shape[0] # dims need to match D = A.shape[1] if D <= self.d: raise InvalidParametersException( 'D <= d: {} < {}'.format(D, self.d)) if B.shape[1] <= self.d: raise InvalidParametersException( 'M <= d: {} < {}'.format(B.shape[1], self.d)) return self.call(np.copy(A), np.copy(B)) # guarantee A, B unchanged def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): assert not (fixedA and fixedB) # this would be stupid, so fail fast sketch_nmuls = self._get_nmuls(A.shape[0], A.shape[1], B.shape[1], self.d, fixedA=fixedA, fixedB=fixedB) N, D = A.shape D, M = B.shape sketched_matmul_nmuls = N * self.d * M return {KEY_NMULTIPLIES: sketch_nmuls + sketched_matmul_nmuls} def _get_nmuls(self, N, D, M, d, fixedA=False, fixedB=False): # default nmuls = sketching with dense matrix nmuls = 0 if not fixedA: nmuls += N * D * d if not fixedB: nmuls += M * D * d return nmuls class RandGaussSketch(SketchedMatmul): def sketch(self, A, B): D = A.shape[1] V = np.random.randn(D, self.d).astype(np.float32) # dividing by expected norm is more similar to theory papers, # but no reason this should actually be better AFAIK # V /= np.sqrt(D) V /= np.linalg.norm(V, axis=0) A = A @ V B = V.T @ B return A, B class RandOrthoGaussSketch(SketchedMatmul): def sketch(self, A, B): D = A.shape[1] V = np.random.randn(D, self.d).astype(np.float32) V, _ = np.linalg.qr(V) A = A @ V B = V.T @ B return A, B class RandRademacherSketch(SketchedMatmul): def sketch(self, A, B): D = A.shape[1] V = np.random.randint(2, size=(D, self.d)).astype(np.float32) * 2 - 1 V /= np.sqrt(D) A = A @ V B = V.T @ B return A, B class HadamardSketch(SketchedMatmul): def sketch(self, A, B): D = A.shape[1] use_D = 1 << int(np.ceil(np.log2(D))) V = scipy.linalg.hadamard(use_D)[:D, :self.d].astype(np.float32) V /= np.linalg.norm(V, axis=0) # V /= np.sqrt(2) # V *= np.sqrt(2) # V *= np.sqrt(D / self.d) # V *= (D / self.d) ** .25 A = A @ V B = V.T @ B return A, B class SketchSqSample(SketchedMatmul): def sketch(self, A, B): return sketch_sq_sample(A, B, self.d) def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_sketch_sq_sample(N, D, M, d) class FdAmm(SketchedMatmul): def sketch(self, A, B): return fd_amm_sketches(A, B, self.d) def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_fd_amm_sketches(N, D, M, d) class CooccurSketch(SketchedMatmul): def sketch(self, A, B): return cooccur_sketches(A, B, self.d) def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_cooccur_sketches(N, D, M, d) class FastJlSketch(SketchedMatmul): def sketch(self, A, B): return fastjl_sketches(A, B, self.d) def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_fastjl_sketches(N, D, M, d) class HashJlSketch(SketchedMatmul): def sketch(self, A, B): return hash_sketches(A, B, self.d) def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_hash_sketches(N, D, M, d) class OsnapSketch(SketchedMatmul): def sketch(self, A, B): return osnap_sketches(A, B, self.d, s=OSNAP_DEFAULT_S) # def get_params(self): # return {'d': self.d, 's': OSNAP_DEFAULT_S} def _get_nmuls(self, N, D, M, d, **sink): return _nmultiplies_osnap_sketches(N, D, M, d) class SvdSketch(SketchedMatmul): __slots__ = 'd niters Ua SVTa Ub SVTb'.split() def __init__(self, d, niters=5): self.d = d self.niters = niters self.reset_for_new_task() def get_params(self): return {'d': self.d, 'niters': self.niters} def _check_mat_shape(self, M): if M is None: return False # if np.min(M.shape) < self.d: if np.max(M.shape) < self.d: raise InvalidParametersException( 'shape has entry < d: {} < {}'.format(M.shape, self.d)) return True def set_A(self, A): # if A is None: # return if self._check_mat_shape(A): self.Ua, self. SVTa = svd_sketch(A, self.d) def set_B(self, B): if self._check_mat_shape(B): self.Ub, self.SVTb = svd_sketch(B, self.d) def reset_for_new_task(self): self.Ua = None self.SVTa = None self.Ub = None self.SVTb = None # def __call__(self, A=None, B=None): # assert A.shape[1] == B.shape[0] # dims need to match # if A.shape[1] < self.d: # raise InvalidParametersException('D < d') def call(self, A=None, B=None): if self.Ua is None: self.set_A(A) if self.Ub is None: self.set_B(B) D = self.Ua.shape[1] if D < self.d: raise InvalidParametersException( 'D < d: {} < {}'.format(D, self.d)) # verify sketch size isn't cheating # print("A.shape", A.shape) # print("B.shape", B.shape) # print("self.Ua.shape: ", self.Ua.shape) # print("self.SVTa.shape: ", self.SVTa.shape) # print("self.Ub.shape: ", self.Ub.shape) # print("self.SVTb.shape: ", self.SVTb.shape) # print("self.d: ", self.d) assert self.Ua.shape[1] <= self.d assert self.SVTa.shape[0] <= self.d assert self.SVTb.shape[0] <= self.d assert self.Ub.shape[1] <= self.d # innermost parens important so that matmuls actually use low rank # outer parens help if B ncols < A nrows (which is true for us) return self.Ua @ ((self.SVTa @ self.Ub) @ self.SVTb) def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): # XXX this will break if not called right after self.call() total = 0 d = self.d N, D = A.shape _, M = B.shape if not fixedA: total += _nmultiplies_svd_sketch(N, D, d, niters=self.niters) if not fixedB: total += _nmultiplies_svd_sketch(D, M, d, niters=self.niters) total += d * D * d # SVTa @ UB, d x D @ D x d total += d * d * M # (above) @ SVTb, d x d @ d x M total += N * d * M # Ua @ (above), N x d @ d x M return {KEY_NMULTIPLIES: total} @_memory.cache def _fitted_pca(X, n_components): pca = PCA(n_components=n_components) return pca.fit(X) class TrainedPcaSketch(ApproxMatmul): __slots__ = 'pca d A B V'.split() def __init__(self, d): # self.pca = PCA(n_components=d) self.d = d self.reset_for_new_task() def reset_for_new_task(self): self.A = None self.B = None def fit(self, A, B, Y=None): # Y = A @ B if not specified D, M = B.shape print("called fit on TrainedPcaSketch!") if D < self.d: raise InvalidParametersException( 'D < d: {} < {}'.format(D, self.d)) if M < self.d: raise InvalidParametersException( 'M < d: {} < {}'.format(M, self.d)) self.pca = _fitted_pca(A, n_components=self.d) self.V = self.pca.components_.T # print("components V.T @ V =\n", self.V.T @ self.V) # yep, orthonormal def set_A(self, A): self.A = A @ self.V def set_B(self, B): self.B = self.V.T @ B def __call__(self, A, B): assert A.shape[1] == B.shape[0] # dims need to match if B.shape[1] < self.d: raise InvalidParametersException( 'M < d: {} < {}'.format(B.shape[1], self.d)) if (self.A is None): self.set_A(A) if (self.B is None): self.set_B(B) return self.A @ self.B def get_params(self): return {'d': self.d} def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): N, D = A.shape D, M = B.shape d = self.d nmuls = N * d * M # assuming matrices already sketched if not fixedA: nmuls += N * D * d if not fixedB: nmuls += D * M * d return {KEY_NMULTIPLIES: nmuls} @_memory.cache def _fitted_sparse_pca(X, d, unscaled_alpha, **kwargs): # this seems to work better than initializing with MiniBatchSparsePCA, # svd of cov mat, or basically anything else I tried U, _, Vt = randomized_svd(X, n_components=d, random_state=123) U = U[:, :d] V = Vt.T[:d] # SparsePCA (and all the sklearn dictionary learning stuff) # internally uses sum of squared errs for each sample, and L1 norm # of parameter matrix; to make alpha meaningful across datasets, # want to scale by number of examples (so it's effectively using MSE) # and divide by L1 norm (which grows linearly with size of parameter # matrix / vector); also scale by variance of data for similar reasons N, D = X.shape alpha = unscaled_alpha * np.var(X - X.mean(axis=0)) * N / D verbose = 1 pca = SparsePCA(n_components=d, alpha=alpha, normalize_components=True, method='lars', U_init=U, V_init=V, max_iter=10, ridge_alpha=max(1, len(X) * X.std() * 10), # ridge_alpha=1e8, verbose=verbose, random_state=123) if verbose > 0: print("fitting sparse pca...") return pca.fit(X) class TrainedSparsePcaSketch(ApproxMatmul): __slots__ = 'pca d alpha nnz can_optimize_transform A B'.split() # def __init__(self, d, alpha, can_optimize_transform=True): def __init__(self, d, alpha, can_optimize_transform=False): self.d = d self.alpha = alpha self.can_optimize_transform = can_optimize_transform self.reset_for_new_task() def reset_for_new_task(self): self.A = None self.B = None def fit(self, A, B, Y=None): # Y = A @ B if not specified D, M = B.shape # if M <= self.d: # raise InvalidParametersException( # 'M <= d: {} < {}'.format(M, self.d)) if D <= self.d: raise InvalidParametersException( 'D < d: {} < {}'.format(D, self.d)) self.pca = _fitted_sparse_pca(A, d=self.d, unscaled_alpha=self.alpha) self.nnz = np.sum(self.pca.components_ != 0) sparsity = np.mean(self.pca.components_ == 0) if self.nnz < self.d: raise InvalidParametersException( "ignoring SparsePCA with nnz < d: " "{} < {}".format(self.nnz, self.d)) if sparsity == 0.: raise InvalidParametersException( "ignoring SparsePCA with no zeros") def set_A(self, A): if self.can_optimize_transform: # uses ridge regression to get coeffs, instead of linear projection # disabled by default because it produces garbage on caltech and # is more expensive than just doing the matmul self.A = self.pca.transform(A) self.A += self.pca.mean_ @ self.pca.components_.T else: self.A = A @ self.pca.components_.T def set_B(self, B): if self.can_optimize_transform: self.B = self.pca.transform(B.T).T self.B += (self.pca.mean_ @ self.pca.components_.T).reshape(-1, 1) else: self.B = (B.T @ self.pca.components_.T).T def __call__(self, A, B): assert A.shape[1] == B.shape[0] # dims need to match N, D = A.shape D, M = B.shape if D <= self.d: raise InvalidParametersException( 'D < d: {} < {}'.format(D, self.d)) fixedA = self.A is not None fixedB = self.B is not None nmuls_naive = N * D * M nmuls_ours = self.get_speed_metrics( A, B, fixedA=fixedA, fixedB=fixedB)[KEY_NMULTIPLIES] if nmuls_naive <= nmuls_ours: raise InvalidParametersException( "naive # of multiplies < sparse sketch # of multiplies: " "{} < {}".format(nmuls_naive, nmuls_ours)) if not fixedA: self.set_A(A) if not fixedB: self.set_B(B) # if N == 700: # if False: print("got to weird dset!") # print("pca means: ", self.pca.mean_[::20]) # print("A means:", A.mean(axis=0)[::20]) # print("B means:", B.mean(axis=1)[::20]) print("pca means sum: ", self.pca.mean_.sum()) print("A means sum: ", A.mean(axis=0).sum()) print("B means sum: ", B.mean(axis=1).sum()) offsets = (self.pca.mean_ @ self.pca.components_.T) print("offsets: ", offsets) print("offsets sum: ", offsets.sum()) # C = (A @ B) # print("true mean of output: ", C.mean()) # print("true std of output: ", C.std()) return self.A @ self.B def get_params(self): return {'d': self.d, 'alpha': self.alpha, 'canCheat': self.can_optimize_transform} def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): N, D = A.shape D, M = B.shape nmuls_sketch_X = N * self.nnz nmuls_sketch_W = M * self.nnz nmuls_make_output = N * self.d * M total_nmuls = nmuls_make_output if not fixedA: total_nmuls += nmuls_sketch_X if not fixedB: total_nmuls += nmuls_sketch_W try: # compute degree of sparsity nnz = self.nnz sparsity = (self.pca.components_ == 0).mean() except AttributeError: # model not fitted yet nnz = -1 sparsity = -1 return {KEY_NMULTIPLIES: total_nmuls, 'nnz': nnz, 'sparsity': sparsity} # ================================================================ drineas06 def _compute_dim_scores(A, B, A_col_norms=None, B_row_norms=None): if A_col_norms is None: A_col_norms = np.linalg.norm(A, axis=0) if B_row_norms is None: B_row_norms = np.linalg.norm(B, axis=1) return A_col_norms * B_row_norms def sketch_sq_sample(A, B, d): scores = _compute_dim_scores(A, B) idxs, weights = importance_sample(scores, d) # idxs, weights = sample_varopt_1d(scores, d) # doesn't help return A[:, idxs] / weights, B[idxs] # weights = np.sqrt(weights) # return A[:, idxs] / weights, B[idxs] / weights.reshape(-1, 1) # probs = scores / np.sum(scores) # D = A.shape[1] # keep_idxs = np.random.choice(D, size=d, p=probs) # # keep_idxs = np.random.choice(D, size=d, p=probs, replace=False) # # keep_idxs = np.random.choice(D, size=d, replace=False) # # keep_idxs = np.arange(D-1) # # keep_idxs = np.arange(1, D) # # keep_idxs = np.arange(D) # weights = np.sqrt(d * probs) # what the paper says; huge errors # # weights = np.sqrt(D * probs) # slightly less bad # # weights = np.sqrt(np.sqrt(d * probs)) # # weights = np.ones(D) # A = np.copy(A) / weights # B = np.copy(B) / weights.reshape(-1, 1) # return np.copy(A[:, keep_idxs]), np.copy(B[keep_idxs]) # return A[:, keep_idxs], B[keep_idxs] # return A, B def _nmultiplies_sketch_sq_sample(N, D, M, d): scores_nmuls = N * D + M * D # sum of sizes of each mat reweight_nmuls = N * d + M * d # sum of sizes of each sampled mat return scores_nmuls + reweight_nmuls # neglect normalization of probs, etc def sketch_sq_deterministic(A, B, d): scores = _compute_dim_scores(A, B) D = A.shape[1] keep_idxs = np.argsort(scores)[::-d] weights = np.sqrt(d * (1. / D)) # uniform prob return A[:, keep_idxs] / weights, B[keep_idxs] / weights.reshape(-1, 1) def test_sketch_sq_sample(): print("test_sketch_sq_sample") N, M, D = 100, 50, 200 np.random.seed(1234) # A = np.random.randint(5, size=(N, D)).astype(np.float32) # B = np.random.randint(5, size=(D, M)).astype(np.float32) # A -= np.mean(A) # B -= np.mean(B) A = np.random.randn(N, D).astype(np.float32) B = np.random.randn(D, M).astype(np.float32) AB = A @ B orig_frob_sq = np.mean(AB * AB) print("true mss: ", orig_frob_sq) prev_normed_err = np.inf for d in (10, 20, 30, 40, 50): A_hat, B_hat = sketch_sq_sample(A, B, d) # A_hat, B_hat = sketch_sq_deterministic(A, B, d) AB_hat = A_hat @ B_hat # print("AB_hat mss: ", (AB_hat * AB_hat).mean()) diffs = AB - AB_hat err_frob_sq = np.mean(diffs * diffs) normed_err_sq = err_frob_sq / orig_frob_sq # print("orig mss: ", orig_frob_sq) print('d = {}, err = {:.3f}'.format(d, normed_err_sq)) assert normed_err_sq < 2. assert normed_err_sq < (prev_normed_err + .05) # should usually hold prev_normed_err = normed_err_sq # ================================================================ sampling # wait, this just returns points summing to the true sample sum # deterministically... def importance_sample(sample_weights, m, replace=False): probs = sample_weights / sample_weights.sum() idxs = np.random.choice( np.arange(len(sample_weights)), p=probs, replace=replace, size=m) weights = 1. / (probs[idxs] * m) return idxs, weights def _invert_permutation(permutation): return np.arange(len(permutation))[np.argsort(permutation)] def _sum_for_tau(x, tau): above_tau = x > tau return x[above_tau].sum() + (x[~above_tau] / tau).sum() def _compute_new_tau(x_sorted_desc, m, tau=0): x = x_sorted_desc current_sum = _sum_for_tau(x, tau) assert current_sum >= m while current_sum > m: x = x[:-1] current_sum = _sum_for_tau(x, tau) def sample_varopt_1d(x, m): # varopt sampling; original paper (see Algorithm 1 on p16): # https://arxiv.org/pdf/0803.0473.pdf # better intuition: # https://datasketches.github.io/docs/Sampling/VarOptSampling.html # # unlike paper, we're just going to do it all at once since that will # be simpler and vectorize way better; basically just recursively # take largest point w_i if w_i > (m / sum_i w_i), with m decremented # by 1 each time; if this doesn't take all the points, importance sample # from the remaining points (with probs proportional to their weights) # # EDIT: this sucks unless really heavy tailed, so probably not a # correct impl? x = np.asarray(x, dtype=np.float32) n = len(x) if m >= n: return np.arange(n) maxval = np.max(x) minval = np.min(x) assert minval >= 0 # needs nonnegative entries if minval == maxval or m == 1: return np.random.choice(np.arange(n), size=m) sort_idxs = np.argsort(x)[::-1] # in descending order x_sorted = x[sort_idxs] unsort_idxs = _invert_permutation(sort_idxs) q = x_sorted * (m / np.sum(x_sorted)) # sums to m # q_tailsums = np.cumsum(q[::-1])[::-1] # next_val = x_sorted[0] head_sz = 0 for i in range(m): if q[0] >= 1.: head_sz += 1 q = q[1:] * ((m - 1) / q[1:].sum()) # TODO just compute tail sums once for renormalization (below) # q_mass_eliminated = q[i] # next_val = q[i + 1] * (m - head_sz) / m * () # renormalize such that tail sums to m - 1 else: break tail_sz = m - head_sz # print("m, head_sz, tail_sz:", m, head_sz, tail_sz) # print("len(q)", len(q)) # probs = q / np.sum(q) probs = x_sorted[head_sz:] / np.sum(x_sorted[head_sz:]) tail_idxs = np.random.choice( np.arange(head_sz, n), p=probs, replace=False, size=tail_sz) idxs = list(tail_idxs) # idxs = tail_idxs # tau = tail_sz / np.sum(x_sorted[head_sz:]) # print("tau: ", tau) # print("x_sorted[:head_sz + 1]: ", x_sorted[:head_sz + 1]) # tau = x_sorted[head_sz] true_probs = probs[tail_idxs - head_sz] * (tail_sz / m) weights = list(1. / (m * true_probs)) # small err; definitely right # weights = [tau] * tail_sz if head_sz > 0: head_idxs = list(np.arange(head_sz)) head_weights = list(np.ones(head_sz)) idxs = head_idxs + idxs weights = head_weights + weights return unsort_idxs[idxs], np.array(weights) # ============================================================ random sketches # sketch both A and B jointly using the same matrix to amortize overhead and # because it seems like this should help accuracy # @numba.jit(nopython=True) def fastjl_sketches(A, B, d, P=None): N, D = A.shape M = B.shape[1] # pad A and B for FHT log2_D = int(np.ceil(np.log2(D))) D_pad = 2 ** log2_D A_pad = np.zeros((N, D_pad), dtype=np.float32) A_pad[:, :D] = A B_pad = np.zeros((D_pad, M), dtype=np.float32) B_pad[:D] = B # construct and apply random signs for each dim randsigns = np.random.randint(0, 2, size=D_pad) * 2 - 1 # scale now instead of scaling FHT mat, so only O(D) multiplies randsigns = randsigns.astype(np.float32) * (1. / np.sqrt(D_pad)) A_pad *= randsigns B_pad *= randsigns.reshape(-1, 1) # # apply fast hadamard transform H = scipy.linalg.hadamard(D_pad, dtype=np.float32) # H = scipy.linalg.hadamard(D_pad, dtype=np.float32) / np.sqrt(D_pad) A_pad = A_pad @ H B_pad = H @ B_pad # dimensionalty reduction if P is None: # logd = np.log2(D_pad) keep_prob = log2_D * log2_D / D_pad # if (keep_prob) >= 1: # print("WARNING: FastJL returning all zeros mat...") P = (np.random.uniform(size=(D_pad, d)) > keep_prob).astype(np.float32) # P *= np.random.randn(*P.shape) * (d / keep_prob) # scaling sigma totally fails; need norm to actually be 1, not just # have expected value of 1 P *= np.random.randn(*P.shape) P *= (1. / np.linalg.norm(P, axis=0)) # print("P shape, Apad shape, Bpad shape: ", P.shape, A_pad.shape, B_pad.shape) return A_pad @ P, P.T @ B_pad def _nmultiplies_fastjl_sketches(N, D, M, d): # avg, not exact, since P sparse # technically adds or subs, but you'd do fma ops regardless for floats log2_D = int(np.ceil(np.log2(D))) D_pad = 2 ** log2_D fht_nmuls = D_pad * np.log2(D_pad) sign_nmuls = D_pad # trickier part; expected number of madds (or similar ops) to mul by P construct_P_nmuls = D_pad * d # assuming only 1 mul for rng + threshold keep_prob = log2_D * log2_D / D_pad nnz_p = min(1, keep_prob) * D_pad # expected nnz per row of P p_nmuls = N * nnz_p * d + d * nnz_p * M return fht_nmuls + sign_nmuls + construct_P_nmuls + p_nmuls @numba.jit(nopython=True) def hash_sketches(A, B, d, scale=1., share_projections=True): N, D = A.shape D, M = B.shape A_hat = np.zeros((N, d), dtype=A.dtype) B_hat = np.zeros((d, M), dtype=B.dtype) for j in range(D): idx = np.random.randint(d) sign = (np.random.randint(0, 2) * 2) - 1 # coeff = sign * scale # worse than predicting mean, esp for small d coeff = sign * scale / np.sqrt(2) # actually pretty decent # coeff = sign * scale * ((d / D) ** .25) # coeff = sign * scale * np.sqrt(d / D) # best for small d / D # coeff = sign * scale * d / D # best for larger d / D A_hat[:, idx] += A[:, j] * coeff if share_projections: B_hat[idx] += B[j] * coeff continue # use a different projection for B idx = np.random.randint(d) sign = (np.random.randint(0, 2) * 2) - 1 B_hat[idx] += B[j] * sign # using unscaled signs preserves norms really well, at least for # random matrices # print("A norm, A_hat norm:", np.linalg.norm(A), np.linalg.norm(A_hat)) # print("B norm, B_hat norm:", np.linalg.norm(B), np.linalg.norm(B_hat)) # A_norm = np.linalg.norm(A) # B_norm = np.linalg.norm(B) # A_hat *= np.linalg.norm(A) / np.linalg.norm(A_hat) # B_hat *= np.linalg.norm(B) / np.linalg.norm(B_hat) return A_hat, B_hat def osnap_sketches(A, B, d, s=OSNAP_DEFAULT_S): N, D = A.shape D, M = B.shape s = max(1, min(d // 2, s)) # handle s too large relative to d A_hat = np.zeros((N, d), dtype=A.dtype) B_hat = np.zeros((d, M), dtype=B.dtype) scale = 1. / np.sqrt(s) # scale = 1. / s # scale = 1 # seems to often work better than dividing by 1/sqrt(s)? # scale = np.sqrt(s) # scale = s subspace_len = (d + s - 1) // s # round up for ss in range(s): start_idx = ss * subspace_len end_idx = min(D, start_idx + subspace_len) A_hat[:, start_idx:end_idx], B_hat[start_idx:end_idx] = \ hash_sketches(A, B, subspace_len, scale=scale) # A_hat /= np.linalg.norm(A_hat, axis=) return A_hat, B_hat def _nmultiplies_hash_sketches(N, D, M, d): # technically adds or subs, but you'd do fma ops regardless for floats return N * D + D * M def _nmultiplies_osnap_sketches(N, D, M, d, s=4): return 4 * _nmultiplies_hash_sketches(N, D, M, d) def test_rand_sketches(): print("test_svd_sketches") N, M, D = 100, 80, 50 np.random.seed(1234) A = np.random.randint(5, size=(N, D)).astype(np.float32) B = np.random.randint(5, size=(D, M)).astype(np.float32) A -= np.mean(A) B -= np.mean(B) AB = A @ B orig_frob_sq = np.sum(AB * AB) prev_normed_err = np.inf # for d in [10]: for d in (1, 2, 4, 8, 16, 32): # (Ua, SVTa), (Ub, SVTb) = svd_sketches(A, B, d) # AB_hat = Ua @ (SVTa @ Ub) @ SVTb A_hat, B_hat = fastjl_sketches(A, B, d) # A_hat, B_hat = hash_sketches(A, B, d) # sharing projections helps # A_hat, B_hat = hash_sketches(A, B, d, share_projections=False) # A_hat, B_hat = osnap_sketches(A, B, d) AB_hat = A_hat @ B_hat # print("fused mats shapes: ") # print(Ua.shape, SVTa.shape, Ub.shape, SVTb.shape) diffs = AB - AB_hat err_frob_sq = np.sum(diffs * diffs) normed_err_sq = err_frob_sq / orig_frob_sq print('d = {}, err = {:.5f}'.format(d, normed_err_sq)) # assert normed_err_sq < 1. # assert normed_err_sq < prev_normed_err + .001 prev_normed_err = normed_err_sq # ================================================================ Rand SVD def svd_sketch(A, d, niters=5, **kwargs): # assert A.shape[0] >= d # assert A.shape[1] >= d assert np.max(A.shape) >= d # can't truncate to larger size U, S, Vt = randomized_svd(A, n_components=d, n_iter=niters, **kwargs) # print("Vt shape: ", Vt.shape) # print("S: ", S) return (U, np.diag(S) @ Vt) def _nmultiplies_svd_sketch(N, D, d, niters): # # "In contrast, randomized schemes can produce an approximate SVD using # # only O(mn log(k) + (m + n)k2) flops" -Halko et al. 2010 # # https://arxiv.org/pdf/0909.4061.pdf # iter_cost = N * D * int(np.ceil(np.log2(d))) # iter_cost += (N + D) * d * d # return iter_cost * niters # # assumes algorithm 4.4 in above; sklearn randomized_svd source # # code says it implements algorithm 4.3, but paper says 4.3 should actually # # be implemented as 4.4 in practice. Also 4x4's complexity is much easier # # to understand and counting multiplies is at best a rough estimate # # regardless. # # # # shapes: # # A: N x D # # A*: D x N # # Omega: D x d # # Y0 = A @ Omega: N x d # # Q0: N x d # # R0: d x d # # Y_tilde_j: # # gauss_mat_cost = D * d # # Y0_cost = N * D * d # Y0_cost = N * D * int(np.ceil(np.log2(d))) # subsampled FFT; see text # Y0_cost += _nmultiplies_qr(N, d) # Yj_tilde_cost = D * N * d + _nmultiplies_qr(N, d) # Yj_cost = # okay, sklearn says it uses algorithm 4.3 in Halko et al. 2010 [1], # so we're going to go with that # [1] https://arxiv.org/pdf/0909.4061.pdf # shapes: # A: N x D # A.T: D x N # G (Omega): D x d # A @ G: N x d # A.T @ (AG) D x d # A @ (A.T@A@G) N x d # Q0: N x d # R0: d x d Omega_cost = D * d A_Omega_cost = N * D * d # each iter: premul by A.T, then A; assumes no LU or QR for stability iter_cost = D * N * d + N * D * d return Omega_cost + A_Omega_cost + iter_cost * niters def svd_sketches(A, B, d, **kwargs): return svd_sketch(A, d, **kwargs), svd_sketch(B, d, **kwargs) # Ua, Sa, VTa = randomized_svd(A, n_components=d, **kwargs) # Ub, Sb, VTb = randomized_svd(B, n_components=d, **kwargs) # print("truncated svd mat shapes:") # print(Ua.shape, Sa.shape, VTa.shape) # print(Ub.shape, Sb.shape, VTb.shape) # return (Ua, np.diag(Sa) @ VTa), (Ub, np.diag(Sb) @ VTb) def test_svd_sketches(): print("test_svd_sketches") N, M, D = 100, 80, 50 np.random.seed(1234) A = np.random.randint(5, size=(N, D)).astype(np.float32) B = np.random.randint(5, size=(D, M)).astype(np.float32) A -= np.mean(A) B -= np.mean(B) AB = A @ B orig_frob_sq = np.sum(AB * AB) prev_normed_err = np.inf # for d in [10]: for d in (1, 2, 4, 8, 16, 32): (Ua, SVTa), (Ub, SVTb) = svd_sketches(A, B, d) AB_hat = Ua @ (SVTa @ Ub) @ SVTb # print("fused mats shapes: ") # print(Ua.shape, SVTa.shape, Ub.shape, SVTb.shape) diffs = AB - AB_hat err_frob_sq = np.sum(diffs * diffs) normed_err_sq = err_frob_sq / orig_frob_sq print('d = {}, err = {:.5f}'.format(d, normed_err_sq)) assert normed_err_sq < 1. assert normed_err_sq < prev_normed_err prev_normed_err = normed_err_sq # ================================================================ FD methods # TODO impl fast-FD, which zeros out half the entries def frequent_directions(A, d, variant=None): N, D = A.shape H = np.zeros((d, D)) assert N >= d assert D >= d # for i in range(N): H[:d - 1] = A[:d - 1] for i in range(d - 1, N): H[-1] = A[i] try: U, S, Vt = np.linalg.svd(H, full_matrices=False) # d x d, d, d x D except np.linalg.LinAlgError as e: print("SVD failed at iter ", i - (d - 1)) print("H shape: ", H.shape) print("A shape: ", A.shape) print("d: ", d) # print("svd mat shape: ", U.shape, S.shape, Vt.shape) raise e # cutoff = S[d - 1] # S is returned as a vector, not a diagonal mat if variant == 'robust': raise NotImplementedError() else: S = np.sqrt((S - S[-1]) ** 2) # note that last entry is dth entry # print("new S shape: ", S.shape) # H = np.diag(S) @ Vt # d x D H = Vt * S.reshape(-1, 1) # d x D; equivalent to np.diag(S) @ Vt return H def fast_frequent_directions(A, d, variant=None, alpha=.5): N, D = A.shape # H = np.zeros((d, D)) H = np.copy(A[:d]) assert N >= d assert D >= d cutoff_idx = int(d * (1 - alpha)) cutoff_idx = min(d - 1, cutoff_idx) # always zero at least last element ntrailing_zeros = d - cutoff_idx i = d while i < N: try: U, S, Vt = np.linalg.svd(H, full_matrices=False) # d x d, d, d x D except np.linalg.LinAlgError as e: print("SVD failed at iter ", i - (d - 1)) print("H shape: ", H.shape) print("A shape: ", A.shape) print("d: ", d) # print("svd mat shape: ", U.shape, S.shape, Vt.shape) raise e cutoff = S[cutoff_idx] if variant == 'parametrized': raise NotImplementedError() else: S = np.sqrt(np.maximum(S - cutoff, 0) ** 2) S = np.sqrt((S - S[-1]) ** 2) # note that last entry is dth entry # print("new S shape: ", S.shape) # H = np.diag(S) @ Vt # d x D H = Vt * S.reshape(-1, 1) # d x D; equivalent to np.diag(S) @ Vt # replace zeroed-out rows of H with next rows of A end_dim = min(N, i + ntrailing_zeros) nrows_to_copy = end_dim - i end_row = cutoff_idx + nrows_to_copy assert nrows_to_copy <= ntrailing_zeros assert end_row <= d H[-nrows_to_copy:] = A[i:end_dim] i = end_dim return H def parametrized_fd_sketches(A, B, d): # from "Improved Practical Matrix Sketching with Guarantees" A_hat = fast_frequent_directions(A.T, d, variant='parametrized', alpha=.2) B_hat = fast_frequent_directions(B.T, d, variant='parametrized', alpha=.2) return A_hat.T, B_hat.T def fd_amm_sketches(A, B, d): # print("A shape: ", A.shape) # print("B shape: ", B.shape) G = np.hstack((A.T, B)) # D x (N + M) H = frequent_directions(G, d) assert H.shape == (d, A.shape[0] + B.shape[1]) C = H[:, :A.shape[0]] # d x N D = H[:, A.shape[0]:] # d x M return C.T, D def fast_fd_amm_sketches(A, B, d): # print("A shape: ", A.shape) # print("B shape: ", B.shape) G = np.hstack((A.T, B)) # D x (N + M) H = fast_frequent_directions(G, d) assert H.shape == (d, A.shape[0] + B.shape[1]) C = H[:, :A.shape[0]] # d x N D = H[:, A.shape[0]:] # d x M return C.T, D def _nmultiplies_frequent_directions(N, D, d): niters = N - d + 1 iter_svd_cost = _nmultiplies_svd(d, D) iter_reweight_cost = d * D iter_cost = iter_svd_cost + iter_reweight_cost return niters * iter_cost def _nmultiplies_fast_frequent_directions(N, D, d): niters = int(np.ceil(N / d)) iter_svd_cost = _nmultiplies_svd(d, D) iter_reweight_cost = d * D iter_cost = iter_svd_cost + iter_reweight_cost return niters * iter_cost def _nmultiplies_fd_amm_sketches(N, D, M, d): N, D = D, N + M # matrices get concatenated return _nmultiplies_frequent_directions(N, D, d) def test_fd_amm_sketches(): print("test_fd_amm_sketches") N, M, D = 100, 80, 50 np.random.seed(1234) A = np.random.randint(5, size=(N, D)).astype(np.float32) B = np.random.randint(5, size=(D, M)).astype(np.float32) # A -= np.mean(A) # B -= np.mean(B) AB = A @ B orig_frob_sq = np.sum(AB * AB) prev_normed_err = np.inf for d in (1, 2, 4, 8, 16, 32): A_hat, B_hat = fd_amm_sketches(A, B, d) AB_hat = A_hat @ B_hat diffs = AB - AB_hat err_frob_sq = np.sum(diffs * diffs) normed_err_sq = err_frob_sq / orig_frob_sq print('d = {}, err = {:.5f}'.format(d, normed_err_sq)) assert normed_err_sq < 1.05 assert normed_err_sq < prev_normed_err prev_normed_err = normed_err_sq # ================================================================ Co-occurring def cooccur_sketches(A, B, d): N, D = A.shape B = B.T M, _ = B.shape assert B.shape[1] == D # assert N >= d # not enough rows in specified A matrix # assert M >= d # not enough cols in specified B matrix # add new rows to A or B so that R from QR factorization is at least d x d if N < d: A_new = np.zeros((d, D), dtype=A.dtype) A_new[:N] = A A = A_new if M < d: B_new = np.zeros((d, D), dtype=B.dtype) B_new[:M] = B B = B_new X = np.copy(A[:, :d]) # N x d Y = np.copy(B[:, :d]) # M x d # mid_idx = d - 2 # does this make it work better for large d? EDIT: nope mid_idx = d // 2 ntrailing_zeros = d - mid_idx i = d while i < D: Qx, Rx = np.linalg.qr(X) # N x d, d x d Qy, Ry = np.linalg.qr(Y) # M x d, d x d prod = Rx @ Ry.T # d x d U, S, Vt = np.linalg.svd(prod, full_matrices=False) # d x d, d, d x d cutoff = S[mid_idx] S = np.sqrt(np.maximum(S - cutoff, 0)) # print("prod.shape", prod.shape) # print("orig X.shape", X.shape) # print("orig Y.shape", Y.shape) X = Qx @ (U * S) # equivalent to U @ np.diag(S) Y = Qy @ (Vt.T * S) # equivalent to Vt.T @ np.diag(S) # print("X.shape", X.shape) # print("Qx.shape", Qx.shape) # print("U.shape", U.shape) # replace zeroed-out cols of X and Y with new cols of A and B end_dim = min(D, i + ntrailing_zeros) ncols_to_copy = end_dim - i end_col = mid_idx + ncols_to_copy assert ncols_to_copy <= ntrailing_zeros assert end_col <= d X[:, mid_idx:end_col] = A[:, i:end_dim] Y[:, mid_idx:end_col] = B[:, i:end_dim] i = end_dim return X[:N], Y[:M].T # slicing is because we may have zero-padded def _nmultiplies_cooccur_sketches(N, D, M, d): niters = int(np.ceil(D / d)) iter_qr_cost = _nmultiplies_qr(N, d) + _nmultiplies_qr(M, d) iter_RRt_cost = d * d * d iter_svd_cost = _nmultiplies_svd(d, d) iter_reweight_cost = N * d + M * d iter_update_x_y_cost = (N * d * d) + (M * d * d) iter_cost = (iter_qr_cost + iter_RRt_cost + iter_svd_cost + iter_reweight_cost + iter_update_x_y_cost) return niters * iter_cost def test_cooccur_sketches(): print("test_cooccur_sketches") # so this doesn't have monotonically better acc as d increases; seems to # run into issues with d being a large fraction of D, possibly because # then it doesn't have many iterations and it's just zeroing out a ton of # the singular vectors N, M, D = 100, 80, 50 # N, M, D = 100, 80, 160 np.random.seed(1234) A = np.random.randint(5, size=(N, D)).astype(np.float32) B = np.random.randint(5, size=(D, M)).astype(np.float32) A -= np.mean(A) B -= np.mean(B) AB = A @ B orig_frob_sq = np.sum(AB * AB) # prev_normed_err = np.inf # for d in [4]: for d in (2, 4, 8, 16, 32): # A_hat, B_hat = fd_amm_sketches(A, B, d) A_hat, B_hat = cooccur_sketches(A, B, d) AB_hat = A_hat @ B_hat # print("fused mats shapes: ") # print(Ua.shape, SVTa.shape, Ub.shape, SVTb.shape) diffs = AB - AB_hat err_frob_sq = np.sum(diffs * diffs) normed_err_sq = err_frob_sq / orig_frob_sq print('d = {}, err = {:.5f}'.format(d, normed_err_sq)) assert normed_err_sq < 1. # assert normed_err_sq < prev_normed_err # prev_normed_err = normed_err_sq # ================================================================ main # def main(): # pass if __name__ == '__main__': np.set_printoptions(formatter={'float': lambda f: "{:.3}".format(f)}) # test_sketch_sq_sample() # test_svd_sketches() # test_fd_amm_sketches() # test_cooccur_sketches() test_rand_sketches() # # N = 1000 # # N = 100 # N = 20 # # N = 10 # M = 10 # # M = 5 # x = np.arange(N) # # x *= x # # x *= x # # x = np.sqrt(x) # # x = 1.1 ** x # # x = 1.15 ** x # x = 2 ** x # # print("x = ", x) # # idxs, weights = sample_varopt_1d(x, M) # idxs, weights = importance_sample(x, M) # y = x[idxs] * weights # xsum, ysum = x.sum(), y.sum() # # print("idxs = ", idxs) # print("vals = ", x[idxs]) # print("weights = ", weights) # print("vals * weights", y) # # print("true sum, sample sum: ", xsum, ysum) # print("sum rel err: ", (xsum - ysum) / xsum)
#!/bin/env/python import numpy as np def energy(A): if A.ndim < 2 or len(A) < 2: return 0 diffs = A - A.mean(axis=0) return np.sum(diffs * diffs) def run_trial(N=100, D=3, seed=None): if seed is not None: np.random.seed(seed) w0, w = np.random.randn(2, D) X = np.random.randn(N, D) X1 = X[(X @ w) > 0] X2 = X[(X @ w) <= 0] U = X[(X @ w0) > 0] V = X[(X @ w0) <= 0] U1 = U[(U @ w) > 0] U2 = U[(U @ w) <= 0] V1 = V[(V @ w) > 0] V2 = V[(V @ w) <= 0] energy_0 = energy(X) energy_w = energy(X1) + energy(X2) energy_w0 = energy(U) + energy(V) energy_w0_w = energy(U1) + energy(U2) + energy(V1) + energy(V2) gain1 = energy_0 - energy_w gain2 = energy_w0 - energy_w0_w if gain1 < gain2: print("N, D, seed = ", N, D, seed) print("energy_0:", energy_0) print("energy_w:", energy_w) print("energy_w0:", energy_w0) print("energy_w0_w:", energy_w0_w) print("gain1:", gain1) print("gain2:", gain2) print("w0:\n", w0) print("w: \n", w) # print("X\t({:.3f}):\n{}".format(energy(X), X)) # print("X1\t({:.3f}):\n{}".format(energy(X1), X1)) # print("X2\t({:.3f}):\n{}".format(energy(X2), X2)) # print("U\t({:.3f}):\n{}".format(energy(U), U)) # print("U1\t({:.3f}):\n{}".format(energy(U1), U1)) # print("U2\t({:.3f}):\n{}".format(energy(U2), U2)) # print("V\t({:.3f}):\n{}".format(energy(V), V)) # print("V1\t({:.3f}):\n{}".format(energy(V1), V1)) # print("V2\t({:.3f}):\n{}".format(energy(V2), V2)) print("X energy: \t{:.3f}".format(energy(X))) print("X1 energy: \t{:.3f}".format(energy(X1))) print("X2 energy: \t{:.3f}".format(energy(X2))) print("U energy: \t{:.3f}".format(energy(U))) print("U1 energy: \t{:.3f}".format(energy(U1))) print("U2 energy: \t{:.3f}".format(energy(U2))) print("V energy: \t{:.3f}".format(energy(V))) print("V1 energy: \t{:.3f}".format(energy(V1))) print("V2 energy: \t{:.3f}".format(energy(V2))) if D == 2: import matplotlib.pyplot as plt _, axes = plt.subplots(2, 2, figsize=(7.5, 7)) # plt.scatter(X[:, 0], X[:, 1]) for ax in axes.ravel(): ax.set_xlim([-2.5, 2.5]) ax.set_ylim([-2.5, 2.5]) # ax.plot([0, w0[0]], [0, w0[1]]) # ax.plot([0, w[0]], [0, w[1]]) axes[0, 0].set_title("X") axes[0, 0].scatter(X[:, 0], X[:, 1]) axes[0, 1].set_title("U and V (split on w0)") axes[0, 1].plot([0, w0[0]], [0, w0[1]]) axes[0, 1].scatter(U[:, 0], U[:, 1]) axes[0, 1].scatter(V[:, 0], V[:, 1]) axes[1, 0].set_title("X1 and X2 (split on w)") axes[1, 0].plot([0, w[0]], [0, w[1]]) axes[1, 0].scatter(X1[:, 0], X1[:, 1]) axes[1, 0].scatter(X2[:, 0], X2[:, 1]) axes[1, 1].set_title("U1, U2, V1, V2 (split on w0 and w)") axes[1, 1].plot([0, w0[0]], [0, w0[1]]) axes[1, 1].plot([0, w[0]], [0, w[1]]) axes[1, 1].scatter(U1[:, 0], U1[:, 1]) axes[1, 1].scatter(U2[:, 0], U2[:, 1]) axes[1, 1].scatter(V1[:, 0], V1[:, 1]) axes[1, 1].scatter(V2[:, 0], V2[:, 1]) plt.tight_layout() plt.show() assert gain1 >= gain2 def main(): ntrials = 100 # for N in [4, 8, 16, 32, 64, 128, 256]: for N in [64, 128, 256]: # for D in [1, 2, 3, 5, 10, 100]: for D in [100, 200]: for trial in range(ntrials): run_trial(N=N, D=D, seed=trial) if __name__ == '__main__': np.set_printoptions(precision=3) main()
#!/bin/env/python import collections import os import numpy as np import matplotlib as mpl import matplotlib.pyplot as plt import seaborn as sb import pandas as pd import pathlib as pl # from . import files from . import amm_results2 as res # from . import amm_methods as ameth # sb.set_context('poster') # sb.set_context('talk') # sb.set_cmap('tab10') FIGS_SAVE_DIR = pl.Path('../figs/amm') USE_FONT = 'DejaVu Sans' mpl.rcParams['font.family'] = 'sans-serif' mpl.rcParams['font.sans-serif'] = [USE_FONT] # to avoid type3 fonts; 42 = truetype, which is more flexible than type1 mpl.rcParams['pdf.fonttype'] = 42 mpl.rcParams['ps.fonttype'] = 42 def fix_ticks(): # recover from seaborn white style messing this up plt.rcParams['xtick.bottom'] = True plt.rcParams['ytick.left'] = True if not os.path.exists(FIGS_SAVE_DIR): FIGS_SAVE_DIR.mkdir(parents=True) def set_seaborn_style(stylename): sb.set_style(stylename) fix_ticks() def save_fig(name): # plt.savefig(os.path.join(FIGS_SAVE_DIR, name + '.png'), # dpi=300, bbox_inches='tight') plt.savefig(os.path.join(FIGS_SAVE_DIR, name + '.pdf'), bbox_inches='tight') def _xlabel_for_xmetric(x_metric): return {'d': 'Sketch Size', 'secs': 'Time (s)', 'muls': 'Number of Multiplies', 'nlookups': 'Number of Lookups', 'ops': 'Number of Operations', 'Latency': 'Latency (ms)', 'Speedup': 'Speedup Over Exact Matrix Multiply', 'NormalizedTime': 'Normalized Latency', 'Throughput': 'Throughput (elements/s)'}[x_metric] def _ylabel_for_xmetric(y_metric): if y_metric == 'Relative Accuracy': return 'Normalized\nAccuracy' if y_metric == 'Accuracy': return 'Classification\nAccuracy' return y_metric def add_ylabels_on_right(axes, fmt, vals): for i, ax in enumerate(axes): lbl = fmt.format(vals[i]) ax2 = ax.twinx() ax2.get_xaxis().set_visible(False) ax2.yaxis.set_label_position('right') ax2.set_ylabel(lbl, fontsize=14, family=USE_FONT, labelpad=5) sb.despine(ax=ax2, top=True, left=True, bottom=True, right=True) plt.setp(ax2.get_xticklabels(), visible=False) plt.setp(ax2.get_yticklabels(), visible=False) ax2.set_yticks([]) ax2.tick_params(axis='y', which='y', length=0) def scan_speed_fig(save=True): # ================================ data cleaning df = res.scan_timings() name_map = collections.OrderedDict() # name_map['mithral scan'] = 'Mithral' name_map['mithral scan'] = 'MADDNESS' # name_map['mithral scan'] = 'Maddness' # name_map['bolt scan uint8'] = 'Bolt\nCheating' name_map['bolt scan safe uint16'] = 'Bolt' name_map['popcount scan'] = 'Popcount' name_map['pq scan'] = 'PQ / OPQ' df = res.rename_values_in_col(df, 'algo', name_map) df = res.melt_times(df) # alright, can't get stds to show without really screwing with stuff # times = np.array(df['time']) # times += np.random.randn(len(df['time'])) * .1 # get 1px for stds # # mask = df['algo'] == 'PQ / OPQ' # mask = df['B'] == 64 # df['time'].loc[mask] = times[mask] df['thruput'] = df['N'] * df['M'] / df['time'] df['thruput'] /= 1e6 # just use units of billions; times are in ms # df['thruput'] *= 1e3 # times are in ms # ================================ fig creation sb.set_context("talk") # fig, ax = plt.subplots(1, 1, figsize=(8, 5)) fig, ax = plt.subplots(1, 1, figsize=(8, 4)) axes = [ax] sb.barplot(data=df, x='algo', y='thruput', units='timing_trial', hue='B', hue_order=[8, 16, 32, 64], order=name_map.values(), ax=ax, ci='sd') # ------------------------ clean up / format axes for ax in axes[:-1]: # remove x labels except for bottom axis plt.setp(ax.get_xticklabels(), visible=False) ax.get_xaxis().set_visible(False) handles, labels = axes[0].get_legend_handles_labels() labels = ['8B Codes', '16B Codes', '32B Codes', '64B Codes'] # labels = ['8 Bytes', '16 Bytes', '32 Bytes', '64 Bytes'] # labels = ['8B', '16B', '32B', '64B'] plt.figlegend(handles, labels, loc='lower center', ncol=4, fontsize=14) for ax in axes: ax.set_ylabel('Billion Dot Products/s', family=USE_FONT) ax.get_legend().remove() # ------------------------ have bottom / top axes print title, x info axes[0].set_title('Speed of f() Functions for Different Encoding Sizes', y=1.04, family=USE_FONT, fontsize=20) # # get and set them again so we can make the first one bold; can't make # # it bold beforehand because need a tick lbl object, not a string # xlabels = list(axes[-1].get_xticklabels()) # xlabels[0].set_weight('bold') # # axes[-1].set_xticklabels(xlabels, rotation=60, ha='right') # axes[-1].set_xticklabels(xlabels) axes[-1].tick_params(axis='x', which='major', pad=4) axes[-1].set_xlabel("", labelpad=-30) ax.xaxis.set_ticks_position('none') # ------------------------ save / show plot plt.tight_layout() # plt.subplots_adjust(bottom=.21) plt.subplots_adjust(bottom=.23) if save: save_fig('scan_speed') else: plt.show() def encode_speed_fig(save=True): # ================================ data cleaning df = res.encode_timings() df = df.loc[df['algo'] != 'mithral encode i16'] # print("df ours f32: ", df.loc[df['algo'].str.lower().str.strip() == 'mithral encode f32']) # print("df ours f32: ", df.loc[df['algo'].str.lower().str.strip() == 'mithral encode i8']) # print(df) # # # print(df['B']) # # # print(df['C']) # import sys; sys.exit() name_map = collections.OrderedDict() # name_map['mithral encode i8'] = r'$\bf{Mithral}$ $\bf{i8}$') # name_map['mithral encode i8'] = r'$\bf{Mithral}$ $\bf{i8}$') # name_map['mithral encode i8'] = 'Mithral i8' # name_map['mithral encode i16'] = 'Mithral i16' # no i16 in plot # name_map['mithral encode f32'] = 'Mithral f32' # name_map['mithral encode i8'] = 'MADDNESS i8' # name_map['mithral encode f32'] = 'MADDNESS f32' name_map['mithral encode f32'] = 'MADDNESS' name_map['bolt encode'] = 'Bolt' name_map['pq encode'] = 'PQ' name_map['opq encode'] = 'OPQ' df = res.rename_values_in_col(df, 'algo', name_map) df = res.melt_times(df, ntimes=5) order = 'MADDNESS Bolt PQ OPQ'.split() # df['thruput'] = df['N'] * df['D'] / df['time'] # df['thruput'] = df['N'] / (df['time'] * .001) # rows/sec time_secs = (df['time'] * .001) df['elemsz'] = 4 df['elemsz'].loc[df['algo'].str.endswith('i8')] = 1 df['elemsz'].loc[df['algo'].str.endswith('i16')] = 2 df['thruput'] = df['N'] * df['D'] * df['elemsz'] / time_secs # GB/sec df['thruput'] /= 1e9 # convert to GB # df['thruput'] /= 1e6 # just use units of billions; times are in ms # full_byte_per_codebook = df['algo'].isin(['PQ', 'OPQ']) # df['B'] = df['C'].values / 2 # # cvals = df['C'].loc[full_byte_per_codebook] # df['B'].loc[full_byte_per_codebook] = df['C'].loc[full_byte_per_codebook] # df['B'] = df['B'].astype(np.int) # # print("df.cols: ", df.columns) # print(df) # # # print(df['B']) # # # print(df['C']) # import sys; sys.exit() # ================================ fig creation sb.set_context('talk') # sb.set_style('darkgrid') # sb.set_style('white') set_seaborn_style('white') # use_nbytes = [8, 16, 32, 64] use_nbytes = [8, 16, 32] # fig, axes = plt.subplots(len(use_nbytes), 1, figsize=(6, 8), sharey=True) # fig, axes = plt.subplots(len(use_nbytes), 1, figsize=(6, 6.5), sharey=True) # fig, axes = plt.subplots(len(use_nbytes), 1, figsize=(6, 7), sharey=True) fig, axes = plt.subplots(len(use_nbytes), 1, figsize=(6, 6.5), sharey=True) for i, nbytes in enumerate(use_nbytes): data = df.loc[df['B'] == nbytes] # print("df.cols: ", df.columns) # print(data) # # # print(df['B']) # # # print(df['C']) # import sys; sys.exit() order = name_map.values() dashes = {name: ([] if name.lower().startswith('maddness') else mpl.rcParams['lines.dashed_pattern']) for name in order} # dashes = None # sb.lineplot(data=data, x='D', y='thruput', hue='algo', # sb.lineplot(data=data, x='D', y='thruput', hue='algo', units='timing_trial', sb.lineplot(data=data, x='D', y='thruput', hue='algo', # ax=axes[i], ci='sd', estimator=None, hue_order=order, ax=axes[i], ci='sd', estimator='mean', hue_order=order, # ax=axes[i], ci=None, estimator='mean', hue_order=order, style='algo', style_order=order, dashes=dashes, palette=my_colors_list) # import sys; sys.exit() # ------------------------ axis cleanup axes[0].set_title('Speed of g() Functions\nfor Different Encoding Sizes', y=1.04, family=USE_FONT, fontsize=16) handles, labels = axes[0].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm df column name # plt.figlegend(handles, labels, loc='lower center', ncol=3, fontsize=13) plt.figlegend(handles, labels, loc='lower center', ncol=4, fontsize=13) for ax in axes: # ax.semilogx() ax.semilogy() ax.set_ylim([.02, 1000]) # ax.set_yticks([.1, 1, 10, 100, 1000]) ax.set_yticks([.1, 10, 1000]) ax.get_legend().remove() # ax.set_ylabel('Billions of\nScalars Encoded/s', # ax.set_ylabel('Scalars Encoded/s\n(Billions)', # ax.set_ylabel('Scalars Encoded\nper Second (Billions)', # ax.set_ylabel('Scalars Encoded\nper Second', # ax.set_ylabel('Scalars Encoded/s', # ax.set_ylabel('Rows Encoded/s', ax.set_ylabel('Encoding\nSpeed (GB/s)', family=USE_FONT, fontsize=14) for ax in axes[:-1]: # remove x labels except for bottom axis ax.tick_params(axis='x', which='x', length=0) plt.setp(ax.get_xticklabels(), visible=False) ax.set_xlabel("", visible=False) # ax.get_xaxis().set_visible(False) # ax.get_xticklabels().set_visible(False) axes[-1].set_xlabel('Number of Columns in Matrix A', family=USE_FONT, fontsize=14) # add byte counts on the right add_ylabels_on_right(axes, "{}B Encodings", use_nbytes) plt.tight_layout() # plt.subplots_adjust(bottom=.18, hspace=.15) # plt.subplots_adjust(bottom=.19, hspace=.15) plt.subplots_adjust(bottom=.17, hspace=.15) # plt.subplots_adjust(bottom=.21, hspace=.15) if save: save_fig('encode_speed') else: plt.show() def lut_speed_fig(save=True): # ================================ data cleaning df = res.lut_timings() name_map = collections.OrderedDict() # name_map['mithral lut dense'] = '$\bf{Mithral}$' # name_map['mithral lut sparse'] = '$\bf{Mithral}$' name_map['mithral lut dense'] = 'MADDNESS' name_map['mithral lut sparse'] = 'MADDNESS' name_map['bolt lut'] = 'Bolt' name_map['pq lut'] = 'PQ' name_map['opq lut'] = 'OPQ' df = res.rename_values_in_col(df, 'algo', name_map) # print(df[:20]) # df['lutconst'] = df['lutconst'].str.strip().astype(np.float).astype(np.int) # print("df.dtypes", df.dtypes) # import sys; sys.exit() names = list(df['algo']) consts = np.array(df['lutconst']) # print("len(names)", len(names)) # print("len(consts)", len(consts)) mithral_const_to_name = collections.OrderedDict() mithral_const_to_name[-1] = 'MADDNESS, L = ∞' mithral_const_to_name[4] = 'MADDNESS, L = 4' mithral_const_to_name[2] = 'MADDNESS, L = 2' mithral_names = list(mithral_const_to_name.values()) # add lut constant into the name for mithral variations new_names = [] ismithral = [] for i, name in enumerate(names): if not name.startswith('Mithral'): new_names.append(name) ismithral.append(False) continue # const = consts[i] # const = "{:d}".format(int(const)) if const > 0 else "∞" # new_names.append(f"{name}, L = {const}") new_names.append(mithral_const_to_name[int(consts[i])]) ismithral.append(True) # print("len(new_names)", len(new_names)) df['algo'] = new_names df['ismithral'] = ismithral df = res.melt_times(df, ntimes=5) # df = res.melt_times(df, ntimes=3) # TODO rerun with ntrials=5 # print(df) df['thruput'] = df['N'] * df['D'] / df['time'] # df['thruput'] /= 1e6 # just use units of billions; times are in ms # # TODO rm once we have updated results # mask = df['algo'].isin(('PQ', 'OPQ')) # df['B'] = -1 # create placeholder col # df['B'].loc[mask] = df['C'].loc[mask] # df['B'].loc[~mask] = df['C'].loc[~mask] / 2 # ================================ fig creation sb.set_context('talk') # sb.set_style('darkgrid') # sb.set_style('white') set_seaborn_style('white') # use_nbytes = [8, 16, 32, 64] use_nbytes = [8, 16, 32] fig, axes = plt.subplots(len(use_nbytes), 1, figsize=(6, 8), sharey=True) order = [mithral_names[2], 'Bolt', mithral_names[1], 'PQ', mithral_names[0], 'OPQ'] dashes = {k: ('-' if k in mithral_names else '--') for k in order} # dashes = {k: ('solid' if k in mithral_names else 'dashed') for k in order} # dashes = {k: (None if k in mithral_names else [3, 3]) for k in order} # dashes = True # print(dashes) # import sys; sys.exit() for i, nbytes in enumerate(use_nbytes): data = df.loc[df['B'] == nbytes] ax = axes[i] # print(f"------------------------ {nbytes}B") # manual version # for algo in order: # subdf = data.loc[df['algo'] == algo] # print("plotting algo: ", algo) # x = subdf['D'].as_matrix() # y = subdf['thruput'].as_matrix() # sort_idxs = np.argsort(x) # x, y = x[sort_idxs], y[sort_idxs] # ax.plot(x, y, dashes[algo], label=algo) dashes = {name: ([] if name.lower().startswith('mithral') else mpl.rcParams['lines.dashed_pattern']) for name in order} sb.lineplot(data=data, x='D', y='thruput', hue='algo', units='timing_trial', ax=axes[i], ci='sd', estimator=None, hue_order=order, style='algo', style_order=order, dashes=dashes) # sb.lineplot(data=data, x='D', y='thruput', hue='algo', units='timing_trial', # hue_order=order, # # hue_order=order, style='algo', style_order=order, # # dashes=True, # style='ismithral', style_order=[True, False], dashes=True, # ax=axes[i], ci='sd', estimator=None) # ------------------------ axis cleanup axes[0].set_title('Speed of h() Functions\nfor Different Encoding Sizes', y=1.04, family=USE_FONT, fontsize=18) # for ax in axes: # print("ax handles, labels: ") # print(ax.get_legend_handles_labels()) handles, labels = axes[-1].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm df column name # handles, labels = handles[:-3], labels[:-3] # rm ismithral plt.figlegend(handles, labels, loc='lower center', ncol=3, fontsize=13) for ax in axes: # ax.semilogx() ax.semilogy() ax.get_legend().remove() ax.set_ylabel('Scalars Encoded/s', family=USE_FONT, fontsize=14) for ax in axes[:-1]: # remove x labels except for bottom axis ax.tick_params(axis='x', which='x', length=0) plt.setp(ax.get_xticklabels(), visible=False) ax.set_xlabel("", visible=False) axes[-1].set_xlabel('Number of Rows in Matrix B', family=USE_FONT, fontsize=14) # add byte counts on the right add_ylabels_on_right(axes, "{}B Encodings", use_nbytes) plt.tight_layout() plt.subplots_adjust(bottom=.18, hspace=.15) if save: save_fig('lut_speed') else: plt.show() def lotsa_colors_cmap(value): assert 0 <= value <= 1 # if this throws, I don't understand cmaps if value < .3333: return plt.get_cmap('tab20')(3 * value) elif value < .6666: return plt.get_cmap('tab20b')((3 * value) - 1) else: return plt.get_cmap('tab20c')((3 * value) - 2) # def my_tab10(value): # assert 0 <= value <= 1 # value = int(value * 10) # perm = [3, 1, 2, 4, 5, 6, 7, 8, 9] # make red first, then orange # value = perm[value] # return plt.get_cmap('tab10')((value / 10.) + .01) # def my_cmap(value): my_colors_list = (plt.get_cmap('Set1').colors + plt.get_cmap('Set3').colors[:1] # skip light yellow + plt.get_cmap('Set3').colors[2:] + plt.get_cmap('Dark2').colors[:6]) # my_colors_list = my_colors_list[:5] + () my_colors_list[6:] # rm bright yellow # new_yellow = (240./255, 230./255, 140./255) new_yellow = (204. / 255, 204. / 255, 0. / 255) # print(type(my_colors_list)) # print(my_colors_list) my_colors_list = my_colors_list[:5] + (new_yellow,) + my_colors_list[6:] # print(type(my_colors_list)) # print(my_colors_list) # import sys; sys.exit() # DEFAULT_PLOT_METHODS = ('Mithral', 'MithralPQ', 'Brute Force', 'Bolt', # DEFAULT_PLOT_METHODS = ('MADDNESS', 'MADDNESS-PQ', 'Exact', 'Bolt', # 'FastJL', 'HashJL', 'OSNAP', 'PCA', 'SparsePCA', # 'Rademacher', 'RandGauss', 'OrthoGauss') DEFAULT_PLOT_METHODS = ( 'MADDNESS', 'MADDNESS-PQ', 'Exact', 'ScalarQuantize', 'Bolt', # 'MADDNESS', 'Exact', 'ScalarQuantize', 'Bolt', # 'FastJL', 'HashJL', 'PCA', 'RandGauss', 'SparsePCA') 'FastJL', 'HashJL', 'PCA', 'SparsePCA') # 'FastJL', 'HashJL', 'PCA', 'SparsePCA') # 'MADDNESS', 'MADDNESS-PQ', 'Exact', 'Bolt', # 'FastJL', 'HashJL', 'PCA', 'RandGauss', 'SparsePCA') def lineplot(data, ax, x_metric, y_metric, units=None, scatter=False, # plot_methods=None): plot_methods=DEFAULT_PLOT_METHODS, first_two_same_marker=True, **kwargs): estimator = 'mean' if units is None else None if plot_methods is not None: data = data.loc[data['method'].isin(set(plot_methods))] order = plot_methods else: # order = 'Ours Bolt Exact PQ SVD FD-AMM CD'.split() # order = [m for m in order if m in data['Method'].unique()] order = list(data['method'].unique()) # move_methods_to_front = ['Ours', 'OursPQ', 'Brute Force'] # move_methods_to_front = ['Mithral', 'MithralPQ', 'Brute Force'] mithral_methods = [method for method in order # if method.lower().startswith('mithral')][::-1] if method.lower().startswith('maddness')][::-1] move_methods_to_front = mithral_methods[:] # move_methods_to_front.append('Brute Force') move_methods_to_front.append('Exact') for elem in move_methods_to_front[:]: if elem in order: order.remove(elem) else: move_methods_to_front.remove(elem) order = move_methods_to_front + sorted(order) order = [method for method in order if method in data['method'].unique()] # order = plot_methods # order = list(data['method'].unique()) # have to specify markers or seaborn freaks out because it doesn't # have enough of them # filled_markers = ('o', 'v', '^', '<', '>', '8', 's', 'p', '*', 'h', # 'H', 'D', 'd', 'P', 'X') # use_markers = ('*', '*', 's') + ( initial_markers = ('D', 'D', 's') if first_two_same_marker else ('D', 's') use_markers = initial_markers + ( 'o', 'v', '^', '<', '>', '8', 'p', 'h', 'd', 'P', 'X', '*', 'D') if scatter: # sb.violinplot(cut=0, saturation=1, linewidth=.001, scale='width', inner='box', # data['Speedup'] *= 1 + (np.random.randn(len(data['Speedup'])) / 100) sb.scatterplot(alpha=.25, # seems to suck the least data=data, x=x_metric, y=y_metric, hue='method', style='method', style_order=order, hue_order=order, markers=use_markers, estimator=estimator, # units=units, estimator=estimator, markers=use_markers, palette=my_colors_list, ax=ax) # sb.boxplot(linewidth=.1, width=2, whis=999, # sb.stripplot(alpha=.25, orient='v', jitter=False, # data=data, x=x_metric, y=y_metric, hue='method', hue_order=order, # palette=my_colors_list, ax=ax) return kwargs.setdefault('ci', 'sd') sb.lineplot(data=data, x=x_metric, y=y_metric, hue='method', # style='method', style_order=order[::-1], hue_order=order[::-1], style='method', style_order=order, hue_order=order, markers=use_markers, estimator=estimator, # units=units, estimator=estimator, markers=use_markers, dashes=False, palette=my_colors_list, ax=ax, **kwargs) lines = ax.get_lines() for i, line in enumerate(lines): line.set_zorder(10 - i) # def cifar_fig(save=False, x_metric='Throughput', y_metric='Accuracy'): def cifar_fig(save=False, x_metric='Speedup', y_metric='Accuracy'): df10 = res.cifar10_amm() df100 = res.cifar100_amm() sb.set_context('poster') # fig, axes = plt.subplots(2, 1, figsize=(11, 13.5), sharex=True) # fig, axes = plt.subplots(2, 1, figsize=(11, 10), sharex=True) fig, axes = plt.subplots(2, 1, figsize=(11, 8.5), sharex=True) # plot_methods = ['Mithral', 'MithralPQ', 'Brute Force', 'Bolt', # plot_methods = ['MADDNESS', 'MADDNESS-PQ', 'Exact', 'Bolt', # 'FastJL', 'HashJL', 'OSNAP', 'PCA', 'SparsePCA', # 'Rademacher', 'RandGauss', 'OrthoGauss'] # # df10 = df10.loc[df10['method'].isin(set(plot_methods))] # df100 = df100.loc[df100['method'].isin(set(plot_methods))] # df10 = df10.loc[df10['method'] != 'OrthoGauss'] # df100 = df100.loc[df100['method'] != 'OrthoGauss'] lineplot(df10, axes[0], x_metric=x_metric, y_metric=y_metric) lineplot(df100, axes[1], x_metric=x_metric, y_metric=y_metric) # plt.suptitle('Sketch size vs Classification Accuracy') xlbl = _xlabel_for_xmetric(x_metric) # plt.suptitle('{} vs {}'.format(xlbl, y_metric)) plt.suptitle('Approximating Softmax Classifiers', family=USE_FONT) axes[0].set_title('CIFAR-10', family=USE_FONT) for ax in axes: ax.set_ylabel(_ylabel_for_xmetric(y_metric), family=USE_FONT) axes[0].set_xlabel(None) axes[1].set_xlabel(xlbl, family=USE_FONT) axes[1].set_title('CIFAR-100', family=USE_FONT) handles, labels = axes[0].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title # axes[0].legend(handles, labels, fontsize='small') # axes[1].legend(handles, labels, fontsize='small') # plt.figlegend(handles, labels, loc='lower center', ncol=1) # plt.figlegend(handles, labels, loc='center right', ncol=1) for ax in axes.ravel(): ax.get_legend().remove() if y_metric == 'Accuracy': axes[0].set_ylim([.09, .96]) axes[1].set_ylim([.009, .73]) elif y_metric == '1 - NMSE': axes[0].set_ylim([0, 1.02]) axes[1].set_ylim([0, 1.02]) # axes[1].get_legend().remove() # axes[1].get_legend().remove() plt.figlegend(handles, labels, loc='lower center', ncol=3) # if x_metric in ('muls', 'ops', 'nlookups', 'Latency', 'Throughput'): axes[0].semilogx() for ax in axes: if x_metric == 'Speedup': ax.set_xlim([.94, ax.get_xlim()[1]]) elif x_metric == 'NormalizedTime': ax.set_xlim([ax.get_xlim()[0], 1.06]) plt.tight_layout() # plt.subplots_adjust(top=.91, bottom=.24) plt.subplots_adjust(top=.89, bottom=.32) # plt.subplots_adjust(top=.95, bottom=.1) save_fig('cifar_{}_{}'.format(x_metric, y_metric)) # save_fig('cifar_{}_{}_no_maddnesspq'.format(x_metric, y_metric)) def fig1(save=False, x_metric='Speedup', y_metric='Accuracy'): df10 = res.cifar10_amm() df100 = res.cifar100_amm() sb.set_context('poster') fig, axes = plt.subplots(2, 1, figsize=(11, 10), sharex=True) # df10['method'] = df10['method'].str.replace('Mithral', 'HashMul') # replace_names_dict = {'Mithral': 'Ours', replace_names_dict = {'MADDNESS': 'Ours', # 'SparsePCA': '2nd best (Mairal et al.)', # 'HashJL': '3rd best (Dasgupta et al.)', 'SparsePCA': 'Mairal et al.', 'HashJL': 'Dasgupta et al.', 'Exact': 'Exact Matrix Multiply' } # print("--- about to run the rename we care about") df10 = res.rename_values_in_col(df10, 'method', replace_names_dict) df100 = res.rename_values_in_col(df100, 'method', replace_names_dict) # df10['method'] = df10['method'].str.replace(replace_names_dict) # df100['method'] = df100['method'].str.replace(replace_names_dict) # print('df10 methods: ', df10['method'].unique()) # import sys; sys.exit() # plot_methods = ['Ours', '2nd best', '3rd best', 'Exact Matrix Multiply'] # plot_methods = ['Ours', 'Mairal et al.', 'Dasgupta et al.', 'Exact Matrix Multiply'] plot_methods = ['Ours', 'Exact Matrix Multiply', 'Mairal et al.', 'Dasgupta et al.'] # plot_methods = ['Ours', '3rd best', '2nd best', 'Exact Matrix Multiply'] # plot_methods = ['Mithral', 'SparsePCA', 'HashJL', 'Brute Force'] # df10 = df10.loc[df10['method'].isin(set(plot_methods))] # df100 = df100.loc[df100['method'].isin(set(plot_methods))] # df10 = df10.loc[df10['method'] != 'OrthoGauss'] # df100 = df100.loc[df100['method'] != 'OrthoGauss'] lineplot(df10, axes[0], x_metric=x_metric, y_metric=y_metric, plot_methods=plot_methods, ci=None, first_two_same_marker=False) lineplot(df100, axes[1], x_metric=x_metric, y_metric=y_metric, plot_methods=plot_methods, ci=None, first_two_same_marker=False) # plt.suptitle('Sketch size vs Classification Accuracy') xlbl = _xlabel_for_xmetric(x_metric) # plt.suptitle('{} vs {}'.format(xlbl, y_metric)) plt.suptitle('Approximating Softmax Classifiers', family=USE_FONT) axes[0].set_title('CIFAR-10', family=USE_FONT) for ax in axes: ax.set_ylabel(_ylabel_for_xmetric(y_metric), family=USE_FONT) axes[0].set_xlabel(None) axes[1].set_xlabel(xlbl, family=USE_FONT) axes[1].set_title('CIFAR-100', family=USE_FONT) handles, labels = axes[0].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title # axes[0].legend(handles, labels, fontsize='small') # axes[1].legend(handles, labels, fontsize='small') # plt.figlegend(handles, labels, loc='lower center', ncol=1) # plt.figlegend(handles, labels, loc='center right', ncol=1) for ax in axes.ravel(): ax.get_legend().remove() if y_metric == 'Accuracy': axes[0].set_ylim([.09, .96]) axes[1].set_ylim([.009, .73]) elif y_metric == '1 - NMSE': axes[0].set_ylim([0, 1.02]) axes[1].set_ylim([0, 1.02]) # axes[1].get_legend().remove() # axes[1].get_legend().remove() plt.figlegend(handles, labels, loc='lower center', ncol=2) # if x_metric in ('muls', 'ops', 'nlookups', 'Latency', 'Throughput'): axes[0].semilogx() for ax in axes: if x_metric == 'Speedup': ax.set_xlim([.94, ax.get_xlim()[1]]) elif x_metric == 'NormalizedTime': ax.set_xlim([ax.get_xlim()[0], 1.06]) plt.tight_layout() plt.subplots_adjust(top=.89, bottom=.23) save_fig('fig1') def caltech_fig(x_metric='Speedup', y_metric='1 - NMSE'): # df = res.caltech_amm() # df = res.caltech_amm() df0 = res.caltech_amm(filt='sobel') df1 = res.caltech_amm(filt='dog5x5') # print("df cols: ", df.columns) sb.set_context('poster') # fig, ax = plt.subplots(1, 1, figsize=(11, 6)) fig, axes = plt.subplots(2, 1, figsize=(12, 8)) # axes = [ax] # is_mithral = df['method'].str.startswith('Mithral') # is_exact = df['method'] == 'Brute Force' # others_to_keep = df['method'].isin(['Brute Force', 'PCA', 'SparsePCA']) # others_to_keep = df['method'].isin(['PCA', 'SparsePCA']) # df = df.loc[is_mithral | others_to_keep] # others suck too hard # df = df.loc[~(df['method'].isin(['Mithral, L = 2', 'Mithral, L = 4']))] # df['method'].loc[df['method'] == 'Mithral, L = ∞'] = 'Mithral' # print("df0 uniq methods: ", df0['method'].unique()) # print("df1 uniq methods: ", df1['method'].unique()) # import sys; sys.exit() # keep_methods = ['Mithral', 'MithralPQ', 'SparsePCA', 'PCA', 'OSNAP'] # keep_methods = ['Mithral', 'MithralPQ', 'SparsePCA', 'PCA', 'HashJL', 'OSNAP', 'FastJL'] # keep_methods = ['Mithral', 'MithralPQ', 'SparsePCA', 'PCA'] # keep_methods = ['MADDNESS', 'MADDNESS-PQ', 'SparsePCA', 'PCA'] # even scalar quantize is slower than custom exact matmul; note that # in the 5x5 plot, it's occluded by maddness (near perfect mse, but # slightly to the left of 1x speedup) # keep_methods = ['MADDNESS', 'MADDNESS-PQ', 'ScalarQuantize', 'SparsePCA'] keep_methods = ['MADDNESS', 'MADDNESS-PQ', 'SparsePCA'] df0 = df0.loc[df0['method'].isin(keep_methods)] df1 = df1.loc[df1['method'].isin(keep_methods)] # print("df0 kept methods: ", df0['method'].unique()) # print("df1 kept methods: ", df1['method'].unique()) # print("df1 scalar quantize numbers: ", df1.loc[df1['method'] == 'ScalarQuantize']) # import sys; sys.exit() # print("df1:\n", df1.loc[(df1['method'] == 'MithralPQ') & df1['task_id'].str.contains('509')]) # import sys; sys.exit() # lineplot(df, ax, x_metric=x_metric, y_metric=y_metric, units=None) lineplot(df0, axes[0], x_metric=x_metric, y_metric=y_metric, plot_methods=keep_methods) lineplot(df1, axes[1], x_metric=x_metric, y_metric=y_metric, plot_methods=keep_methods) handles, labels = axes[-1].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title # plt.figlegend(handles, labels, loc='lower center', ncol=2) # plt.figlegend(handles, labels, loc='lower center', ncol=4) plt.figlegend(handles, labels, loc='lower center', ncol=len(keep_methods)) # plt.suptitle('Approximating an Image Filter') for ax in axes: ax.set_xlabel(_xlabel_for_xmetric(x_metric), fontsize=20) ax.set_ylabel(y_metric) ax.get_legend().remove() ax.set_ylim([-.01, 1.01]) ax.plot([1, 1], ax.get_ylim(), 'k--') # for ax in axes[:-1]: # # remove x labels except for bottom axis # plt.setp(ax.get_xticklabels(), visible=False) # ax.get_xaxis().set_visible(False) axes[0].set_title('Approximating a Sobel Filter', y=1.02, fontsize=28) axes[1].set_title('Approximating a Gaussian Filter', y=1.02, fontsize=28) # plt.subplots_adjust(top=.91, bottom=.37) plt.tight_layout() # plt.subplots_adjust(bottom=.26, hspace=.72) # with ncol=2 plt.subplots_adjust(bottom=.22, hspace=.7) # with ncol=2 # plt.subplots_adjust(top=.95, bottom=.1) save_fig('caltech_{}_{}'.format(x_metric, '1 - NMSE')) # save_fig('caltech_sobel_{}_{}'.format(x_metric, '1 - NMSE')) # save_fig('caltech_dog_{}_{}'.format(x_metric, '1 - NMSE')) # def ucr_fig(x_metric='Speedup', y_metric='Accuracy'): # def ucr_fig(x_metric='Speedup', y_metric='Change in Accuracy'): def ucr_fig(x_metric='Speedup', y_metric='Relative Accuracy'): # df = res.ucr_amm() # df = res.ucr_amm(k=64) # df = res.ucr_amm(k=128) # df = res.ucr_amm(k=256) df0 = res.ucr_amm(k=64) df1 = res.ucr_amm(k=128) df2 = res.ucr_amm(k=256) sb.set_context('poster') # fig, ax = plt.subplots(1, 1, figsize=(11, 8)) fig, axes = plt.subplots(3, 1, figsize=(12, 13), sharex=True) # axes = [ax] # df = df.loc[df['task_id'].str.lower().str.contains('starlight')] # df = df.loc[df['method'] == 'Mithral'] # # df = df.loc[df['method'] == 'MithralPQ'] # # df = df.loc[df['ncodebooks'] == 4] # df = df['Accuracy acc_orig acc_orig_1nn ncodebooks method task_id'.split() + ['Relative Accuracy']] # df.reset_index(inplace=True, drop=True) # print(df) # import sys; sys.exit() # df['Change in Accuracy'] = df['Accuracy'] - df['acc-1nn-raw'] # print("uniq N, D, M: ") # print(df['N'].unique()) # print(df['D'].unique()) # print(df['M'].unique()) # df_brute = df.loc[df['method'] == 'Brute Force'] # print("uniq times from brute force: ", df_brute['time'].unique()) # print("df Brute:\n", df_brute['N D M method normalized_mse Accuracy time'.split()]) # import sys; sys.exit() # df['acc'] # # TODO put in results cleaning after debug # if 'Accuracy' in df.columns: # # df['Relative Accuracy'] = df['Accuracy'] / (df['acc_orig'] + 1e-20) # # # note that relative accuracy can actually be higher if errors # # # happen to compensate for incorrect classification sometimes # # print("max relative acc: ", df['Relative Accuracy'].values.max()) # # # assert df['Relative Accuracy'].values.max() <= 1.000001 # # acc_orig field is supposed to capture this, but I messed it up for # # 1nn so this will also work # tid2acc = {} # exactdf = df.loc[df['method'] == 'Brute Force'] # for tid in df['task_id'].unique(): # subdf = exactdf.loc[exactdf['task_id'] == tid] # if subdf.shape[0] != 1: # print(f"tid = {tid} gives bad subdf:\n", subdf) # tid2acc[tid] = subdf['Accuracy'].values[0] # df['BaseAccuracy'] = [tid2acc[tid] for tid in df['task_id']] # df['Relative Accuracy'] = df['Accuracy'] / df['BaseAccuracy'] # df = df.loc[~(df['method'].isin(['Mithral, L = 2', 'Mithral, L = 4']))] # # df['method'].loc[df['method'] == 'Mithral, L = ∞'] = 'Mithral' # df0 = df0.loc[df0['method'] != 'Brute Force'] # df1 = df1.loc[df1['method'] != 'Brute Force'] # df2 = df2.loc[df2['method'] != 'Brute Force'] # print(df.columns) # import sys; sys.exit() def clean_df(df): df['Change in Accuracy'] = df['Accuracy'] - df['acc-1nn-raw'] # is_mithral = df['method'].str.startswith('Mithral') # is_mithral = df['method'] == 'Mithral' is_mithral = df['method'] == 'MADDNESS' # # is_exact = df['method'] == 'Brute Force' others_to_keep = df['method'].isin([ 'PCA', 'SparsePCA', 'Bolt', 'HashJL', 'OSNAP']) # others_to_keep = df['method'].isin(['PCA', 'SparsePCA']) return df.loc[is_mithral | others_to_keep] df0 = clean_df(df0) df1 = clean_df(df1) df2 = clean_df(df2) # df = df.loc[df['method'] == 'Brute Force'] # df['not_mse'] = 1. - df['normalized_mse'] # df = df.loc[df['not_mse'] < 2] lineplot(df0, axes[0], x_metric=x_metric, y_metric=y_metric, scatter=True) lineplot(df1, axes[1], x_metric=x_metric, y_metric=y_metric, scatter=True) lineplot(df2, axes[2], x_metric=x_metric, y_metric=y_metric, scatter=True) plt.suptitle('Approximating an RBF Kernel Classifier') axes[-1].set_xlabel(_xlabel_for_xmetric(x_metric)) # ax.set_ylabel('1. - NMSE') handles, labels = axes[-1].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title plt.figlegend(handles, labels, loc='lower center', ncol=3) for ax in axes: ax.set_ylabel(_ylabel_for_xmetric(y_metric)) ax.get_legend().remove() ax.semilogx() ax.set_xlim([.9, ax.get_xlim()[1]]) # ax.set_ylim([.2, 1.1]) # plt.plot([1, 1], ax.get_ylim(), 'k--') plt.tight_layout() plt.subplots_adjust(top=.94, bottom=.25) # plt.subplots_adjust(top=.95, bottom=.1) save_fig('ucr_{}_{}'.format(x_metric, y_metric)) def ucr_fig2(x_metric='Speedup', y_metric='Relative Accuracy', # problem='softmax'): problem='rbf'): # df0 = res.ucr_amm(k=64) # df1 = res.ucr_amm(k=128) # df2 = res.ucr_amm(k=256) df = res.ucr_amm(k=128, problem=problem) sb.set_context('poster') # fig, axes = plt.subplots(3, 1, figsize=(12, 13), sharex=True) fig, axes = plt.subplots(3, 1, figsize=(12, 12), sharex=True) # df = res.ucr_amm(k=128, problem='rbf') # df_bolt = df.loc[df['method'] == 'Bolt'] # print("number of uniq bolt speedups:") # print(df_bolt['Speedup'].unique().size) # import sys; sys.exit() def clean_df(df): df['Change in Accuracy'] = df['Accuracy'] - df['acc-1nn-raw'] return df # # is_mithral = df['method'].str.startswith('Mithral') # is_mithral = df['method'] == 'Mithral' # # # is_exact = df['method'] == 'Brute Force' # others_to_keep = df['method'].isin([ # 'PCA', 'SparsePCA', 'Bolt', 'HashJL', 'OSNAP']) # # others_to_keep = df['method'].isin(['PCA', 'SparsePCA']) # return df.loc[is_mithral | others_to_keep] def frac_above_thresh(df, thresh): return res.frac_above_thresh( df, x_metric, y_metric, 'method', 'task_id', thresh) df = clean_df(df) # df0['frac_above_thresh'] = frac_above_thresh(df, .5) # df_bolt = df.loc[df['method'] == 'Bolt'] # print("number of uniq bolt speedups:") # print(df_bolt['Speedup'].unique().size) # import sys; sys.exit() # df = df.loc[df['method'] == 'SparsePCA'] # print(df.groupby('task_id')['Speedup'].count()) # import sys; sys.exit() y_frac_thresholds = [.5, .75, .95] df0 = frac_above_thresh(df, y_frac_thresholds[0]) df1 = frac_above_thresh(df, y_frac_thresholds[1]) df2 = frac_above_thresh(df, y_frac_thresholds[2]) # # print(df0['frac_above_thresh']) # print(df0) # # for row in df0.iterrows(): # # print(row) # # print(df0.unstack(0)) # print("df cols: ", df.columns) # print("df0 cols: ", df0.columns) # print("uniq methods: ", df['method'].unique()) # df = df.loc[df['method'] == 'Brute Force'] # df['not_mse'] = 1. - df['normalized_mse'] # df = df.loc[df['not_mse'] < 2] ycol = 'frac_above_thresh' lineplot(df0, axes[0], x_metric=x_metric, y_metric=ycol, scatter=False) lineplot(df1, axes[1], x_metric=x_metric, y_metric=ycol, scatter=False) lineplot(df2, axes[2], x_metric=x_metric, y_metric=ycol, scatter=False) kind = 'a Softmax' if problem == 'softmax' else 'an RBF Kernel' plt.suptitle(f'Approximating {kind} Classifier') axes[-1].set_xlabel(_xlabel_for_xmetric(x_metric)) # ax.set_ylabel('1. - NMSE') handles, labels = axes[-1].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title plt.figlegend(handles, labels, loc='lower center', ncol=3) for i, ax in enumerate(axes): # ax.set_ylabel(_ylabel_for_xmetric(y_metric)) # ax.set_ylabel("Fraction of Datasets\nWith Relative Acc > " # f"{y_frac_thresholds[i]}") # ax.set_ylabel(f"Fraction with Relative\nAccuracy> {y_frac_thresholds[i]}") ax.set_ylabel(f"Fraction > {y_frac_thresholds[i]}") ax.get_legend().remove() ax.semilogx() ax.set_xlim([.9, ax.get_xlim()[1]]) ax.set_ylim([0, 1.03]) # ax.set_ylim([.2, 1.1]) # plt.plot([1, 1], ax.get_ylim(), 'k--') plt.tight_layout() # plt.subplots_adjust(top=.94, bottom=.25) plt.subplots_adjust(top=.94, bottom=.22) # plt.subplots_adjust(top=.95, bottom=.1) save_fig('ucr2_{}_{}_{}'.format(x_metric, y_metric, problem)) def main(): scan_speed_fig() encode_speed_fig() lut_speed_fig() fig1() ucr_fig2() caltech_fig() # caltech_fig(y_metric='1 - NMSE') # caltech_fig(x_metric='ops', y_metric='1 - NMSE') cifar_fig() # cifar_fig(y_metric='1 - NMSE') # cifar_fig(x_metric='ops') # cifar_fig(x_metric='ops', y_metric='1 - NMSE') # ucr_fig2(x_metric='ops', y_metric='1 - NMSE') # ucr_fig2(x_metric='ops') # cifar_fig(y_metric='1 - NMSE') # ucr_fig2() # ucr_fig2(y_metric='1 - NMSE') if __name__ == '__main__': main()
#!/bin/env/python from . import amm, vq_amm METHOD_EXACT = 'Exact' METHOD_SCALAR_QUANTIZE = 'ScalarQuantize' METHOD_SKETCH_SQ_SAMPLE = 'SketchSqSample' METHOD_SVD = 'SVD' # truncated SVD run on the matrix at test time METHOD_FD_AMM = 'FD-AMM' METHOD_COOCCUR = 'CooccurSketch' METHOD_PCA = 'PCA' # PCA projection, with PCA basis learned at train time METHOD_SPARSE_PCA = 'SparsePCA' # like above, using sklearn SparsePCA METHOD_RANDGAUSS = 'RandGauss' METHOD_ORTHOGAUSS = 'OrthoGauss' METHOD_HADAMARD = 'Hadamard' METHOD_RADEMACHER = 'Rademacher' METHOD_FASTJL = 'FastJL' METHOD_HASHJL = 'HashJL' METHOD_OSNAP = 'OSNAP' METHOD_OPQ = 'OPQ' METHOD_BOLT = 'Bolt' METHOD_BOLT_PERM = 'Bolt+Perm' METHOD_BOLT_CORRPERM = 'Bolt+CorrPerm' METHOD_BOLT_SPLITS = 'BoltSplits' METHOD_BOLT_MULTISPLITS = 'Bolt+MultiSplits' METHOD_BOLT_PERM_MULTISPLITS = 'Bolt+Perm+MultiSplits' METHOD_PQ = 'PQ' METHOD_PQ_PERM = 'PQ+Perm' METHOD_PQ_MULTISPLITS = 'PQ+MultiSplits' METHOD_PQ_PERM_MULTISPLITS = 'PQ+Perm+MultiSplits' METHOD_MITHRALPQ = 'MithralPQ' METHOD_OLD_MITHRALPQ = 'OldMithralPQ' METHOD_MITHRAL = 'Mithral' # these are for trying out different perm options METHOD_BOLT_GEHT_COV_TOPK = 'Bolt_CovTopk' METHOD_BOLT_GEHT_COV_SAMP = 'Bolt_CovSamp' METHOD_BOLT_GEHT_COR_TOPK = 'Bolt_CorTopk' METHOD_BOLT_GEHT_COR_SAMP = 'Bolt_CorSamp' # DEFAULT_METHODS = (METHOD_EXACT, METHOD_SVD, METHOD_FD_AMM, # METHOD_COOCCUR, METHOD_PCA, METHOD_PQ, # METHOD_BOLT, METHOD_MITHRALPQ) METHOD_TO_ESTIMATOR = { METHOD_EXACT: amm.ExactMatMul, METHOD_SCALAR_QUANTIZE: amm.QuantizedMatmul, METHOD_SKETCH_SQ_SAMPLE: amm.SketchSqSample, METHOD_SVD: amm.SvdSketch, METHOD_FD_AMM: amm.FdAmm, METHOD_COOCCUR: amm.CooccurSketch, METHOD_PCA: amm.TrainedPcaSketch, METHOD_SPARSE_PCA: amm.TrainedSparsePcaSketch, METHOD_RANDGAUSS: amm.RandGaussSketch, METHOD_ORTHOGAUSS: amm.RandOrthoGaussSketch, METHOD_HADAMARD: amm.HadamardSketch, METHOD_RADEMACHER: amm.RandRademacherSketch, METHOD_FASTJL: amm.FastJlSketch, METHOD_HASHJL: amm.HashJlSketch, METHOD_OSNAP: amm.OsnapSketch, METHOD_PQ: vq_amm.PQMatmul, METHOD_BOLT: vq_amm.BoltMatmul, METHOD_OPQ: vq_amm.OPQMatmul, METHOD_BOLT_CORRPERM: vq_amm.GEHTBoltMatmul_CorrTopk, METHOD_BOLT_GEHT_COV_TOPK: vq_amm.GEHTBoltMatmul_CovTopk, METHOD_BOLT_GEHT_COV_SAMP: vq_amm.GEHTBoltMatmul_CovSamp, METHOD_BOLT_GEHT_COR_TOPK: vq_amm.GEHTBoltMatmul_CorrTopk, METHOD_BOLT_GEHT_COR_SAMP: vq_amm.GEHTBoltMatmul_CorrSamp, METHOD_BOLT_PERM: vq_amm.GEHTBoltMatmul_CovTopk, METHOD_BOLT_SPLITS: vq_amm.BoltSplits, METHOD_BOLT_MULTISPLITS: vq_amm.BoltMultiSplits, METHOD_BOLT_PERM_MULTISPLITS: vq_amm.BoltPermMultiSplits, METHOD_PQ_PERM: vq_amm.PQPerm, METHOD_PQ_MULTISPLITS: vq_amm.PQMultiSplits, METHOD_PQ_PERM_MULTISPLITS: vq_amm.PQPermMultiSplits, METHOD_OLD_MITHRALPQ: vq_amm.OldMithralPQ, METHOD_MITHRALPQ: vq_amm.MithralPQ, METHOD_MITHRAL: vq_amm.MithralMatmul } ALL_METHODS = sorted(list(METHOD_TO_ESTIMATOR.keys())) ALL_METHODS.remove(METHOD_SKETCH_SQ_SAMPLE), # always terrible results ALL_METHODS.remove(METHOD_OPQ) # takes forever to train, more muls than exact # these were just for playing with different permuation options ALL_METHODS.remove(METHOD_BOLT_GEHT_COV_TOPK) ALL_METHODS.remove(METHOD_BOLT_GEHT_COV_SAMP) ALL_METHODS.remove(METHOD_BOLT_GEHT_COR_TOPK) ALL_METHODS.remove(METHOD_BOLT_GEHT_COR_SAMP) RANDOM_SKETCHING_METHODS = ( METHOD_FASTJL, METHOD_HASHJL, METHOD_OSNAP, METHOD_RANDGAUSS, METHOD_ORTHOGAUSS, METHOD_RADEMACHER) DENSE_SKETCH_METHODS = (METHOD_PCA, METHOD_FASTJL, METHOD_RANDGAUSS, METHOD_HADAMARD, METHOD_ORTHOGAUSS, METHOD_RADEMACHER) FAST_SKETCH_METHODS = RANDOM_SKETCHING_METHODS + ( METHOD_HADAMARD, METHOD_PCA, METHOD_SPARSE_PCA) # SLOW_SKETCH_METHODS = (METHOD_SVD, METHOD_FD_AMM, METHOD_COOCCUR) SLOW_SKETCH_METHODS = (METHOD_FD_AMM, METHOD_COOCCUR, METHOD_SVD) SKETCH_METHODS = FAST_SKETCH_METHODS + SLOW_SKETCH_METHODS # VQ_METHODS = (METHOD_PQ, METHOD_BOLT, METHOD_OPQ) # VQ_METHODS = (METHOD_PQ, METHOD_BOLT) BOLT_METHODS = (METHOD_BOLT, METHOD_BOLT_PERM, METHOD_BOLT_CORRPERM, METHOD_BOLT_SPLITS, METHOD_BOLT_MULTISPLITS, METHOD_BOLT_PERM_MULTISPLITS) PQ_METHODS = (METHOD_PQ, METHOD_PQ_PERM, METHOD_PQ_MULTISPLITS, METHOD_PQ_PERM_MULTISPLITS) MITHRAL_METHODS = (METHOD_MITHRALPQ, METHOD_MITHRAL, METHOD_OLD_MITHRALPQ) VQ_METHODS = PQ_METHODS + BOLT_METHODS + MITHRAL_METHODS NONDETERMINISTIC_METHODS = (METHOD_SKETCH_SQ_SAMPLE, METHOD_SVD) + VQ_METHODS # USE_METHODS = (FAST_SKETCH_METHODS + # (METHOD_EXACT, METHOD_BOLT, METHOD_MITHRALPQ, METHOD_MITHRAL)) USE_METHODS = ((METHOD_EXACT, METHOD_BOLT, METHOD_MITHRALPQ, METHOD_MITHRAL) + FAST_SKETCH_METHODS) USE_CALTECH_METHODS = list(USE_METHODS) USE_CALTECH_METHODS.remove(METHOD_BOLT) # Bolt *can't* be faster
#!/usr/bin/env python from __future__ import print_function import numpy as np import pprint microbench_output = \ """ ncodebooks = 4 amm bolt N, D, M, ncodebooks: 10000, 512, 10, 4 (5x20): 7.574 (4.225e+07/s), 7.582 (4.221e+07/s), 7.584 (4.219e+07/s), 7.587 (4.218e+07/s), 7.579 (4.222e+07/s), amm bolt N, D, M, ncodebooks: 10000, 512, 100, 4 (5x20): 7.747 (1.652e+08/s), 7.743 (1.653e+08/s), 7.757 (1.650e+08/s), 7.758 (1.650e+08/s), 7.743 (1.653e+08/s), amm bolt N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 26.029 (2.749e+08/s), 26.028 (2.749e+08/s), 26.013 (2.751e+08/s), 26.010 (2.751e+08/s), 26.063 (2.745e+08/s), amm bolt N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 1.931 (8.167e+08/s), 1.924 (8.197e+08/s), 1.925 (8.193e+08/s), 1.925 (8.193e+08/s), 1.929 (8.176e+08/s), ncodebooks = 8 amm bolt N, D, M, ncodebooks: 10000, 512, 10, 8 (5x20): 6.912 (4.630e+07/s), 6.919 (4.625e+07/s), 6.912 (4.630e+07/s), 6.909 (4.632e+07/s), 6.911 (4.630e+07/s), amm bolt N, D, M, ncodebooks: 10000, 512, 100, 8 (5x20): 7.169 (1.785e+08/s), 7.207 (1.776e+08/s), 7.200 (1.778e+08/s), 7.205 (1.777e+08/s), 7.205 (1.777e+08/s), amm bolt N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 24.550 (2.914e+08/s), 24.514 (2.919e+08/s), 24.485 (2.922e+08/s), 24.470 (2.924e+08/s), 24.474 (2.923e+08/s), amm bolt N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 2.445 (6.450e+08/s), 2.454 (6.427e+08/s), 2.436 (6.474e+08/s), 2.448 (6.442e+08/s), 2.446 (6.448e+08/s), ncodebooks = 16 amm bolt N, D, M, ncodebooks: 10000, 512, 10, 16 (5x20): 6.350 (5.039e+07/s), 6.350 (5.039e+07/s), 6.347 (5.042e+07/s), 6.356 (5.035e+07/s), 6.438 (4.970e+07/s), amm bolt N, D, M, ncodebooks: 10000, 512, 100, 16 (5x20): 7.340 (1.744e+08/s), 7.270 (1.761e+08/s), 7.302 (1.753e+08/s), 7.277 (1.759e+08/s), 7.299 (1.754e+08/s), amm bolt N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 28.217 (2.536e+08/s), 28.063 (2.550e+08/s), 28.082 (2.548e+08/s), 28.086 (2.547e+08/s), 28.070 (2.549e+08/s), amm bolt N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 3.525 (4.474e+08/s), 3.529 (4.469e+08/s), 3.525 (4.474e+08/s), 3.530 (4.468e+08/s), 3.527 (4.471e+08/s), ncodebooks = 32 amm bolt N, D, M, ncodebooks: 10000, 512, 10, 32 (5x20): 6.036 (5.302e+07/s), 6.070 (5.272e+07/s), 6.085 (5.259e+07/s), 6.158 (5.196e+07/s), 6.176 (5.181e+07/s), amm bolt N, D, M, ncodebooks: 10000, 512, 100, 32 (5x20): 7.473 (1.713e+08/s), 7.478 (1.712e+08/s), 7.571 (1.691e+08/s), 7.567 (1.692e+08/s), 7.571 (1.691e+08/s), amm bolt N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 36.693 (1.950e+08/s), 36.721 (1.948e+08/s), 36.753 (1.947e+08/s), 36.805 (1.944e+08/s), 37.216 (1.923e+08/s), ncodebooks = 64 amm bolt N, D, M, ncodebooks: 10000, 512, 10, 64 (5x20): 6.962 (4.596e+07/s), 6.959 (4.598e+07/s), 6.954 (4.602e+07/s), 6.959 (4.598e+07/s), 6.964 (4.595e+07/s), amm bolt N, D, M, ncodebooks: 10000, 512, 100, 64 (5x20): 8.539 (1.499e+08/s), 8.598 (1.489e+08/s), 8.484 (1.509e+08/s), 8.572 (1.493e+08/s), 8.527 (1.501e+08/s), amm bolt N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 64.087 (1.116e+08/s), 64.096 (1.116e+08/s), 64.638 (1.107e+08/s), 64.099 (1.116e+08/s), 64.079 (1.117e+08/s), ncodebooks = 4 ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 10, 4 (5x20): 0.021 (4.770e+09/s), 0.021 (4.770e+09/s), 0.021 (4.770e+09/s), 0.021 (4.770e+09/s), 0.021 (4.770e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 10, 4 (5x20): 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 10, 4 (5x20): 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 100, 4 (5x20): 0.077 (1.301e+10/s), 0.077 (1.301e+10/s), 0.076 (1.318e+10/s), 0.080 (1.252e+10/s), 0.077 (1.301e+10/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 100, 4 (5x20): 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.016 (1.252e+09/s), 0.017 (1.178e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 100, 4 (5x20): 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), 0.000 (inf/s), ---- f32 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.999 (2.686e+09/s), 0.974 (2.755e+09/s), 1.001 (2.681e+09/s), 1.000 (2.683e+09/s), 0.999 (2.686e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.614 (7.284e+08/s), 0.611 (7.320e+08/s), 0.598 (7.479e+08/s), 0.613 (7.296e+08/s), 0.601 (7.441e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), ---- i16 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.604 (4.443e+09/s), 0.603 (4.450e+09/s), 0.579 (4.635e+09/s), 0.604 (4.443e+09/s), 0.605 (4.435e+09/s), i16 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.257 (1.740e+09/s), 0.280 (1.597e+09/s), 0.265 (1.688e+09/s), 0.254 (1.761e+09/s), 0.254 (1.761e+09/s), i16 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 4 (5x20): 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), 0.024 (1.863e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.083 (1.188e+09/s), 0.083 (1.188e+09/s), 0.085 (1.160e+09/s), 0.084 (1.174e+09/s), 0.084 (1.174e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.077 (1.281e+09/s), 0.077 (1.281e+09/s), 0.076 (1.298e+09/s), 0.076 (1.298e+09/s), 0.076 (1.298e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), ---- i8 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.034 (2.901e+09/s), 0.029 (3.401e+09/s), 0.029 (3.401e+09/s), 0.030 (3.287e+09/s), 0.030 (3.287e+09/s), i8 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.023 (4.288e+09/s), 0.023 (4.288e+09/s), 0.023 (4.288e+09/s), 0.023 (4.288e+09/s), 0.023 (4.288e+09/s), i8 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 4 (5x20): 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), 0.004 (2.466e+10/s), ncodebooks = 8 ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 10, 8 (5x20): 0.043 (2.329e+09/s), 0.043 (2.329e+09/s), 0.043 (2.329e+09/s), 0.043 (2.329e+09/s), 0.043 (2.329e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 10, 8 (5x20): 0.031 (1.292e+09/s), 0.032 (1.252e+09/s), 0.033 (1.214e+09/s), 0.034 (1.178e+09/s), 0.034 (1.178e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 10, 8 (5x20): 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 100, 8 (5x20): 0.154 (6.504e+09/s), 0.162 (6.183e+09/s), 0.155 (6.462e+09/s), 0.155 (6.462e+09/s), 0.162 (6.183e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 100, 8 (5x20): 0.035 (1.145e+09/s), 0.033 (1.214e+09/s), 0.032 (1.252e+09/s), 0.034 (1.178e+09/s), 0.034 (1.178e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 100, 8 (5x20): 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), 0.001 (4.006e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 1.810 (1.483e+09/s), 1.790 (1.499e+09/s), 1.797 (1.493e+09/s), 1.809 (1.483e+09/s), 1.810 (1.483e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 1.395 (6.412e+08/s), 1.371 (6.524e+08/s), 1.394 (6.417e+08/s), 1.394 (6.417e+08/s), 1.393 (6.421e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), ---- i16 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 1.102 (2.435e+09/s), 1.106 (2.426e+09/s), 1.091 (2.460e+09/s), 1.101 (2.437e+09/s), 1.129 (2.377e+09/s), i16 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 0.681 (1.313e+09/s), 0.653 (1.370e+09/s), 0.654 (1.368e+09/s), 0.653 (1.370e+09/s), 0.653 (1.370e+09/s), i16 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 8 (5x20): 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), 0.041 (2.182e+10/s), 0.043 (2.080e+10/s), 0.043 (2.080e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.173 (5.701e+08/s), 0.172 (5.734e+08/s), 0.173 (5.701e+08/s), 0.185 (5.331e+08/s), 0.173 (5.701e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.160 (1.233e+09/s), 0.176 (1.121e+09/s), 0.185 (1.066e+09/s), 0.165 (1.195e+09/s), 0.161 (1.225e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), ---- i8 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.059 (1.672e+09/s), 0.059 (1.672e+09/s), 0.059 (1.672e+09/s), 0.059 (1.672e+09/s), 0.059 (1.672e+09/s), i8 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.049 (4.025e+09/s), 0.050 (3.945e+09/s), 0.049 (4.025e+09/s), 0.048 (4.109e+09/s), 0.048 (4.109e+09/s), i8 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 8 (5x20): 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), 0.008 (2.466e+10/s), ncodebooks = 16 ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 10, 16 (5x20): 0.094 (1.066e+09/s), 0.093 (1.077e+09/s), 0.100 (1.002e+09/s), 0.100 (1.002e+09/s), 0.097 (1.033e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 10, 16 (5x20): 0.065 (1.233e+09/s), 0.066 (1.214e+09/s), 0.066 (1.214e+09/s), 0.065 (1.233e+09/s), 0.066 (1.214e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 10, 16 (5x20): 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 100, 16 (5x20): 0.367 (2.729e+09/s), 0.372 (2.692e+09/s), 0.374 (2.678e+09/s), 0.377 (2.657e+09/s), 0.374 (2.678e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 100, 16 (5x20): 0.067 (1.196e+09/s), 0.064 (1.252e+09/s), 0.064 (1.252e+09/s), 0.064 (1.252e+09/s), 0.064 (1.252e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 100, 16 (5x20): 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), 0.003 (2.671e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 3.597 (7.460e+08/s), 3.607 (7.439e+08/s), 3.599 (7.456e+08/s), 3.610 (7.433e+08/s), 3.614 (7.425e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 2.761 (6.479e+08/s), 2.761 (6.479e+08/s), 2.760 (6.482e+08/s), 2.751 (6.503e+08/s), 2.763 (6.475e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 0.103 (1.737e+10/s), 0.105 (1.704e+10/s), 0.123 (1.454e+10/s), 0.128 (1.398e+10/s), 0.123 (1.454e+10/s), ---- i16 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 2.233 (1.202e+09/s), 2.261 (1.187e+09/s), 2.207 (1.216e+09/s), 2.207 (1.216e+09/s), 2.261 (1.187e+09/s), i16 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 1.417 (1.262e+09/s), 1.563 (1.145e+09/s), 1.514 (1.182e+09/s), 1.464 (1.222e+09/s), 1.483 (1.206e+09/s), i16 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 16 (5x20): 0.136 (1.315e+10/s), 0.130 (1.376e+10/s), 0.147 (1.217e+10/s), 0.133 (1.345e+10/s), 0.134 (1.335e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.397 (2.484e+08/s), 0.407 (2.423e+08/s), 0.395 (2.497e+08/s), 0.388 (2.542e+08/s), 0.385 (2.562e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.369 (1.069e+09/s), 0.368 (1.072e+09/s), 0.377 (1.046e+09/s), 0.375 (1.052e+09/s), 0.408 (9.669e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), ---- i8 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.131 (7.529e+08/s), 0.131 (7.529e+08/s), 0.131 (7.529e+08/s), 0.131 (7.529e+08/s), 0.131 (7.529e+08/s), i8 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.103 (3.830e+09/s), 0.103 (3.830e+09/s), 0.103 (3.830e+09/s), 0.103 (3.830e+09/s), 0.104 (3.793e+09/s), i8 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 16 (5x20): 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), 0.019 (2.076e+10/s), ncodebooks = 32 ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 10, 32 (5x20): 0.201 (4.983e+08/s), 0.194 (5.163e+08/s), 0.205 (4.886e+08/s), 0.201 (4.983e+08/s), 0.200 (5.008e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 10, 32 (5x20): 0.142 (1.129e+09/s), 0.143 (1.121e+09/s), 0.144 (1.113e+09/s), 0.142 (1.129e+09/s), 0.161 (9.954e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 10, 32 (5x20): 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 100, 32 (5x20): 0.762 (1.314e+09/s), 0.781 (1.282e+09/s), 0.756 (1.325e+09/s), 0.753 (1.330e+09/s), 0.798 (1.255e+09/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 100, 32 (5x20): 0.183 (8.757e+08/s), 0.149 (1.076e+09/s), 0.154 (1.041e+09/s), 0.150 (1.068e+09/s), 0.147 (1.090e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 100, 32 (5x20): 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), 0.007 (2.289e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 7.958 (3.372e+08/s), 7.142 (3.757e+08/s), 7.148 (3.754e+08/s), 7.114 (3.772e+08/s), 7.135 (3.761e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 5.589 (6.402e+08/s), 5.642 (6.341e+08/s), 5.563 (6.432e+08/s), 5.592 (6.398e+08/s), 5.579 (6.413e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 0.341 (1.049e+10/s), 0.330 (1.084e+10/s), 0.327 (1.094e+10/s), 0.327 (1.094e+10/s), 0.328 (1.091e+10/s), ---- i16 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 4.369 (6.142e+08/s), 4.357 (6.159e+08/s), 4.537 (5.914e+08/s), 4.361 (6.153e+08/s), 4.406 (6.090e+08/s), i16 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 2.888 (1.239e+09/s), 2.889 (1.238e+09/s), 2.898 (1.235e+09/s), 2.898 (1.235e+09/s), 2.909 (1.230e+09/s), i16 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 32 (5x20): 0.329 (1.087e+10/s), 0.326 (1.098e+10/s), 0.331 (1.081e+10/s), 0.328 (1.091e+10/s), 0.345 (1.037e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.781 (1.263e+08/s), 0.785 (1.256e+08/s), 0.793 (1.244e+08/s), 0.788 (1.252e+08/s), 0.787 (1.253e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.814 (9.693e+08/s), 0.828 (9.529e+08/s), 0.755 (1.045e+09/s), 0.766 (1.030e+09/s), 0.768 (1.027e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.045 (1.753e+10/s), 0.041 (1.924e+10/s), 0.041 (1.924e+10/s), 0.046 (1.715e+10/s), 0.041 (1.924e+10/s), ---- i8 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.320 (3.082e+08/s), 0.303 (3.255e+08/s), 0.301 (3.277e+08/s), 0.321 (3.072e+08/s), 0.301 (3.277e+08/s), i8 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.279 (2.828e+09/s), 0.260 (3.035e+09/s), 0.263 (3.000e+09/s), 0.221 (3.570e+09/s), 0.242 (3.260e+09/s), i8 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 32 (5x20): 0.061 (1.293e+10/s), 0.044 (1.793e+10/s), 0.041 (1.924e+10/s), 0.041 (1.924e+10/s), 0.040 (1.972e+10/s), ncodebooks = 64 ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 10, 64 (5x20): 0.454 (2.206e+08/s), 0.497 (2.015e+08/s), 0.489 (2.048e+08/s), 0.486 (2.061e+08/s), 0.457 (2.192e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 10, 64 (5x20): 0.349 (9.184e+08/s), 0.344 (9.317e+08/s), 0.385 (8.325e+08/s), 0.377 (8.502e+08/s), 0.344 (9.317e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 10, 64 (5x20): 0.019 (1.687e+10/s), 0.019 (1.687e+10/s), 0.019 (1.687e+10/s), 0.019 (1.687e+10/s), 0.020 (1.603e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 10000, 512, 100, 64 (5x20): 1.586 (6.315e+08/s), 1.530 (6.546e+08/s), 1.531 (6.542e+08/s), 1.529 (6.551e+08/s), 1.539 (6.508e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 10000, 512, 100, 64 (5x20): 0.405 (7.914e+08/s), 0.408 (7.856e+08/s), 0.449 (7.138e+08/s), 0.403 (7.953e+08/s), 0.411 (7.798e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 10000, 512, 100, 64 (5x20): 0.020 (1.603e+10/s), 0.020 (1.603e+10/s), 0.019 (1.687e+10/s), 0.019 (1.687e+10/s), 0.019 (1.687e+10/s), ---- f32 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 14.943 (1.796e+08/s), 15.205 (1.765e+08/s), 14.912 (1.799e+08/s), 14.951 (1.795e+08/s), 14.981 (1.791e+08/s), f32 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 11.376 (6.290e+08/s), 11.305 (6.330e+08/s), 11.313 (6.325e+08/s), 11.315 (6.324e+08/s), 11.312 (6.326e+08/s), f32 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 0.877 (8.159e+09/s), 0.822 (8.705e+09/s), 0.845 (8.468e+09/s), 0.849 (8.428e+09/s), 0.836 (8.559e+09/s), ---- i16 amm mithral N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 9.459 (2.837e+08/s), 9.458 (2.837e+08/s), 9.420 (2.849e+08/s), 9.457 (2.837e+08/s), 9.452 (2.839e+08/s), i16 amm mithral enc N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 5.819 (1.230e+09/s), 5.820 (1.230e+09/s), 5.824 (1.229e+09/s), 5.845 (1.224e+09/s), 5.901 (1.213e+09/s), i16 amm mithral zipb N, D, M, ncodebooks: 223590, 96, 12, 64 (5x20): 0.818 (8.748e+09/s), 0.823 (8.695e+09/s), 0.803 (8.911e+09/s), 0.818 (8.748e+09/s), 0.851 (8.409e+09/s), ---- f32 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 1.571 (6.278e+07/s), 1.571 (6.278e+07/s), 1.573 (6.270e+07/s), 1.574 (6.266e+07/s), 1.571 (6.278e+07/s), f32 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 1.479 (1.067e+09/s), 1.473 (1.071e+09/s), 1.475 (1.070e+09/s), 1.476 (1.069e+09/s), 1.473 (1.071e+09/s), f32 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 0.114 (1.384e+10/s), 0.115 (1.372e+10/s), 0.115 (1.372e+10/s), 0.110 (1.435e+10/s), 0.115 (1.372e+10/s), ---- i8 amm mithral N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 0.561 (1.758e+08/s), 0.560 (1.761e+08/s), 0.561 (1.758e+08/s), 0.560 (1.761e+08/s), 0.560 (1.761e+08/s), i8 amm mithral enc N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 0.453 (3.483e+09/s), 0.492 (3.207e+09/s), 0.470 (3.357e+09/s), 0.464 (3.401e+09/s), 0.494 (3.194e+09/s), i8 amm mithral zipb N, D, M, ncodebooks: 49284, 27, 2, 64 (5x20): 0.114 (1.384e+10/s), 0.120 (1.315e+10/s), 0.116 (1.360e+10/s), 0.114 (1.384e+10/s), 0.114 (1.384e+10/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 2 (5x20): 3.827 (2.613e+07/s), 3.815 (2.621e+07/s), 3.830 (2.611e+07/s), 3.858 (2.592e+07/s), 3.832 (2.610e+07/s), our sketch matmul N, D, M, d: 10000, 512, 10, 2 (5x20): 1.080 (9.259e+07/s), 1.041 (9.606e+07/s), 1.049 (9.533e+07/s), 1.049 (9.533e+07/s), 1.045 (9.569e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 4 (5x20): 3.505 (2.853e+07/s), 3.568 (2.803e+07/s), 3.541 (2.824e+07/s), 3.431 (2.915e+07/s), 3.234 (3.092e+07/s), our sketch matmul N, D, M, d: 10000, 512, 10, 4 (5x20): 2.081 (4.805e+07/s), 2.135 (4.684e+07/s), 2.083 (4.801e+07/s), 2.077 (4.815e+07/s), 2.079 (4.810e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 8 (5x20): 3.772 (2.651e+07/s), 3.641 (2.746e+07/s), 3.617 (2.765e+07/s), 3.616 (2.765e+07/s), 4.002 (2.499e+07/s), our sketch matmul N, D, M, d: 10000, 512, 10, 8 (5x20): 2.864 (3.492e+07/s), 2.861 (3.495e+07/s), 2.901 (3.447e+07/s), 3.017 (3.315e+07/s), 2.880 (3.472e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 16 (5x20): 4.535 (2.205e+07/s), 4.565 (2.191e+07/s), 4.475 (2.235e+07/s), 4.476 (2.234e+07/s), 4.480 (2.232e+07/s), our sketch matmul N, D, M, d: 10000, 512, 10, 16 (5x20): 5.217 (1.917e+07/s), 5.185 (1.929e+07/s), 5.243 (1.907e+07/s), 5.256 (1.903e+07/s), 5.184 (1.929e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 32 (5x20): 6.537 (1.530e+07/s), 6.527 (1.532e+07/s), 6.517 (1.534e+07/s), 6.507 (1.537e+07/s), 6.512 (1.536e+07/s), our sketch matmul N, D, M, d: 10000, 512, 10, 32 (5x20): 9.143 (1.094e+07/s), 9.119 (1.097e+07/s), 9.137 (1.094e+07/s), 9.110 (1.098e+07/s), 9.128 (1.096e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 64 (5x20): 10.156 (9.846e+06/s), 10.136 (9.866e+06/s), 10.143 (9.859e+06/s), 10.146 (9.856e+06/s), 10.147 (9.855e+06/s), our sketch matmul N, D, M, d: 10000, 512, 10, 64 (5x20): 17.739 (5.637e+06/s), 17.767 (5.628e+06/s), 17.641 (5.669e+06/s), 17.647 (5.667e+06/s), 17.640 (5.669e+06/s), blas sketch matmul N, D, M, d: 10000, 512, 10, 128 (5x20): 17.149 (5.831e+06/s), 17.183 (5.820e+06/s), 17.144 (5.833e+06/s), 17.109 (5.845e+06/s), 17.182 (5.820e+06/s), our sketch matmul N, D, M, d: 10000, 512, 10, 128 (5x20): 35.289 (2.834e+06/s), 35.025 (2.855e+06/s), 35.294 (2.833e+06/s), 35.022 (2.855e+06/s), 35.071 (2.851e+06/s), blas matmul N, D, M: 10000, 512, 10 (5x20): 4.174 (2.396e+07/s), 4.136 (2.418e+07/s), 4.164 (2.402e+07/s), 4.198 (2.382e+07/s), 4.188 (2.388e+07/s), our matmul N, D, M: 10000, 512, 10 (5x20): 3.546 (2.820e+07/s), 3.546 (2.820e+07/s), 3.553 (2.815e+07/s), 3.555 (2.813e+07/s), 3.560 (2.809e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 2 (5x20): 4.085 (2.448e+08/s), 4.091 (2.444e+08/s), 4.055 (2.466e+08/s), 4.045 (2.472e+08/s), 4.057 (2.465e+08/s), our sketch matmul N, D, M, d: 10000, 512, 100, 2 (5x20): 1.322 (7.564e+08/s), 1.337 (7.479e+08/s), 1.336 (7.485e+08/s), 1.323 (7.559e+08/s), 1.322 (7.564e+08/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 4 (5x20): 3.631 (2.754e+08/s), 3.843 (2.602e+08/s), 3.798 (2.633e+08/s), 3.848 (2.599e+08/s), 3.847 (2.599e+08/s), our sketch matmul N, D, M, d: 10000, 512, 100, 4 (5x20): 2.626 (3.808e+08/s), 2.491 (4.014e+08/s), 2.510 (3.984e+08/s), 2.589 (3.862e+08/s), 2.480 (4.032e+08/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 8 (5x20): 4.275 (2.339e+08/s), 4.313 (2.319e+08/s), 4.333 (2.308e+08/s), 4.289 (2.332e+08/s), 4.130 (2.421e+08/s), our sketch matmul N, D, M, d: 10000, 512, 100, 8 (5x20): 3.405 (2.937e+08/s), 3.571 (2.800e+08/s), 3.405 (2.937e+08/s), 3.423 (2.921e+08/s), 3.405 (2.937e+08/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 16 (5x20): 5.392 (1.855e+08/s), 5.316 (1.881e+08/s), 5.283 (1.893e+08/s), 5.281 (1.894e+08/s), 5.184 (1.929e+08/s), our sketch matmul N, D, M, d: 10000, 512, 100, 16 (5x20): 6.046 (1.654e+08/s), 6.047 (1.654e+08/s), 6.076 (1.646e+08/s), 6.071 (1.647e+08/s), 6.044 (1.655e+08/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 32 (5x20): 7.291 (1.372e+08/s), 7.293 (1.371e+08/s), 7.308 (1.368e+08/s), 7.296 (1.371e+08/s), 7.294 (1.371e+08/s), our sketch matmul N, D, M, d: 10000, 512, 100, 32 (5x20): 10.697 (9.348e+07/s), 10.584 (9.448e+07/s), 10.599 (9.435e+07/s), 10.611 (9.424e+07/s), 10.594 (9.439e+07/s), blas sketch matmul N, D, M, d: 10000, 512, 100, 64 (5x20): 11.586 (8.631e+07/s), 11.528 (8.675e+07/s), 11.528 (8.675e+07/s), 11.535 (8.669e+07/s), 11.530 (8.673e+07/s), our sketch matmul N, D, M, d: 10000, 512, 100, 64 (5x20): 20.459 (4.888e+07/s), 20.514 (4.875e+07/s), 20.542 (4.868e+07/s), 20.429 (4.895e+07/s), 20.532 (4.870e+07/s), blas matmul N, D, M: 10000, 512, 100 (5x20): 13.506 (7.404e+07/s), 13.432 (7.445e+07/s), 13.467 (7.426e+07/s), 13.464 (7.427e+07/s), 13.484 (7.416e+07/s), our matmul N, D, M: 10000, 512, 100 (5x20): 27.160 (3.682e+07/s), 27.135 (3.685e+07/s), 27.260 (3.668e+07/s), 27.213 (3.675e+07/s), 27.268 (3.667e+07/s), blas sketch matmul N, D, M, d: 223590, 96, 12, 2 (5x20): 17.987 (1.492e+08/s), 17.601 (1.524e+08/s), 18.118 (1.481e+08/s), 17.847 (1.503e+08/s), 17.977 (1.493e+08/s), our sketch matmul N, D, M, d: 223590, 96, 12, 2 (5x20): 5.117 (5.243e+08/s), 5.115 (5.246e+08/s), 5.102 (5.259e+08/s), 5.088 (5.273e+08/s), 5.111 (5.250e+08/s), blas sketch matmul N, D, M, d: 223590, 96, 12, 4 (5x20): 11.524 (2.328e+08/s), 12.362 (2.170e+08/s), 11.828 (2.268e+08/s), 11.793 (2.275e+08/s), 11.785 (2.277e+08/s), our sketch matmul N, D, M, d: 223590, 96, 12, 4 (5x20): 9.979 (2.689e+08/s), 10.007 (2.681e+08/s), 10.010 (2.680e+08/s), 10.010 (2.680e+08/s), 9.973 (2.690e+08/s), blas sketch matmul N, D, M, d: 223590, 96, 12, 8 (5x20): 19.261 (1.393e+08/s), 19.116 (1.404e+08/s), 19.205 (1.397e+08/s), 19.342 (1.387e+08/s), 19.189 (1.398e+08/s), our sketch matmul N, D, M, d: 223590, 96, 12, 8 (5x20): 14.543 (1.845e+08/s), 14.510 (1.849e+08/s), 14.570 (1.842e+08/s), 14.556 (1.843e+08/s), 14.509 (1.849e+08/s), blas matmul N, D, M: 223590, 96, 12 (5x20): 19.189 (1.398e+08/s), 19.231 (1.395e+08/s), 19.378 (1.385e+08/s), 19.348 (1.387e+08/s), 19.390 (1.384e+08/s), our matmul N, D, M: 223590, 96, 12 (5x20): 16.242 (1.652e+08/s), 16.194 (1.657e+08/s), 16.197 (1.657e+08/s), 16.230 (1.653e+08/s), 16.238 (1.652e+08/s), blas sketch matmul N, D, M, d: 49284, 27, 2, 2 (5x20): 0.375 (2.628e+08/s), 0.373 (2.643e+08/s), 0.380 (2.594e+08/s), 0.380 (2.594e+08/s), 0.378 (2.608e+08/s), our sketch matmul N, D, M, d: 49284, 27, 2, 2 (5x20): 0.219 (4.501e+08/s), 0.220 (4.480e+08/s), 0.219 (4.501e+08/s), 0.216 (4.563e+08/s), 0.203 (4.856e+08/s), blas matmul N, D, M: 49284, 27, 2 (5x20): 0.327 (3.014e+08/s), 0.318 (3.100e+08/s), 0.319 (3.090e+08/s), 0.328 (3.005e+08/s), 0.328 (3.005e+08/s), our matmul N, D, M: 49284, 27, 2 (5x20): 0.186 (5.299e+08/s), 0.181 (5.446e+08/s), 0.183 (5.386e+08/s), 0.174 (5.665e+08/s), 0.173 (5.698e+08/s), """ def _load_matmul_times_for_n_d_m(startswith): lines = microbench_output.split('\n') matmul_lines = [line for line in lines if line.startswith(startswith)] matmul_shape_to_times = {} matmul_shape_to_thruputs = {} for line in matmul_lines: start_idx = line.find(':') + 1 end_idx = line.find('(') nmd_str = line[start_idx:end_idx] N, D, M = [int(substr) for substr in nmd_str.split(',')[:3]] speeds_str = line[line.find('):') + 2:] speed_pairs = speeds_str.split(',')[:5] # print("N, D, M: ", N, D, M) # print("speed pairs: ", speed_pairs) times = [] thruputs = [] for pair in speed_pairs: pair = pair.strip() if not len(pair): continue # handle trailing comma on line # print("pair: ", pair) pair = pair.strip() time_str, thruput_str = pair.split() times.append(float(time_str)) thruput_str = thruput_str.strip('()s/') thruputs.append(float(thruput_str)) key = (N, D, M) matmul_shape_to_times[key] = times matmul_shape_to_thruputs[key] = thruputs # print("what we're getting from func:") # pprint.pprint(matmul_shape_to_times) # pprint.pprint(matmul_shape_to_thruputs) return matmul_shape_to_times, matmul_shape_to_thruputs def _load_sketch_times_for_n_d_m(startswith): # print("loading sketch times for ", startswith) lines = microbench_output.split('\n') matmul_lines = [line for line in lines if line.startswith(startswith)] matmul_shape_to_times = {} matmul_shape_to_thruputs = {} for line in matmul_lines: start_idx = line.find(':') + 1 end_idx = line.find('(') nmd_str = line[start_idx:end_idx] N, D, M, d = [int(substr) for substr in nmd_str.split(',')[:4]] speeds_str = line[line.find('):') + 2:] speed_pairs = speeds_str.split(',')[:5] # print("N, D, M: ", N, D, M) # print("speed pairs: ", speed_pairs) times = [] thruputs = [] for pair in speed_pairs: pair = pair.strip() if not len(pair): continue # handle trailing comma on line # print("pair: ", pair) pair = pair.strip() time_str, thruput_str = pair.split() times.append(float(time_str)) thruput_str = thruput_str.strip('()s/') thruputs.append(float(thruput_str)) key = (N, D, M, d) matmul_shape_to_times[key] = times matmul_shape_to_thruputs[key] = thruputs # pprint.pprint(matmul_shape_to_times) # pprint.pprint(matmul_shape_to_thruputs) return matmul_shape_to_times, matmul_shape_to_thruputs def load_matmul_times_for_n_d_m(key1='blas matmul', key2='our matmul', sketches=False): if sketches: # print("results from blas:") shape2lat0, shape2thruput0 = _load_sketch_times_for_n_d_m(key1) # print("results from ours:") shape2lat1, shape2thruput1 = _load_sketch_times_for_n_d_m(key2) else: # print("results from blas:") shape2lat0, shape2thruput0 = _load_matmul_times_for_n_d_m(key1) # print("results from ours:") shape2lat1, shape2thruput1 = _load_matmul_times_for_n_d_m(key2) # take minimum of time from eigen blas and our sgemm shape2lat = {} for k in shape2lat0: vals0 = shape2lat0.get(k, [1e20]) vals1 = shape2lat1.get(k, [1e20]) mean0, mean1 = np.mean(vals0), np.mean(vals1) if mean0 < mean1: shape2lat[k] = shape2lat0[k] else: shape2lat[k] = shape2lat1[k] shape2thruput = {} for k in shape2thruput0: vals0 = shape2thruput0.get(k, [-1e20]) vals1 = shape2thruput1.get(k, [-1e20]) # print("k, vals0, vals1: ", k) # print(vals0) # print(vals1) mean0, mean1 = np.mean(vals0), np.mean(vals1) if mean0 > mean1: shape2thruput[k] = shape2thruput0[k] else: shape2thruput[k] = shape2thruput1[k] # print("what we're returning:") # pprint.pprint(shape2lat) # pprint.pprint(shape2thruput) return shape2lat, shape2thruput def _load_vq_times_for_n_d_m(startswith): lines = microbench_output.split('\n') lines = [line for line in lines if line.startswith(startswith)] shape_ncodebooks_to_times = {} shape_ncodebooks_to_thruputs = {} for line in lines: start_idx = line.find(':') + 1 end_idx = line.find('(') nmd_str = line[start_idx:end_idx] N, D, M, C = [int(substr) for substr in nmd_str.split(',')[:4]] speeds_str = line[line.find('):') + 2:] speed_pairs = speeds_str.split(',')[:5] times = [] thruputs = [] for pair in speed_pairs: pair = pair.strip() if not len(pair): continue # handle trailing comma on line time_str, thruput_str = pair.split() times.append(float(time_str)) thruput_str = thruput_str.strip('()s/') thruputs.append(float(thruput_str)) key = (N, D, M, C) shape_ncodebooks_to_times[key] = times shape_ncodebooks_to_thruputs[key] = thruputs # print("startswith: ", startswith) # if 'bolt' in startswith: # print("bolt speed dicts:") # pprint.pprint(shape_ncodebooks_to_times) # pprint.pprint(shape_ncodebooks_to_thruputs) return shape_ncodebooks_to_times, shape_ncodebooks_to_thruputs # def load_multisplit_times_for_n_d_m(): # return _load_vq_times_for_n_d_m('famm mithral') def load_bolt_times_for_n_d_m(): return _load_vq_times_for_n_d_m('amm bolt') def load_mithral_f32_times_for_n_d_m(): # two spaces so it doesn't try to read enc and zip times return _load_vq_times_for_n_d_m('f32 amm mithral ') def load_mithral_i16_times_for_n_d_m(): return _load_vq_times_for_n_d_m('i16 amm mithral ') def load_mithral_i8_times_for_n_d_m(): return _load_vq_times_for_n_d_m('i8 amm mithral ') def load_mithral_times_for_n_d_m(): return _load_vq_times_for_n_d_m('f32 amm mithral ') def load_sketch_times_for_n_d_m(): return load_matmul_times_for_n_d_m( 'blas sketch matmul', 'our sketch matmul', sketches=True) def main(): # load_matmul_times_for_n_d_m() # load_multisplit_times_for_n_d_m() # load_bolt_times_for_n_d_m() # pprint.pprint(load_sketch_times_for_n_d_m()) # pprint.pprint(load_multisplit_times_for_n_d_m()) # pprint.pprint(load_mithral_times_for_n_d_m()) ret = load_matmul_times_for_n_d_m() print("matmul latencies, thruputs") pprint.pprint(ret) # pprint.pprint(load_bolt_times_for_n_d_m()) if __name__ == '__main__': main()
#!/usr/bin/env python import itertools import numpy as np from sklearn import cluster from scipy import signal # import types import kmc2 # state-of-the-art kmeans initialization (as of NIPS 2016) from joblib import Memory _memory = Memory('.', verbose=0) # ================================================================ misc def is_dict(x): return isinstance(x, dict) def is_list_or_tuple(x): return isinstance(x, (list, tuple)) def as_list_or_tuple(x): return x if is_list_or_tuple(x) else [x] def is_scalar_seq(x): try: [float(element) for element in x] return True except TypeError: return False def as_scalar_seq(x): if is_scalar_seq(x): return x try: _ = float(x) return [x] except TypeError: raise TypeError("Couldn't convert value '{}' to sequence " "of scalars".format(x)) def is_string(x): return isinstance(x, (str,)) def flatten_list_of_lists(l): return list(itertools.chain.from_iterable(l)) def element_size_bytes(x): return np.dtype(x.dtype).itemsize def invert_permutation(permutation): return np.arange(len(permutation))[np.argsort(permutation)] # ================================================================ image def conv2d(img, filt, pad='same'): # assert pad in ('same',) # TODO support valid # mode = 'constant' if len(img.shape) == 2: return signal.correlate2d(img, filt, mode=pad) # img is more than 2d; do a 2d conv for each channel and sum results assert len(img.shape) == 3 out = np.zeros(img.shape[:2], dtype=np.float32) for c in range(img.shape[2]): f = filt[:, :, c] if len(filt.shape) == 3 else filt out += signal.correlate2d(img[:, :, c], f, mode=pad) return out # def filter_img(img, filt): # out = conv2d(img, filt) # return out / np.max(out) # ================================================================ distance def dists_sq(X, q): diffs = X - q return np.sum(diffs * diffs, axis=-1) def dists_l1(X, q): diffs = np.abs(X - q) return np.sum(diffs, axis=-1) def sq_dists_to_vectors(X, queries, rowNorms=None, queryNorms=None): Q = queries.shape[0] mat_size = X.shape[0] * Q mat_size_bytes = element_size_bytes(X[0] + queries[0]) if mat_size_bytes > int(1e9): print("WARNING: sq_dists_to_vectors: attempting to create a matrix" \ "of size {} ({}B)".format(mat_size, mat_size_bytes)) if rowNorms is None: rowNorms = np.sum(X * X, axis=1, keepdims=True) if queryNorms is None: queryNorms = np.sum(queries * queries, axis=1) dotProds = np.dot(X, queries.T) return (-2 * dotProds) + rowNorms + queryNorms # len(X) x len(queries) def all_eq(x, y): if len(x) != len(y): return False if len(x) == 0: return True return np.max(np.abs(x - y)) < .001 def top_k_idxs(elements, k, smaller_better=True, axis=-1): if smaller_better: # return indices of lowest elements which_nn = np.arange(k) return np.argpartition(elements, kth=which_nn, axis=axis)[:k] else: # return indices of highest elements which_nn = len(elements) - 1 - np.arange(k) return np.argpartition(elements, kth=which_nn, axis=axis)[-k:][::-1] def compute_true_knn(X, Q, k=1000, print_every=5, block_sz=128): nqueries = Q.shape[0] nblocks = int(np.ceil(nqueries / float(block_sz))) truth = np.full((nqueries, k), -999, dtype=np.int32) if nqueries <= block_sz: dists = sq_dists_to_vectors(Q, X) assert dists.shape == (Q.shape[0], X.shape[0]) for i in range(nqueries): truth[i, :] = top_k_idxs(dists[i, :], k) # truth[i, :] = top_k_idxs(dists[:, i], k) return truth for b in range(nblocks): # recurse to fill in knn for each block start = b * block_sz end = min(start + block_sz, nqueries) rows = Q[start:end, :] truth[start:end, :] = compute_true_knn(X, rows, k=k, block_sz=block_sz) if b % print_every == 0: print("computing top k for query block " "{} (queries {}-{})...".format(b, start, end)) # for i in range(nqueries): # if i % print_every == 0: # print "computing top k for query {}...".format(i) # truth[i, :] = top_k_idxs(dists[i, :], k) print("done") assert np.all(truth != -999) return truth def knn(X, q, k, dist_func=dists_sq): dists = dist_func(X, q) idxs = top_k_idxs(dists, k) return idxs, dists[idxs] @_memory.cache def kmeans(X, k, max_iter=16, init='kmc2', return_sse=False): X = X.astype(np.float32) # handle fewer nonzero rows than centroids (mostly just don't choke # if X all zeros, which happens when run in PQ with tiny subspaces) rowsums = X.sum(axis=1) nonzero_mask = rowsums != 0 nnz_rows = np.sum(nonzero_mask) if nnz_rows < k: print("X.shape: ", X.shape) print("k: ", k) print("nnz_rows: ", nnz_rows) centroids = np.zeros((k, X.shape[1]), dtype=X.dtype) labels = np.full(X.shape[0], nnz_rows, dtype=np.int) if nnz_rows > 0: # special case, because can't have slice of size 0 # make a centroid out of each nonzero row, and assign only those # rows to that centroid; all other rows get assigned to next # centroid after those, which is all zeros centroids[nnz_rows] = X[nonzero_mask] labels[nonzero_mask] = np.arange(nnz_rows) if return_sse: return centroids, labels, 0 return centroids, labels # if k is huge, initialize centers with cartesian product of centroids # in two subspaces sqrt_k = int(np.ceil(np.sqrt(k))) if k >= 16 and init == 'subspaces': print("kmeans: clustering in subspaces first; k, sqrt(k) =" " {}, {}".format(k, sqrt_k)) _, D = X.shape centroids0, _ = kmeans(X[:, :D/2], sqrt_k, max_iter=1) centroids1, _ = kmeans(X[:, D/2:], sqrt_k, max_iter=1) seeds = np.empty((sqrt_k * sqrt_k, D), dtype=np.float32) for i in range(sqrt_k): for j in range(sqrt_k): row = i * sqrt_k + j seeds[row, :D/2] = centroids0[i] seeds[row, D/2:] = centroids1[j] seeds = seeds[:k] # rounded up sqrt(k), so probably has extra rows elif init == 'kmc2': try: seeds = kmc2.kmc2(X, k).astype(np.float32) except ValueError: # can happen if dist of 0 to centroid print("WARNING: couldn't use kmc2 initialization") seeds = 'k-means++' if k < max_iter else 'random' else: raise ValueError("init parameter must be one of {'kmc2', 'subspaces'}") est = cluster.MiniBatchKMeans( k, init=seeds, max_iter=max_iter, n_init=1).fit(X) if return_sse: return est.cluster_centers_, est.labels_, est.inertia_ return est.cluster_centers_, est.labels_ def orthonormalize_rows(A): Q, R = np.linalg.qr(A.T) return Q.T def random_rotation(D): rows = np.random.randn(D, D) return orthonormalize_rows(rows) def hamming_dist(v1, v2): return np.count_nonzero(v1 != v2) def hamming_dists(X, q): return np.array([hamming_dist(row, q) for row in X]) if __name__ == '__main__': a = np.random.randn(10) sort_idxs = np.argsort(a)[::-1] print(a) print(top_k_idxs(a, 3, smaller_better=False)) print(sort_idxs[:3])
#!/usr/bin/env python import numpy as np import numba import zstandard as zstd # pip install zstandard # ================================================================ Funcs def nbits_cost(diffs, signed=True): """ >>> [nbits_cost(i) for i in [0, 1, 2, 3, 4, 5, 7, 8, 9]] [0, 2, 3, 3, 4, 4, 4, 5, 5] >>> [nbits_cost(i) for i in [-1, -2, -3, -4, -5, -7, -8, -9]] [1, 2, 3, 3, 4, 4, 4, 5] >>> nbits_cost([]) array([], dtype=int32) >>> nbits_cost([0, 2, 1, 0]) array([0, 3, 2, 0], dtype=int32) >>> nbits_cost([0, 2, 1, 3, 4, 0], signed=False) array([0, 2, 1, 2, 3, 0], dtype=int32) """ if diffs is None: return None diffs = np.asarray(diffs, dtype=np.int32) if diffs.size == 0: return np.array([], dtype=np.int32) if not signed: assert np.all(diffs >= 0) pos_idxs = diffs > 0 nbits = np.zeros(diffs.shape, dtype=np.int32) nbits[pos_idxs] = np.floor(np.log2(diffs[pos_idxs])) + 1 nbits[~pos_idxs] = 0 return nbits # shape = diffs.shape # diffs = diffs.ravel() # zero_idxs = (diffs == 0) # # nbits[zero_idxs] = 0 # nbits = np.zeros(len(diffs), dtype=np.int32) # diffs = diffs[~zero_idxs] # equiv_diffs = np.abs(diffs) + (diffs >= 0).astype(np.int32) # +1 if < 0 # # assert np.all(np.abs(diffs) > 0) # # assert np.all(equiv_diffs > 0) # nbits[~zero_idxs] = np.ceil(np.log2(equiv_diffs)) + 1 # nbits = np.asarray(nbits, dtype=np.int32) # next line can't handle scalar # assert np.all(nbits >= 0) shape = diffs.shape diffs = diffs.ravel() equiv_diffs = np.abs(diffs) + (diffs >= 0).astype(np.int32) # +1 if < 0 nbits = np.ceil(np.log2(equiv_diffs)) + 1 nbits = np.asarray(nbits, dtype=np.int32) # next line can't handle scalar nbits[diffs == 0] = 0 assert np.all(nbits >= 0) return nbits.reshape(shape) if nbits.size > 1 else nbits[0] # unpack if scalar @numba.njit(fastmath=True) def zigzag_encode(x): """ >>> [zigzag_encode(i) for i in [0,1,-1,2,-2,3,-3]] [0, 1, 2, 3, 4, 5, 6] >>> zigzag_encode([0,1,-1,2,-2,3,-3]) array([0, 1, 2, 3, 4, 5, 6], dtype=int32) """ x = np.asarray(x, dtype=np.int32) return (np.abs(x) << 1) - (x > 0).astype(np.int32) @numba.njit(fastmath=True) def zigzag_decode(x): return np.bitwise_xor(x >> 1, -np.bitwise_and(x, 1)) def quantize(X, nbits=16, minval=None, maxval=None): minval = np.min(X) if minval is None else minval maxval = np.max(X) if maxval is None else maxval unsigned_max = (1 << nbits) - 1 dtype_min = 1 << (nbits - 1) scale = float(unsigned_max) / maxval X = np.maximum(0, X - minval) X = np.minimum(unsigned_max, X * scale) X -= dtype_min # center at 0 dtype = {16: np.int16, 12: np.int16, 8: np.int8}[nbits] return X.astype(dtype) # ================================================================ def zstd_compress(buff, comp=None): comp = zstd.ZstdCompressor() if comp is None else comp if isinstance(buff, str): buff = bytes(buff, encoding='utf8') return comp.compress(buff) def zstd_decompress(buff, decomp=None): decomp = zstd.ZstdDecompressor() if decomp is None else decomp return decomp.decompress(decomp) # ============================================================== sprintz # except without the predictive coding part because we do that manually; # we also omit the run-length encoding because the author says that's a # huge pain to code and won't change the results much for our fast-changing # time series; also we don't do the grouping thing since it only # affects the decoding speed (it could affect the ratio slightly if the # number of variables were really low and not a multiple of 8, but neither # is the case for us) # def bitpack_vec(x, nbits_per_element): # n = len(x) # total_nbits = n * nbits_per_element # bitvec = np.zeros(total_nbits, dtype=np.bool) # for i, val in enumerate(x): # start_idx = i * nbits_per_element # for b in range(nbits_per_element): # bit = (val >> b) & 1 # bitvec[start_idx + b] = bit # return np.packbits(bitvec) # def bitunpack(X, nbits_per_element): # was_1d = X.ndim == 1 # X = np.atleast_2d(X) # N, D = X.shape # ret = np.unpackbits(X, axis=1) # if was_1d: # ret = ret.squeeze() # return ret # @numba.njit(fastmath=True) def bitpack(X, nbits_per_element): was_1d = X.ndim == 1 X = np.atleast_2d(X) N, D = X.shape # orig_elemsz = X.dtype.itemsize orig_elemsz_bits = 8 * X.dtype.itemsize assert X.dtype in (np.uint8, np.uint16) assert X.dtype in (np.uint8, np.uint16) if nbits_per_element == orig_elemsz_bits: ret = X elif X.dtype == np.uint8: # print("N, D, nbits: ", N, D, nbits_per_element) # shape = X.shape X = X.ravel() # unpacked = np.unpackbits(X, count=nbits_per_element, bitorder='little', axis=-1) unpacked = np.unpackbits(X, bitorder='little', axis=-1) # print("unpacked initial shape: ", unpacked.shape) unpacked = unpacked.reshape(N * D, 8)[:, :nbits_per_element] # print("unpacked new shape: ", unpacked.shape) ret = np.packbits(unpacked.reshape(N, -1), axis=1) # ret = ret.reshape(N, -1) # print("ret.shape: ", ret.shape) else: # X_low = (X & 0xff)[:, :, np.newaxis] # X_high = ((X & 0xff00) >> 8)[:, :, np.newaxis] # X_combined = np.concatenate([X_low, X_high], axis=-1) # X = X[:, :, np.newaxis] # X = np.concatenate([X, X], axis=-1) # X[:, :, 0] = X[:, :, 0] & 0xff # X[:, :, 1] = (X[:, :, 1] & 0xff00) >> 8 # X = X.reshape(N, 2 * D).astype(np.uint8) X = np.ascontiguousarray(X).view(np.uint8).reshape(N, 2 * D) # print("X shape: ", X.shape) unpacked = np.unpackbits(X, axis=1, bitorder='little') unpacked = unpacked.reshape(N, orig_elemsz_bits, D) # unpacked = unpacked[:, ::-1, :] # low bits in low idxs unpacked = np.ascontiguousarray(unpacked[:, :nbits_per_element]) ret = np.packbits(unpacked.reshape(N, -1)) # nbits_per_row = D * nbits_per_element # bitmat = np.zeros((N, nbits_per_row), dtype=np.uint8) # for j in range(D): # col = X[:, j] # start_idx = j * nbits_per_element # for b in range(nbits_per_element): # bit = (col >> b) & 1 # bitmat[:, start_idx + b] = bit # ret = np.packbits(bitmat, axis=1) if was_1d: ret = ret.squeeze() return ret @numba.njit(fastmath=True) def _sprintz_header_sz(headers, header_elem_nbits): _, D = headers.shape header_row_sz = int(np.ceil(D * header_elem_nbits / 8)) rows_total_nbits = headers.sum(axis=1) # zero_rows = rows_total_nbits == 0 # header_sz = np.sum(nzero_rows) # one byte for run length # pair_sums = zero_rows + header_sz = 0 prev_was_zero = False for row in rows_total_nbits: is_zero = row == 0 if is_zero: if prev_was_zero: continue else: header_sz += 1 # start of run else: header_sz += header_row_sz prev_was_zero = is_zero return header_sz # def sprintz_packed_size(X, nbits=None, just_return_sz=False, postproc='zstd'): def sprintz_packed_size(X, nbits=None, just_return_sz=True, postproc=None): if nbits is None: nbits = {1: 8, 2: 16}.get(X.dtype.itemsize, 16) unsigned_dtype = {8: np.uint8, 16: np.uint16}[nbits] window_len = 8 pad_nrows = X.shape[0] % window_len if pad_nrows != 0: pad_rows = np.zeros((pad_nrows, X.shape[1]), dtype=X.dtype) X = np.vstack([X, pad_rows]) N, D = X.shape if X.dtype.itemsize > 2: # basically just catching floats # print("sprintz: quantizing X...WTF") X = quantize(X, nbits=nbits) if np.min(X) < 0: # print("sprintz: zigzag_encoding X!") X = zigzag_encode(X).astype(unsigned_dtype) # else: # print("sprintz: not zigzag_encoding X!") header_elem_nbits = {8: 3, 16: 4}[nbits] X_nbits = nbits_cost(X, signed=False) X_nbits = np.asfarray(X_nbits).reshape(N // window_len, window_len, -1) block_nbits = X_nbits.max(axis=1).astype(np.uint8) block_nbits[block_nbits == (nbits - 1)] = nbits headers = block_nbits if just_return_sz: payload_sz = int(block_nbits.sum() * window_len / 8) header_sz = _sprintz_header_sz(headers, header_elem_nbits) # print("header sz: ", header_sz) return header_sz + payload_sz nwindows = N // window_len payloads = [] for i in range(nwindows): start_idx = i * window_len end_idx = start_idx + window_len X_slice = X[start_idx:end_idx] for j in range(D): col = X_slice[:, j] payloads.append(bitpack(col, headers[i, j])) headers = bitpack(headers, header_elem_nbits) payloads = np.hstack(payloads) if postproc is None: return headers.nbytes + payloads.nbytes elif postproc == 'zstd': return len(zstd_compress(headers)) + len(zstd_compress(payloads)) # # nbits_slice = nbits_cost(X_slice, signed=False) # nbits_slice = X_nbits[start_idx:end_idx] # max_nbits = nbits_slice.max(axis=0) # headers[i] = np.minimum(max_nbits, nbits - 1) # 8->7, 16->15 # max_nbits[max_nbits == nbits - 1] = nbits # 7->8, 15->16 # for j in range(D): # col = X_slice[:, j] # payloads.append(bitpack(col, max_nbits[j])) # headers = bitpack(headers, header_elem_nbits) # payloads = np.hstack(payloads) # header_bytes = headers.tobytes() # # payload_bytes = headers.tobytes() # blosc.compress(buff, typesize=elem_sz, # cname=compressor, shuffle=shuffle) # if __name__ == '__main__': import doctest doctest.testmod()
#!/usr/bin/env python from __future__ import print_function import os import numpy as np import pandas as pd from io import StringIO from . import amm_methods as methods from joblib import Memory _memory = Memory('.', verbose=1) pd.options.mode.chained_assignment = None # suppress stupid warning RESULTS_DIR = os.path.join('results', 'amm') TIMING_RESULTS_DIR = os.path.join(RESULTS_DIR, 'timing') # we log these, but don't need them for the plots AMM_DROP_COLS = ['__pyience_timestamp__', 'y_mean', 'y_std', 'bias', 'raw_mse', 'r', 'alpha', 'ncentroids'] def _read_csv_with_garbage(path, **kwargs): with open(path, 'r') as f: # print("\n".join(f.readlines())) keep_lines = [line.strip() for line in f.readlines() if (',' in line and not line.startswith('-'))] contents = '\n'.join(keep_lines) # print("contents\n", contents) return pd.read_csv(StringIO(contents), **kwargs) def rename_values_in_col(df, col, name_map, drop_others=True): name_map = {k.strip().lower(): v for k, v in name_map.items()} vals = [name_map.get(name.strip().lower(), "") for name in df[col]] valid_vals = set(name_map.values()) # print("valid_vals: ", valid_vals) valid_mask = np.array([val in valid_vals for val in vals]) # print("valid mask: ", valid_mask) df = df.copy() df[col] = vals if drop_others: df = df.loc[valid_mask] return df # print(df) def melt_observation_cols(df, cols, var_name=None, value_name=None): """like pd.melt, but assumes only 1 observation var instead of 1 id var""" independent_vars = [col for col in df.columns if col not in set(cols)] return pd.melt(df, id_vars=independent_vars, value_vars=cols, var_name=var_name, value_name='time') def melt_times(df, ntimes=5): observation_vars = 't0 t1 t2 t3 t4'.split() observation_vars = observation_vars[:ntimes] return melt_observation_cols( df, observation_vars, var_name='timing_trial', value_name='time') def drop_cols_inplace(df, cols): for col in AMM_DROP_COLS: try: df.drop([col], axis=1, inplace=True) except KeyError: pass return df def frac_above_thresh(df, xvar, yvar, methodvar, unitvar, ythresh): """ (method, xvar) -> [0, 1] Assumes you have a tidy dataframe where method, xvar, yvar, and unit are each a col. """ df = df.copy() # df['frac_above_thresh'] = (df[yvar] > ythresh).astype(np.float) # (method, xvar, [(unit, yvar)]) -> bool df['frac_above_thresh'] = (df[yvar] > ythresh).astype(np.float) # independent_vars = [methodvar, unitvar, xvar] independent_vars = [methodvar, xvar] # return df.groupby(independent_vars)['is_above_thresh'].transform('mean') # last part converts from groupby back to regular df EDIT: no it doesn't # return df.groupby(independent_vars)['frac_above_thresh'].mean().apply(pd.Series) # return df.groupby(independent_vars)['is_above_thresh'].mean() df = df.groupby(independent_vars)['frac_above_thresh'].mean() # this is the magic line; turn multi-index levels into regular cols, with # multi-index value broadcast all the corresponding rows; # WOW this took a long time to figure out... df = df.reset_index(level=independent_vars) return df # tmp = df.groupby(independent_vars)['frac_above_thresh'].mean() # tmp = df.groupby(independent_vars)['frac_above_thresh'].transform('mean') # print("tmp:\n", tmp) # df['frac_above_thresh'] = tmp # return df # return df.groupby(independent_vars)[independent_vars + ['frac_above_thresh']] # ret = df.groupby([methodvar, unitvar, xvar])['__above_thresh__'] # ret.drop(['__above_thresh__'], axis=1, inplace=True) # return ret def encode_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'encode-timing.csv') ORIG_HEADERS = 'algo __ N D C B ___ t0 _0 t1 _1 t2 _2 t3 _3 t4 _4'.split() USE_HEADERS = 'algo N D C B t0 t1 t2 t3 t4'.split() # ORIG_HEADERS = 'algo __ N D C ___ t0 _0 t1 _1 t2 _2'.split() # USE_HEADERS = 'algo N D C t0 t1 t2'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] return df # print(df) def lut_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'lut-timing.csv') ORIG_HEADERS = ('algo __ N D C B lutconst ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'algo N D C B lutconst t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] return df def scan_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'scan-timing.csv') ORIG_HEADERS = 'algo __ N C B M ___ t0 _0 t1 _1 t2 _2 t3 _3 t4 _4'.split() USE_HEADERS = 'algo N C B M t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None, skiprows=1) df = df[USE_HEADERS] return df def mithral_amm_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'amm-mithral-timing.csv') ORIG_HEADERS = ('dset dtype algo __ N D M C lutconst ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'dset dtype algo N D M C lutconst t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] return df def bolt_amm_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'amm-bolt-timing.csv') ORIG_HEADERS = ('dset dtype algo __ N D M C ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'dset dtype algo N D M C t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] df['fixedB'] = df['algo'].str.strip().str.endswith('noenc') df.drop('algo', axis=1, inplace=True) df = df.loc[df['fixedB']] # print("bolt df:\n", df) # import sys; sys.exit() return df def dense_amm_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'amm-dense-timing.csv') ORIG_HEADERS = ('dset algo __ N D M d ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'dset algo N D M d t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] df['algo'] = df['algo'].str.strip() # drop stuff that doesn't have fixedW; we just let the existing methods # use fixedW (same as fixedB in amm.py), instead of having to encode the # smaller matrix # df = df.loc[~df['algo'].isin(['blas sketch matmul', 'our sketch matmul'])] t_sums = (df['t0'] + df['t1'] + df['t2'] + df['t3'] + df['t4']).values / 5 # df['t_avg'] = (df['t0'] + df['t1'] + df['t2'] + df['t3'] + df['t4']) / 5. # # mark whether it's from our gemm or eigen gemm # df['is_ours'] = df['algo'].str.startswith('our') # print("uniq n vals: ", np.unique(df['N'])) sizes = np.empty((len(df), 4), dtype=np.int) sizes[:, 0] = df['N'] sizes[:, 1] = df['D'] sizes[:, 2] = df['M'] sizes[:, 3] = df['d'] as_tuples = [tuple(row) for row in sizes] uniq_tuples = sorted(list(set(as_tuples))) keep_idxs = [] # print("sizes:\n", sizes) # print("uniq_tuples:\n", uniq_tuples) for tup in uniq_tuples: row = np.array(tup) idxs = np.where((sizes == row).sum(axis=1) == sizes.shape[1])[0] best_idx = idxs[np.argmin(t_sums[idxs])] # print(f"{tup} -> {best_idx}") keep_idxs.append(best_idx) df = df.iloc[keep_idxs] rename_dict = {} rename_dict['blas matmul'] = 'Brute Force' rename_dict['our matmul'] = 'Brute Force' rename_dict['blas sketch matmul'] = 'Dense Sketch' rename_dict['our sketch matmul'] = 'Dense Sketch' rename_dict['blas sketch fixedw matmul'] = 'Dense Sketch' rename_dict['our sketch fixedw matmul'] = 'Dense Sketch' df = rename_values_in_col(df, 'algo', rename_dict, drop_others=False) return df def osnap_amm_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'amm-osnap-timing.csv') ORIG_HEADERS = ('dset algo __ N D M d s ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'dset algo N D M d s t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] df.drop('algo', axis=1, inplace=True) return df def sparse_amm_timings(): TIMINGS_PATH = os.path.join(TIMING_RESULTS_DIR, 'amm-sparse-timing.csv') ORIG_HEADERS = ('dset algo __ N D M d frac ___ ' 't0 _0 t1 _1 t2 _2 t3 _3 t4 _4').split() USE_HEADERS = 'dset algo N D M d frac t0 t1 t2 t3 t4'.split() df = _read_csv_with_garbage(TIMINGS_PATH, names=ORIG_HEADERS, header=None) df = df[USE_HEADERS] df.drop('algo', axis=1, inplace=True) return df def scalar_quantized_amm_timings(): timings = [] timings.append(['Cifar10', 10000, 512, 10, 2.013]) timings.append(['Cifar100', 10000, 512, 100, 6.472]) timings.append(['Ucr128', 1000, 320, 128, .4808]) timings.append(['Caltech3x3', 49284, 27, 2, .894]) timings.append(['Caltech5x5', 48400, 75, 2, 1.409]) dicts = [{'dset': l[0], 'N': l[1], 'D': l[2], 'M': l[3], 'time': l[4]} for l in timings] # df = pd.DataFrame.from_records(dicts) # print("scalar_quantized_amm_timings: ") # print(df) return pd.DataFrame.from_records(dicts) # import sys; sys.exit() # df = pd.DataFrame.from_records(timings) # output from ./GEMMsBenchmark after defining FBGEMM_MEASURE_TIME_BREAKDOWN # in include/fbgemm/Fbgemm.h; recorded here since not in a results file ''' M, N, K, Type, Packing (us), Kernel (us), Postproc (us), Total (us), GOPs 10000, 10, 512, FBGEMM_i8_acc32, 438.069, 1465.04, 43.5959, 2013.31, 50.6 10000, 10, 512, FBGEMM_i8_acc16, 512.8, 1338.1, 69.9, 2115.1, 48.3 10000, 100, 512, FBGEMM_i8_acc32, 473.7, 9203.9, 85.9, 9923.9, 103.1 10000, 100, 512, FBGEMM_i8_acc16, 569.8, 5558.7, 108.5, 6472.2, 158.1 1000, 128, 320, FBGEMM_i8_acc32, 39.5, 724.6, 5.8, 795.2, 101.8 1000, 128, 320, FBGEMM_i8_acc16, 43.5, 404.1, 3.1, 480.8, 168.4 49284, 2, 27, FBGEMM_i8_acc32, 298.5, 226.2, 139.6, 894.0, 5.9 49284, 2, 27, FBGEMM_i8_acc16, 333.6, 650.1, 162.5, 1608.7, 3.3 48400, 2, 75, FBGEMM_i8_acc32, 482.0, 546.0, 141.5, 1409.3, 10.2 48400, 2, 75, FBGEMM_i8_acc16, 438.3, 1228.7, 159.2, 2278.4, 6.4 ''' # def _extract_cols_into_list_of_tuples(df, cols): def _extract_cols_into_list_of_tuples(df, cols): # return [tuple(row) for row in df[cols].iterrows()] ar = np.vstack([df[col] for col in cols]).T # print("ar: \n", ar) ar = np.atleast_2d(ar).astype(np.int) # return [tuple(row) for row in ar] return [sum([hash(-12435 * i + 1) ^ hash(1234567 * val) for i, val in enumerate(row)]) for row in ar] # return [int(hash(tuple(row))) for row in ar] def _join_on_cols(df_left, left_cols, df_right, right_cols, verbose=0): df_left = df_left.copy() df_right = df_right.copy() df_left['__index__'] = _extract_cols_into_list_of_tuples( df_left, left_cols) df_right['__index__'] = _extract_cols_into_list_of_tuples( df_right, right_cols) # dup_cols = set(left_cols) & set(right_cols) # if verbose > 0: # print("_join_on_cols(); dropping duplicate cols from rhs: ", dup_cols) # df_right = df_right.drop(dup_cols, axis=1) df = df_left.merge( df_right, on='__index__', how='left', suffixes=('', '_rhs')) df.drop(['__index__'], axis=1, inplace=True) # df.sort_values(left_cols, axis=0, inplace=True) return df def _join_with_mithral_times(df, timing_dtype='f32'): time_df = mithral_amm_timings() if timing_dtype is not None: time_df = time_df.loc[time_df['dtype'].str.strip() == timing_dtype] df = df.loc[df['method'].str.lower().str.startswith('mithral')] df['ncodebooks'] = df['ncodebooks'].astype(np.int) # time_df.reset_index(inplace=True, drop=True) # df.reset_index(inplace=True, drop=True) # print("time_df with appropriate dtype:\n", time_df) # import sys; sys.exit() # we also report times for subroutines within mithral; can't let it # use any of these; just use rename_values_in_col to drop them and also # get more intuitive debug output # rename_dict = {'amm mithral sparselut': 'Mithral, L = ??', # 'amm mithral nolut': 'Mithral, L = ∞'} name_mithral_dense = 'mithralDense' # only one we use; others arbitrary rename_dict = {'amm mithral sparselut': 'mithralSparse', 'amm mithral denselut': name_mithral_dense, 'amm mithral nolut': 'mithralOffline'} time_df = rename_values_in_col(time_df, 'algo', rename_dict) # give MithralPQ a valid lut const so the join will work (pq is equivalent # to constant of 1) is_mithral_pq = df['method'].str.lower().str.startswith('mithralpq') df.loc[is_mithral_pq, 'lut_work_const'] = 1 df_mpq = df.loc[is_mithral_pq].copy() # there shouldn't be rows that violated this, but there are (probably # from early runs that haven't been overwritten yet) df = df.loc[df['lut_work_const'].values <= df['ncodebooks'].values] # now add in extra rows for mithral with no lut computation (which is # assumed to use dense luts because no reason not to) vs mithral # with dense lut computation as part of the timing is_any_mithral = df['method'].str.lower().str.startswith('mithral') is_mithral = is_any_mithral & (~is_mithral_pq) is_dense = df['lut_work_const'] == -1 df_mithral_dense = df.loc[is_mithral & is_dense].copy() dummy_lutconst = -2 df_mithral_dense['lut_work_const'] = dummy_lutconst time_df['lutconst'].loc[ time_df['algo'] == name_mithral_dense] = dummy_lutconst # add in version of mithralpq with offline lut computation df_mpq = df.loc[df['method'].str.lower().str.startswith('mithralpq')] df_mpq['lut_work_const'] = -1 df = pd.concat([df, df_mithral_dense, df_mpq], axis=0) cols_df = 'N D M ncodebooks lut_work_const'.split() cols_time_df = 'N D M C lutconst'.split() # print("df cols: ", df.columns) # time_df.reset_index(inplace=True, drop=True) # df.reset_index(inplace=True, drop=True) df.sort_values(['method'] + cols_df, axis=0, inplace=True) time_df.sort_values(cols_time_df, axis=0, inplace=True) ret = _join_on_cols(df, cols_df, time_df, cols_time_df) # ret['lut_work_const'].loc[ret['lut_work_const'] == dummy_lutconst] = -1 # show_keys = 'method N D M C ncodebooks lutconst lut_work_const'.split() # print("mithral df:\n", df['method N D M ncodebooks lut_work_const'.split()]) # # print("mithral time df:\n", time_df.loc[time_df['dset'] == 'Cifar10']) # print("mithral time df:\n", time_df.loc[time_df['dset'] == 'Caltech3x3']) # print("joined df:\n", ret[show_keys]) # # print("joined df:\n", ret) # import sys; sys.exit() # one of these fails if the join failed; check if you have redundant # rows in either df or missing rows in the time df assert np.all(ret['C'] == ret['ncodebooks']) assert np.all(ret['lutconst'] == ret['lut_work_const']) return ret def _join_with_bolt_times(df): time_df = bolt_amm_timings() df = df.loc[df['method'].str.lower().str.startswith('bolt')] return _join_on_cols(df, 'N D M ncodebooks'.split(), time_df, 'N D M C'.split()) def _join_with_osnap_times(df): time_df = osnap_amm_timings() # df = df.loc[df['method'].str.lower().str.startswith('osnap')] df = df.loc[df['method'].isin( [methods.METHOD_OSNAP, methods.METHOD_HASHJL])] df['s'] = 1 df['s'].loc[df['method'] == methods.METHOD_OSNAP] = 4 # print("osnap df shape: ", df.shape) df['d'] = df['d'].astype(np.int) # print("time_df:\n", time_df[time_df['dset'] == 'Cifar10']) # note that d < s isn't present in time_df, which makes sense return _join_on_cols(df, 'N D M d s'.split(), time_df, 'N D M d s'.split()) def _join_with_brute_force_times(df): time_df = dense_amm_timings() df = df.loc[df['method'].str.lower().str.startswith('exact')] time_df = time_df.loc[time_df['algo'].str.lower().str.startswith('brute')] # print("df:\n", df) # print("time_df:\n", time_df) return _join_on_cols(df, 'N D M'.split(), time_df, 'N D M'.split()) def _join_with_dense_sketch_times(df): time_df = dense_amm_timings() # print("found methods in df: ", df['method'].unique()) # print("dense sketch methods: ", methods.DENSE_SKETCH_METHODS) df = df.loc[df['method'].isin(methods.DENSE_SKETCH_METHODS)] time_df = time_df.loc[time_df['algo'].str.lower().str.startswith( 'dense sketch')] # print("df:\n", df) # print("time_df:\n", time_df) return _join_on_cols(df, 'N D M d'.split(), time_df, 'N D M d'.split()) def _join_with_scalar_quantize_times(df): time_df = scalar_quantized_amm_timings() df = df.loc[df['method'] == methods.METHOD_SCALAR_QUANTIZE] # print("scalar quantize time df:\n", time_df) # print("scalar quantize acc df:\n", df.columns) # print(df['N D M'.split()]) # df_joint = _join_on_cols(df, 'N D M'.split(), time_df, 'N D M'.split()) # print("joined df: ") # print(df_joint['N D M time'.split()]) # import sys; sys.exit() return _join_on_cols(df, 'N D M'.split(), time_df, 'N D M'.split()) def extract_pareto_frontier_idxs(xvals, yvals): """assumes lower x is better and higher y is better""" assert len(xvals) == len(yvals) sort_idxs = np.argsort(xvals) xvals = xvals[sort_idxs] yvals = yvals[sort_idxs] # orig_idxs = np.arange(len(xvals)) first_idx = sort_idxs[0] curr_thresh = yvals[first_idx] keep_idxs = [first_idx] for i, y in enumerate(yvals[1:]): if y > curr_thresh: curr_thresh = y keep_idxs.append(sort_idxs[i + 1]) return keep_idxs def _join_with_sparse_sketch_times(df, sparse_pareto=True): time_df = sparse_amm_timings() df = df.loc[df['method'].str.lower().str.startswith('sparse')] df['d'] = df['d'].astype(np.int) new_rows = [] for _, row in df.iterrows(): # pprint.pprint(dict(row)) subdf = time_df for key in 'N D M d'.split(): subdf = subdf.loc[subdf[key] == row[key]] if len(subdf) < 1: continue sparsities = subdf['frac'] # print("subdf for N, D, M, D: ", [row[k] for k in 'N D M d'.split()]) # print(subdf) # N, D, M, d = [row[k] for k in 'N D M d'.split()] target_frac = row['sparsity'] small_enough_sparsities_idxs = np.where(sparsities.values <= target_frac)[0] if len(small_enough_sparsities_idxs): take_idx = small_enough_sparsities_idxs[-1] else: # no nonzeros, or at least uselessly few of them take_idx = np.argmin(sparsities.values) time_keys = 't0 t1 t2 t3 t4'.split() times_row = subdf.iloc[take_idx] # times = subdf.loc[take_idx, time_keys] row = dict(row) for key in time_keys: row[key] = float(times_row[key]) row['time'] = sum([float(times_row[key]) for key in time_keys]) / len(time_keys) new_rows.append(row) # return pd.DataFrame.from_records(new_rows) df = pd.DataFrame.from_records(new_rows) if not sparse_pareto: return df # # for dset in df[''] # subdf = df.loc[df['method'] == 'SparsePCA'] # here we have a bunch of hack stuff # print("df columns: ", df.columns) # yvals = 1. - df['normalized_mse'].values subdfs = [] for tid in df['task_id'].unique(): subdf = df.loc[df['task_id'] == tid] xvals = subdf['time'].values if 'acc_amm' in df.columns: yvals = subdf['acc_amm'].values else: yvals = 1. - subdf['normalized_mse'].values idxs = extract_pareto_frontier_idxs(xvals, yvals) subdfs.append(subdf.iloc[idxs]) df = pd.concat(subdfs, axis=0) return df def _clean_method_names_amm(df): key = 'method' if 'method' in df else 'algo' if 'lutconst' in df: df.loc[df['lutconst'] == -2, key] = 'MADDNESS Dense' is_lutconst_neg1 = df['lutconst'] == -1 is_mithral_pq = df['method'] == 'MithralPQ' df.loc[is_lutconst_neg1 & is_mithral_pq, key] = 'MADDNESS-PQ' df.loc[is_lutconst_neg1 & ~is_mithral_pq, key] = 'MADDNESS' df.loc[df['lutconst'] == 1, key] = 'MADDNESS, L = 1' df.loc[df['lutconst'] == 2, key] = 'MADDNESS, L = 2' df.loc[df['lutconst'] == 4, key] = 'MADDNESS, L = 4' # df.loc[df['lutconst'] == -2, key] = 'Mithral Dense' # is_lutconst_neg1 = df['lutconst'] == -1 # is_mithral_pq = df['method'] == 'MithralPQ' # df.loc[is_lutconst_neg1 & is_mithral_pq, key] = 'MithralPQ' # df.loc[is_lutconst_neg1 & ~is_mithral_pq, key] = 'Mithral' # df.loc[df['lutconst'] == 1, key] = 'Mithral, L = 1' # df.loc[df['lutconst'] == 2, key] = 'Mithral, L = 2' # df.loc[df['lutconst'] == 4, key] = 'Mithral, L = 4' # mask = df['lutconst'] == 1 # is_mithral_pq = df[key].str.lower().str.startswith('mithralpq') # mask &= ~is_mithral_pq # df[key][mask] = 'Mithral, L = ∞' # df[key].loc[df[key] == 'Exact'] = 'Brute Force' df[key].loc[df[key] == 'Exact'] = 'Exact' return df def _clean_metrics_amm(df): df = df.rename({'acc_amm': 'Accuracy'}, axis=1) # if 'time' not in df.columns: mask = df['time'].isna() # df.loc['time', mask] = ((df['t0'] + df['t1'] + df['t2'] + df['t3'] + df['t4']).values / 5.)[mask] times = (df['t0'] + df['t1'] + df['t2'] + df['t3'] + df['t4']) / 5. df.loc[mask, 'time'] = times.values[mask] df['Throughput'] = 1e3 * df['N'] * df['M'] / df['time'] # create ops column that sums number of multiplies + lookups df['muls'] = df['muls'].fillna(0) mask = ~df['nlookups'].isna() df['ops'] = df['muls'] # print("debugging df[ops]: ") # # print(type(df['ops'])) # # print(type(df['ops'])) # print(type(df['nlookups'])) df['ops'].loc[mask] += df['nlookups'].loc[mask] # df['nor'] # df_exact = df.loc[df['method'] == 'Brute Force'] df_exact = df.loc[df['method'] == 'Exact'] # print("df_exact\n", df_exact) if 'task_id' in df.columns: nuniq_tasks = len(df['task_id'].unique()) else: nuniq_tasks = 1 # cifar{10,100} assert df_exact.shape[0] == nuniq_tasks base_time = float(df_exact.loc[0, 'time']) df['NormalizedTime'] = df['time'] / base_time df['Speedup'] = 1. / df['NormalizedTime'] df['1 - NMSE'] = 1. - df['normalized_mse'] if 'Accuracy' in df.columns: df['Relative Accuracy'] = df['Accuracy'] / (df['acc_orig'] + 1e-20) # print("df.columns", df.columns) # df['Change in Accuracy'] = df['Accuracy'] - df['acc-1nn-raw'] # if 'acc-1nn-raw' in df.columns: # # # note that relative accuracy can actually be higher if errors # # # happen to compensate for incorrect classification sometimes # # print("max relative acc: ", df['Relative Accuracy'].values.max()) # # # assert df['Relative Accuracy'].values.max() <= 1.000001 # # acc_orig field is supposed to capture this, but I messed it up for # # 1nn so this will also work # tid2acc = {} # exactdf = df.loc[df['method'] == 'Exact'] # for tid in df['task_id'].unique(): # subdf = exactdf.loc[exactdf['task_id'] == tid] # tid2acc[tid] = subdf['Accuracy'].values[0] # df['BaseAccuracy'] = [tid2acc[tid] for tid in df['task_id']] # df['Relative Accuracy'] = df['Accuracy'] / df['BaseAccuracy'] return df def _join_with_times(df, timing_dtype='f32', sparse_pareto=True): df_quant = _join_with_scalar_quantize_times(df) # print("scalar quantize time df:\n", time_df) # print("scalar quantize acc df:\n", df) # print("df scalar quant:\n", df_quant['dset N D M time'.split()]) # import sys; sys.exit() df_bolt = _join_with_bolt_times(df) # # print("df bolt:\n", df_bolt) # df_tmp = df_bolt['N D M C ncodebooks method normalized_mse t0 t1 t2 t3 t4 task_id'.split()] # df_tmp = df_tmp.loc[df_tmp['task_id'].isin(['ucr Yoga k=128', 'ucr Wafer k=128'])] # df_tmp['time'] = (df_tmp['t0'] + df_tmp['t1'] + df_tmp['t2'] + df_tmp['t3'] + df_tmp['t4']) / 5. # print("df tmp:\n", df_tmp) # print("df tmp times:\n", df_tmp[['C', 'time', 'task_id']]) # # tids = df_tmp['task_id'].unique() # # # yep, exactly 6 results per dset # # counts = df_tmp.groupby('task_id')['task_id'].count() # # print("task counts: ", counts) # import sys; sys.exit() # print("df bolt:\n", df_bolt) # looks good # import sys; sys.exit() # assert np.all(df_mithral['lutconst'] == df_mithral['lut_work_const']) # df_mithral = df.loc[df['method'].str.startswith('Mithral')] # df_mithral.to_csv('mithral-caltech-debug.csv') df_mithral = _join_with_mithral_times(df, timing_dtype=timing_dtype) # df_tmp = df_mithral # df_tmp = df_tmp['N D M C ncodebooks lutconst lut_work_const method algo normalized_mse t0 t1'.split()] # # print("mithral rows:\n", df.loc[df['method'].str.startswith('mithral')]) # print("mithralpq rows after join:\n", df_tmp.loc[df_tmp['method'] == 'MithralPQ']) # print("mithral rows after join:\n", df_tmp[:100]) # mismatch_mask = df_tmp['lutconst'] != df_tmp['lut_work_const'] # print("mithral mismatched rows:\n", df_tmp.loc[mismatch_mask]) # print(df_mithral['lutconst', 'lut_work_const']) # import sys; sys.exit() # if this line fails, it's usually because the join with mithral times # failed assert np.all(df_mithral['lutconst'] == df_mithral['lut_work_const']) df_osnap = _join_with_osnap_times(df) df_brute = _join_with_brute_force_times(df) df_sketch = _join_with_dense_sketch_times(df) df_sparse = _join_with_sparse_sketch_times(df, sparse_pareto=sparse_pareto) # dfs = [df_mithral, df_bolt, df_osnap, df_brute, df_sketch, df_sparse] dfs = [df_quant, df_mithral, df_bolt, df_osnap, df_brute, df_sketch, df_sparse] return pd.concat(dfs, axis=0, join='outer', sort=False) def _clean_amm_results_df(df, timing_dtype='f32', sparse_pareto=True): # print("initial methods: ", df['method'].unique()) df = _join_with_times( df, timing_dtype=timing_dtype, sparse_pareto=sparse_pareto) # df['time'] = df['t_avg'] # df = melt_times(df) # print("uniq methods after joining with times:\n", sorted(df['method'].unique())) # import sys; sys.exit() df = _clean_metrics_amm(df) df = df.loc[~df['time'].isna()] # print("uniq methods after rming nan times:\n", sorted(df['method'].unique())) # import sys; sys.exit() df = _clean_method_names_amm(df) return df @_memory.cache def _read_amm_csv(fname, **kwargs): df = pd.read_csv(os.path.join(RESULTS_DIR, fname), **kwargs) drop_cols_inplace(df, AMM_DROP_COLS) return df def cifar10_amm(): # df = pd.read_csv(os.path.join(RESULTS_DIR, 'cifar10.csv')) # drop_cols_inplace(df, AMM_DROP_COLS) df = _read_amm_csv('cifar10.csv') # print("initial uniq methods:\n", sorted(df['method'].unique())) return _clean_amm_results_df(df) def cifar100_amm(): # df = pd.read_csv(os.path.join(RESULTS_DIR, 'cifar100.csv')) # drop_cols_inplace(df, AMM_DROP_COLS) df = _read_amm_csv('cifar100.csv') return _clean_amm_results_df(df) @_memory.cache def caltech_amm(filt='sobel'): """filt must be one of {'sobel','dog5x5'}""" df = _read_amm_csv('caltech_{}.csv'.format(filt)) return _clean_amm_results_df(df, timing_dtype='i8') @_memory.cache def ucr_amm(k=128, problem='rbf'): """k must be one of {64, 128, 256}""" df = _read_amm_csv('ucr_k={}_problem={}.csv'.format(k, problem)) df['origN'] = df['N'].values df['N'] = 1000 # timing is for a test set size of 1000 if problem == 'softmax': df['M'] = k # to get meaningful timing vs acc comparison return _clean_amm_results_df(df, sparse_pareto=False) def main(): # print(encode_timings()) # print(lut_timings()) # print(scan_timings()) # print(bolt_amm_timings()) # print(mithral_amm_timings()) # print(dense_amm_timings()) # print(osnap_amm_timings()) # print(sparse_amm_timings()) # print(cifar10_amm()) # print(cifar100_amm()) # print(caltech_amm(filt='sobel')) # print(caltech_amm(filt='dog5x5')) # print(ucr_amm(k=64)) # df = ucr_amm(k=128, problem='rbf') # df_bolt = df.loc[df['method'] == 'Bolt'] # print("number of uniq bolt speedups:") # print(df_bolt['Speedup'].unique().size) # import sys; sys.exit() # df = df.loc[df['method'] == 'Bolt'] # print("bolt dset counts") # print(df.groupby('task_id')['task_id'].count()) # # print("bolt speedup per dset counts") # # print(df.groupby('task_id')['Speedup'].unique().count()) # print("number of uniq speedups:") # print(df['Speedup'].unique().size) # df = cifar10_amm() # df = cifar100_amm() df = caltech_amm(filt='dog5x5') # df = caltech_amm(filt='sobel') print(sorted(df['method'].unique())) # # df = df.loc[df['method'].isin(['Brute Force', 'Mithral', 'SparsePCA'])] # df = df.loc[df['method'].isin(['Exact', 'ScalarQuantize', 'MADDNESS', 'SparsePCA'])] # df = df.sort_values(['method', 'Speedup'], axis=0) # print(df['method Speedup Accuracy'.split()]) if __name__ == '__main__': main()
#!/usr/bin/env python import abc import numpy as np from . import vquantizers as vq from . import amm KEY_NLOOKUPS = 'nlookups' class VQMatmul(amm.ApproxMatmul, abc.ABC): def __init__(self, ncodebooks, ncentroids=None): self.ncodebooks = ncodebooks self.ncentroids = (self._get_ncentroids() if ncentroids is None else ncentroids) self.enc = self._create_encoder(ncodebooks) self.reset_for_new_task() @abc.abstractmethod def _create_encoder(self, ncodebooks): # to be overriden by subclasses return vq.PQEncoder(ncodebooks=ncodebooks, ncentroids=self.ncentroids, **self._get_encoder_kwargs()) # @abc.abstractmethod def _get_ncentroids(self): pass @abc.abstractmethod def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): pass def _get_encoder_kwargs(self): # to be overriden by subclasses return {} def reset_for_new_task(self): self.A_enc = None self.luts = None def fit(self, A, B, Y=None): _, D = A.shape if D < self.ncodebooks: raise amm.InvalidParametersException( 'D < C: {} < {}'.format(D, self.ncodebooks)) self.enc.fit(A, B.T) def set_A(self, A): self.A_enc = self.enc.encode_X(A) def set_B(self, B): self.luts = self.enc.encode_Q(B.T) def __call__(self, A, B): if self.A_enc is None: self.set_A(A) if self.luts is None: self.set_B(B) return self.enc.dists_enc(self.A_enc, self.luts) def get_params(self): return {'ncodebooks': self.ncodebooks} # ================================================================ PQ class PQMatmul(VQMatmul): def _create_encoder(self, ncodebooks): # to be overriden by subclasses return vq.PQEncoder(ncodebooks=ncodebooks, ncentroids=self.ncentroids, **self._get_encoder_kwargs()) def _get_ncentroids(self): return 256 def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): # data encoding and LUT costs nmuls = 0 nmuls += 0 if fixedA else A.shape[0] * A.shape[1] * self.ncentroids nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids nlookups = A.shape[0] * B.shape[1] * self.ncodebooks return {amm.KEY_NMULTIPLIES: nmuls, KEY_NLOOKUPS: nlookups} # ================================================================ OPQ class OPQMatmul(PQMatmul): def _get_encoder_kwargs(self): return dict(preproc='OPQ') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) rot_nmuls = A.shape[0] * A.shape[1] * A.shape[1] # OPQ rotation cost metrics[amm.KEY_NMULTIPLIES] += rot_nmuls return metrics # ================================================================ Bolt class BoltMatmul(PQMatmul): # def __init__(self, ncodebooks): # self.ncodebooks = ncodebooks # self.ncentroids = 16 # self.enc = self._create_encoder(self.ncodebooks) # self._reset() def _get_ncentroids(self): return 16 def _create_encoder(self, ncodebooks): return vq.PQEncoder(ncodebooks=ncodebooks, ncentroids=self.ncentroids, quantize_lut=True, # quantize_lut=False, # accumulate_how='mean', accumulate_how='sum', upcast_every=-1, # upcast_every=2, # upcast_every=4, # upcast_every=256, # fine as long as using mean # TODO set quantize_lut=True after debug **self._get_encoder_kwargs()) class GEHTBoltMatmul_CovTopk(BoltMatmul): def _get_encoder_kwargs(self): return dict( preproc='GEHT', sample_how='deterministic', stats_mat='cov') class GEHTBoltMatmul_CovSamp(BoltMatmul): def _get_encoder_kwargs(self): return dict( preproc='GEHT', sample_how='importance', stats_mat='cov') class GEHTBoltMatmul_CorrTopk(BoltMatmul): def _get_encoder_kwargs(self): return dict( preproc='GEHT', sample_how='deterministic', stats_mat='corr') class GEHTBoltMatmul_CorrSamp(BoltMatmul): def _get_encoder_kwargs(self): return dict( preproc='GEHT', sample_how='importance', stats_mat='corr') class BoltSplits(BoltMatmul): def _get_encoder_kwargs(self): return dict( preproc='PQ', encode_algo='splits') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics class BoltMultiSplits(BoltMatmul): def _get_encoder_kwargs(self): return dict(encode_algo='multisplits') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics class BoltPermMultiSplits(BoltMatmul): def _get_encoder_kwargs(self): return dict(preproc='GEHT', encode_algo='multisplits') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics class PQPerm(PQMatmul): def _get_encoder_kwargs(self): return dict(preproc='GEHT') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics class PQMultiSplits(PQMatmul): def __init__(self, ncodebooks, ncentroids=256): super().__init__(ncodebooks=ncodebooks, ncentroids=ncentroids) def _get_encoder_kwargs(self): return dict(encode_algo='multisplits') def get_params(self): return {'ncodebooks': self.ncodebooks, 'ncentroids': self.ncentroids} def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics class PQPermMultiSplits(PQMatmul): def __init__(self, ncodebooks, ncentroids=256): super().__init__(ncodebooks=ncodebooks, ncentroids=ncentroids) def _get_encoder_kwargs(self): return dict(preproc='GEHT', encode_algo='multisplits') def get_params(self): return {'ncodebooks': self.ncodebooks, 'ncentroids': self.ncentroids} def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): metrics = super().get_speed_metrics(A, B, fixedA=fixedA, fixedB=fixedB) nmuls = 0 nmuls += 0 if fixedB else B.shape[0] * B.shape[1] * self.ncentroids metrics[amm.KEY_NMULTIPLIES] = nmuls return metrics # ================================================================ Mithral class OldMithralPQ(PQMatmul): # def _get_ncentroids(self): # return 16 def __init__(self, ncodebooks): super().__init__(ncodebooks=ncodebooks, ncentroids=16) def _create_encoder(self, ncodebooks): return vq.PQEncoder(ncodebooks=ncodebooks, ncentroids=self.ncentroids, encode_algo='multisplits', quantize_lut=True, upcast_every=16, # fine as long as using mean accumulate_how='mean') def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): N, D = A.shape D, M = B.shape # data encoding and LUT costs nmuls = 0 nmuls += 0 if fixedA else N * D # offset + scale before quantize nmuls += 0 if fixedB else M * self.ncentroids * D # lookups given encoded data + luts nlookups = N * M * self.ncodebooks return {amm.KEY_NMULTIPLIES: nmuls, KEY_NLOOKUPS: nlookups} class MithralMatmul(VQMatmul): def __init__(self, ncodebooks, lut_work_const=-1): self.lut_work_const = lut_work_const if (lut_work_const is not None) and (lut_work_const > 0) and ( lut_work_const > ncodebooks): raise amm.InvalidParametersException( "lut_work_const > ncodebooks: {} > {}".format( lut_work_const, ncodebooks)) super().__init__(ncodebooks=ncodebooks, ncentroids=16) # def _get_ncentroids(self): # return 16 # def fit(self, A, B, Y=None): # super().fit(self, A, B, Y=Y) def _create_encoder(self, ncodebooks): return vq.MithralEncoder( ncodebooks=ncodebooks, lut_work_const=self.lut_work_const) def get_params(self): return {'ncodebooks': self.ncodebooks, 'lut_work_const': self.lut_work_const} def get_speed_metrics(self, A, B, fixedA=False, fixedB=False): N, D = A.shape D, M = B.shape # data encoding and LUT costs nmuls = 0 nmuls += 0 if fixedA else N * D # offset + scale before quantize nmuls_per_codebook_per_output = self.ncentroids * D nmuls_per_output = nmuls_per_codebook_per_output * self.ncodebooks nmuls += 0 if fixedB else nmuls_per_output * M # lookups given encoded data + luts nlookups = N * M * self.ncodebooks return {amm.KEY_NMULTIPLIES: nmuls, KEY_NLOOKUPS: nlookups} def set_B(self, B): self.luts, self.offset, self.scale = self.enc.encode_Q(B.T) def __call__(self, A, B): if self.A_enc is None: self.set_A(A) if self.luts is None: self.set_B(B) return self.enc.dists_enc(self.A_enc, self.luts, offset=self.offset, scale=self.scale) class MithralPQ(MithralMatmul): def __init__(self, ncodebooks): super().__init__(ncodebooks=ncodebooks, lut_work_const=1)
#!/bin/env/python """utility functions for running experiments""" from __future__ import print_function, absolute_import import datetime import os import itertools import warnings import numpy as np import pandas as pd import sys import sklearn # from sklearn.model_selection import StratifiedKFold from python.files import ensure_dir_exists try: from joblib import Memory memory = Memory('.', verbose=0) cache = memory.cache except Exception: def cache(f): return f # ================================================================ Constants KEY_FINISHED_UPDATING = '__pyn_finished_updating__' KEY_NEW_KEYS = '__pyn_newkeys__' # ================================================================ Types class UsageError(Exception): pass class Options(object): """Wrapper for a collection to signify that each element is one possible parameter value""" def __init__(self, *args): if args is None or len(args) < 1: raise ValueError("No options given!") if len(args) == 1 and hasattr(args, '__len__'): self.values = args[0] # given a list else: self.values = args # given individual objects def __len__(self): return len(self.values) # deliberately don't act like a collection so that we fail fast if # code doesn't know that this is supposed to represent Options, rather # than a collection of values. This is mostly to ensure that Options # are always expanded out when generating sets of parameters. def __getitem__(self, idx): self._raise() def __setitem__(self, idx, item): self._raise() def _raise(self): raise TypeError("Options object is not a collection; use options.values" " to access the collection of individual options") # ================================================================ Funcs # ------------------------------------------------ misc utils def make_immutable(x): """ >>> make_immutable(5) == 5 True >>> make_immutable('a') == 'a' True >>> make_immutable((1, 2)) == (1, 2) True >>> make_immutable([1, 2]) == [1, 2] False >>> make_immutable([1, 2]) == (1, 2) True """ # must either be not a collections or immutable try: {}[x] = 0 # dicts require immutability return x except TypeError: # so it's mutable; either a collection or a # mutable class; if a class, we're hosed, so # assume it's a collection try: # if it's a singleton collection, try returning # first element; this will jump to except # unless x is a collection _ = len(x) # apparently a collection, so make it a tuple return tuple(x) except TypeError: return repr(x) # not a collection; stringify as last resort def as_key(x): return make_immutable(x) # ------------------------------------------------ IO / saving results def now_as_string(): return datetime.datetime.now().strftime("%Y-%m-%dT%H_%M_%S") def save_data_frame(df, save_dir='results', name="", timestamp='copy', cols_in_filename=None, col_kv_fmt="_{}={}", store_index=False, append=True, dedup_cols=None, add_timestamp_col=True, sort_by=None, **sink): if timestamp == 'copy': # save one copy with and without timestamp kwargs = dict(name=name, col_kv_fmt=col_kv_fmt, cols_in_filename=cols_in_filename, dedup_cols=dedup_cols, store_index=store_index, append=append, sort_by=sort_by, add_timestamp_col=add_timestamp_col) backups_dir = os.path.join(save_dir, 'pyience-backups') save_data_frame(df, timestamp=True, save_dir=backups_dir, **kwargs) save_data_frame(df, timestamp=False, save_dir=save_dir, **kwargs) return # construct filename name = name if name else "" if cols_in_filename: cols = list(df.columns.values) # substrs = ["{%s}" % col for col in cols] # name = name_fmt # for ss in substrs: # key = ss[1:-1] for key in cols_in_filename: if key not in cols: warnings.warn("Column '{}' not found in Dataframe." "Excluding it from filename".format(key)) continue # get value associated with this key; ignored if col not constant vals = df[key] nuniq = len(vals.unique()) if nuniq != 1: warnings.warn("Column '{}' has more than one value in Dataframe." "Excluding it from filename".format(key)) continue val = vals[0] fmt = col_kv_fmt if name == "" and not col_kv_fmt.startswith("{"): fmt = col_kv_fmt[1:] name += fmt.format(key, val) ensure_dir_exists(save_dir) raw_timestamp_str = now_as_string() timestamp_str = ("_" + raw_timestamp_str) if timestamp else "" fileName = "{}{}.csv".format(name, timestamp_str).strip("_") save_path = os.path.join(save_dir, fileName) if add_timestamp_col: df['__pyience_timestamp__'] = [raw_timestamp_str] * df.shape[0] if append and os.path.exists(save_path): existing_df = pd.read_csv(save_path) # print("existing_df_cols", existing_df.columns) # print("existing_df_cols", df.columns) # print("dedup_cols", dedup_cols) df = pd.concat([existing_df, df], axis=0, sort=False, ignore_index=True) # print("df secs: ") # print(df['secs']) dedup_cols = set(dedup_cols) & set(list(df.columns)) df.drop_duplicates(subset=dedup_cols, keep='last', inplace=True) df = df.sort_index(axis=1) if sort_by is not None: df.sort_values(sort_by, inplace=True) # also move these cols to the front for legibility, since they're # probably something you care about other_cols = [col for col in df.columns.values if col not in sort_by] df = df[sort_by + other_cols] df.to_csv(save_path, index=store_index) def save_dicts_as_data_frame(d, **kwargs): if not isinstance(d, dict): try: df = pd.DataFrame.from_records(d) except Exception: dfs = [pd.DataFrame.from_records(dd, index=[0]) for dd in d] df = pd.concat(dfs, axis=0, ignore_index=True) else: df = pd.DataFrame.from_records(d, index=[0]) save_data_frame(df, **kwargs) def generate_save_path(params, savedir, subdir_keys=None): subdir = '' # create nested subdirectories with names specified by # the values for the keys in subdir_keys if subdir_keys is not None: subdir_keys = list(subdir_keys) subdir_names = ["{}__{}".format(str(key), str(params[key])) for key in subdir_keys] subdir = os.path.join(*subdir_names) savedir = os.path.join(savedir, subdir) return savedir # ------------------------------------------------ parameter generation def expand_params(params): """dict of kv pairs -> list of dicts with one option selected for each key whose value is an instance of Options.""" # keys with values that are Options; try all combos of these options_keys = [key for key in params if isinstance(params[key], Options)] options_keys = sorted(options_keys) # sort for reproducibility options_vals = [params[key].values for key in options_keys] # keys with values that aren't Options; these are the same every time no_options_keys = [key for key in params if not isinstance(params[key], Options)] no_options_vals = [params[key] for key in no_options_keys] no_options_params = dict(zip(no_options_keys, no_options_vals)) # make a list of all possible combos of values for each key with Options expanded_params_list = [] for v in itertools.product(*options_vals): expanded_params = dict(zip(options_keys, v)) # pick one option for each expanded_params.update(no_options_params) # add in fixed params expanded_params_list.append(expanded_params) return expanded_params_list def update_func_from_dict(d): def f(params, new_keys, d=d): updated = False for k, v in d.items(): if k in new_keys: for kk, vv in v.items(): updated = updated or (kk not in params) params.setdefault(kk, vv) return updated return f def generate_params_combinations(params_list, update_func={}): """Uses update_func to update each dict based on its values (e.g., to add SVM kernel params if it contains "classifier": "SVM")""" if not isinstance(params_list, (list, set, frozenset, tuple)): params_list = [params_list] for params in params_list: params[KEY_NEW_KEYS] = set(params.keys()) if isinstance(update_func, dict): update_func = update_func_from_dict(update_func) while True: new_list = [] for params in params_list: expanded = expand_params(params) new_list += expanded if not update_func: params_list = new_list break allFinished = True for params in new_list: # if these params aren't fully updated, update them; keep # track of which keys are added along the way so we can # pass this set to the update function next time if not params.get(KEY_FINISHED_UPDATING, False): # read which keys were added last time and which keys # are currently present new_keys = params[KEY_NEW_KEYS] existing_keys = frozenset(params.keys()) params.pop(KEY_NEW_KEYS) unfinished = update_func(params, new_keys) # compute and store which keys were added this time new_keys = frozenset(params.keys()) - existing_keys params[KEY_NEW_KEYS] = new_keys if unfinished: allFinished = False params[KEY_FINISHED_UPDATING] = not unfinished params_list = new_list if allFinished: break for p in params_list: p.pop(KEY_FINISHED_UPDATING) p.pop(KEY_NEW_KEYS) return params_list # ------------------------------------------------ cross validation def stratified_split_train_test(X, Y, train_frac=.8, random_state=123): """Returns X_train, X_test, y_train, y_test""" return sklearn.model_selection.train_test_split( X, Y, train_size=train_frac, stratify=Y, random_state=random_state) def split_train_test(X, Y, train_frac=.8, random_state=123): """Returns X_train, X_test, y_train, y_test""" np.random.seed(123) return sklearn.model_selection.train_test_split( X, Y, train_size=train_frac, random_state=random_state) # n_folds = int(train_frac / (2. - train_frac)) # split = StratifiedKFold(Y, n_folds=n_folds, random_state=12345) # train_index, test_index = next(iter(split)) # X, Xtest = X[train_index], X[test_index] # Y, Ytest = Y[train_index], Y[test_index] # return X, Xtest, Y, Ytest # ------------------------------------------------ Command line def _split_kv_arg(arg): key, val = arg.split('=') return key.strip('-'), val def _is_kv_arg(arg): return len(arg.split('=')) == 2 def _clean_flag_arg(arg): return arg.strip('-') def _is_flag_arg(arg): return arg[0] == '-' def _parse_func_call_cmd(s): """ >>> _parse_func_call_cmd("range(5)") array([0, 1, 2, 3, 4]) >>> _parse_func_call_cmd("range(2, -3, -2)") array([ 2, 0, -2]) >>> _parse_func_call_cmd("linspace( -2,-20, 3)") array([ -2., -11., -20.]) >>> _parse_func_call_cmd("logspace(-1, 3, 3)") array([1.e-01, 1.e+01, 1.e+03]) """ fnames = 'randn randint range linspace logspace'.split() nargs = [(1,), (1, 2, 3), (1, 2, 3), (2, 3), (2, 3)] funcs = [np.random.randn, np.random.randint, np.arange, np.linspace, np.logspace] if not isinstance(s, str): return None for fname, argc, func in zip(fnames, nargs, funcs): if not s.startswith(fname + '('): continue if not s.endswith(')'): raise ValueError("You tried to call function '{}', but forgot the" " closing parenthesis".format(fname)) in_parens = s[len(fname) + 1:-1] maybe_args = in_parens.split(',') if len(maybe_args) not in argc: raise ValueError( "You tried to call function '{}', but passed an invalid number" " of arguments: {}. Needed to be one of: {}" .format( fname, len(maybe_args), argc)) try: nums = [int(arg) for arg in maybe_args] return func(*nums) except: # noqa raise ValueError("Command '{}' has arguments that can't be coerced" " into integers".format(s)) return None def _to_appropriate_type(s): """convert string `s` to an int, bool, float, or integer range as appropriate. Returns the original string if it does not appear to be any of these types.""" if s == 'True' or s == 'T': return True elif s == 'False' or s == 'F': return False try: return int(s) except: # noqa pass try: return float(s) except: # noqa pass if len(s.split('..')) in (2, 3): # range vals_as_strs = s.split('..') try: return np.arange(*[int(val) for val in vals_as_strs]) except: # noqa pass as_func_result = _parse_func_call_cmd(s) if as_func_result is not None: return as_func_result return s def parse_cmd_line(argv=None, positional_keys=None, allow_flags=True, infer_types=True): """Parses the list of command line arguments into a dictionary of key-value pairs Parameters ---------- argv : iterable of strings This should be sys.argv if supplied. Otherwise, sys.argv is read. positional_keys : iterable of strings, optional If k strings are specified, the up to the first k arguments will be treated as values to be paired with these keys. Arguments of the form foo=bar will never be treated this way. allow_flags : bool, optional If True, allows arguments of the form --myArg. When passed, this will add {'myArg': True} to the returned dictionary. This is equivalent to myArg=True infer_types : bool, optional If True, attempts to infer the type of each value in the returned dictionary. E.g., instead of returning {'height': '72'}, it will return {'height': 72}. Returns ------- argKV : dict: string -> inferred type or string A dictionary whose keys and values are specified by the command line arguments >>> # ------------------------ positional args only >>> argv = ['pyience.py', 'fooVal', 'barVal'] >>> d = parse_cmd_line(argv, positional_keys=['fooKey', 'barKey']) >>> len(d) 2 >>> d['fooKey'] 'fooVal' >>> d['barKey'] 'barVal' >>> # ------------------------ key-value args >>> argv = ['pyience.py', 'fooVal', 'bletchKey=bletchVal', 'blahKey=blahVal'] >>> d = parse_cmd_line(argv, positional_keys=['fooKey', 'barKey']) >>> len(d) 3 >>> d['fooKey'] 'fooVal' >>> d.get('barKey', 'notHere') 'notHere' >>> d['bletchKey'] 'bletchVal' >>> d['blahKey'] 'blahVal' >>> # ------------------------ flags >>> argv = ['pyience.py', 'fooVal', 'bletchKey=bletchVal', '--myFlag'] >>> d = parse_cmd_line(argv, positional_keys=['fooKey', 'barKey']) >>> d['myFlag'] True >>> # ------------------------ type inference >>> argv = ['pyience.py', '--myFlag', 'foo=1.1', 'bar=7', 'baz=T', 'r=1..5'] >>> d = parse_cmd_line(argv, positional_keys=['fooKey', 'barKey']) >>> len(d) 5 >>> d['myFlag'] True >>> d['foo'] 1.1 >>> d['bar'] 7 >>> d['baz'] True >>> d['r'] array([1, 2, 3, 4]) >>> # ------------------------ no positional args >>> d = parse_cmd_line(argv) >>> len(d) 5 >>> d['myFlag'] True >>> d['foo'] 1.1 """ if argv is None: argv = sys.argv args = argv[1:] # ignore file name num_positional_keys = 0 if positional_keys is not None and len(positional_keys): num_positional_keys = len(positional_keys) # validate input; keyword arguments must come after positional # arguments, and there must be no more positional arguments than # we have keys to associate with them kwargs_started = False flags_started = False for i, arg in enumerate(args): if _is_kv_arg(arg): # it's a keyword argument kwargs_started = True elif _is_flag_arg(arg): flags_started = True else: # it's not a keyword argument or flag arguemnt if kwargs_started: raise UsageError("key=value arguments must come after" "positional arguments!") if flags_started: raise UsageError("flag (e.g., --myFlag) arguments must come" "after positional arguments!") arg_num = i + 1 if arg_num > num_positional_keys: raise UsageError("only expecting " "{} positional arguments!".format( num_positional_keys)) argKV = {} for i, arg in enumerate(args): if _is_kv_arg(arg): key, val = _split_kv_arg(arg) argKV[key] = val elif _is_flag_arg(arg): key = _clean_flag_arg(arg) argKV[key] = 'True' # string so that all vals are strings elif i < num_positional_keys: key = positional_keys[i] argKV[key] = arg else: raise UsageError("couldn't parse argument '{}'".format(arg)) if infer_types: for k, v in argKV.items(): argKV[k] = _to_appropriate_type(v) return argKV # ------------------------------------------------ other stuff def apply_funcs(funcs, data): f = chain(funcs) return f(data) def chain(funcs): if funcs is None or not len(funcs): return lambda x: x def f(*args, **kwargs): res = funcs[0](*args, **kwargs) for func in funcs[1:]: res = func(res) return f def subdict(d, keys): """Returns a new dictionary composed of the (key, value) pairs from d for the keys specified in keys""" return {k: d[k] for k in keys} # ------------------------------------------------ sklearn interop def set_attrs(obj, attrs_dict, require_attrs_exist=False): if require_attrs_exist: keys_and_there = ([(k, k in obj.__dict__) for k in attrs_dict]) missing_keys = [k for (k, there) in keys_and_there if not there] there = zip(*keys_and_there)[1] if not all(there): raise ValueError("Object is missing keys {}".format( missing_keys)) obj.__dict__.update(attrs_dict) # ------------------------------------------------ cross validation def _uniq_element_positions(iterable): """ Returns a mapping of unique elements to positions at which they occur within the iterable """ objs2positions = {} for i, obj in enumerate(iterable): key = as_key(obj) positions = objs2positions.get(key, []) positions.append(i) objs2positions[key] = positions return objs2positions # def _group_start_idxs_eq_split(nelements, ngroups): # group_sz = nelements // ngroups # return np.arange(0, nelements, group_sz, dtype=np.int) def _group_start_end_idxs(nelements, ngroups=-1, fractions=None): hasFracs = fractions is not None and len(fractions) if ngroups <= 1 and not hasFracs: return np.array([0], dtype=np.int), np.array([nelements], dtype=np.int) if not hasFracs: fracs = np.ones(ngroups) fractions = np.asarray(fracs) fractions /= np.max(fracs) cum_fracs = np.cumsum(fractions) end_idxs = (nelements * cum_fracs).astype(np.int) start_idxs = np.r_[0, end_idxs[:-1]] return start_idxs, end_idxs def _split_into_groups(iterable, ngroups=-1, fractions=None, shuffle=True): if shuffle: iterable = np.copy(iterable) np.shuffle(iterable) start_idxs, end_idxs = _group_start_end_idxs(len(iterable), ngroups, fractions) return [iterable[start:end] for start, end in zip(start_idxs, end_idxs)] def cv_partition_idxs(labels, n_folds=5, fractions=None, stratified=True): if fractions is not None and len(fractions): if len(fractions) != n_folds: raise ValueError("Specified fractions of total for {} groups, but " "n_folds is {}; ignoring n_fold".format( len(fractions), n_folds)) if stratified: all_idxs = [[] for i in range(n_folds)] lbl2idxs = _uniq_element_positions(labels) for lbl, idxs in lbl2idxs.items(): if len(idxs) < n_folds: warnings.warn(("Label {} appears only {} times, which is " "less than the number of folds requested, {}" .format(lbl, len(idxs), n_folds)), Warning) idxGroups = _split_into_groups(idxs, n_folds, fractions) for i, group in enumerate(idxGroups): all_idxs[i] += group return all_idxs else: possible_idxs = np.arange(len(labels)) return _split_into_groups(possible_idxs, n_folds, fractions) def cv_split(X, y, n_folds=5, fractions=None, stratified=True): if len(X) != len(y): raise IndexError("len(X) {} != len(y) {}".format(len(X), len(y))) all_idxs = cv_partition_idxs(y, n_folds=n_folds, fractions=fractions, stratified=stratified) X_split = [X[idxs] for idxs in all_idxs] y_split = [y[idxs] for idxs in all_idxs] return X_split, y_split # ================================================================ Main def update(params, new_keys): if 'classifier' in new_keys: params['kernel'] = Options('rbf', 'linear') # we use setdefault here so that we don't overwrite values # passed in at the top level if 'kernel' in new_keys: kernel = params['kernel'] params.setdefault('C', Options(10. ** np.arange(-5, 3))) if kernel == 'rbf': params.setdefault('gamma', Options([1, 10])) return True if new_keys else False def main(): cVals = 10. ** np.arange(-3, 3) d = {"classifier": "SVM", 'C': Options(cVals)} # generate_params_combinations(d, update) combos = generate_params_combinations(d, update) # add a fake outcome variable for combo in combos: combo['runtime'] = np.random.rand() * 10 # print out a dataframe so we can see that this worked import pandas as pd print(pd.DataFrame.from_records(combos)) # woot; it worked if __name__ == '__main__': from doctest import testmod testmod() main()
#!/usr/bin/env python import functools import matplotlib.pyplot as plt import numpy as np import os from scipy.stats.stats import pearsonr import seaborn as sb import time from collections import namedtuple # import datasets import files import product_quantize as pq import pyience as pyn from datasets import neighbors as dsets from utils import kmeans, top_k_idxs from joblib import Memory _memory = Memory('.', verbose=0) np.set_printoptions(precision=3) SAVE_DIR = '../results' # ================================================================ Distances def dists_elemwise_sq(x, q): diffs = x - q return diffs * diffs def dists_elemwise_l1(x, q): return np.abs(x - q) def dists_elemwise_dot(x, q): return x * q # ================================================================ Clustering def load_dataset_object(which_dataset, **load_dataset_kwargs): X_train, Q, X_test, true_nn = dsets.load_dataset( which_dataset, **load_dataset_kwargs) assert Q.shape[-1] == X_train.shape[-1] if isinstance(which_dataset, str): name = files.basename(which_dataset, noext=True) else: name = which_dataset.__name__ # assumes which_dataset is a class return Dataset(Q, X_train, X_test, true_nn, name) Dataset = namedtuple('Dataset', [ 'Q', 'X_train', 'X_test', 'true_nn', 'name']) # ================================================================ Quantizers # ------------------------------------------------ Product Quantization def _learn_centroids(X, ncentroids, nsubvects, subvect_len): ret = np.empty((ncentroids, nsubvects, subvect_len)) for i in range(nsubvects): start_col = i * subvect_len end_col = start_col + subvect_len X_in = X[:, start_col:end_col] centroids, labels = kmeans(X_in, ncentroids) ret[:, i, :] = centroids return ret def _parse_codebook_params(D, code_bits=-1, bits_per_subvect=-1, nsubvects=-1): if nsubvects < 0: nsubvects = code_bits // bits_per_subvect elif code_bits < 1: code_bits = bits_per_subvect * nsubvects elif bits_per_subvect < 1: bits_per_subvect = code_bits // nsubvects ncentroids = int(2 ** bits_per_subvect) subvect_len = D // nsubvects assert code_bits % bits_per_subvect == 0 if D % subvect_len: print("D, nsubvects, subvect_len = ", D, nsubvects, subvect_len) assert D % subvect_len == 0 # TODO rm this constraint return nsubvects, ncentroids, subvect_len def _fit_pq_lut(q, centroids, elemwise_dist_func): _, nsubvects, subvect_len = centroids.shape assert len(q) == nsubvects * subvect_len q = q.reshape((1, nsubvects, subvect_len)) q_dists_ = elemwise_dist_func(centroids, q) q_dists_ = np.sum(q_dists_, axis=-1) return np.asfortranarray(q_dists_) # ncentroids, nsubvects, col-major class PQEncoder(object): def __init__(self, dataset, code_bits=-1, bits_per_subvect=-1, nsubvects=-1, elemwise_dist_func=dists_elemwise_sq): X = dataset.X_train self.elemwise_dist_func = elemwise_dist_func tmp = _parse_codebook_params(X.shape[1], code_bits=code_bits, bits_per_subvect=bits_per_subvect, nsubvects=nsubvects) self.nsubvects, self.ncentroids, self.subvect_len = tmp self.code_bits = int(np.log2(self.ncentroids)) # for fast lookups via indexing into flattened array self.offsets = np.arange(self.nsubvects, dtype=np.int) * self.ncentroids self.centroids = _learn_centroids(X, self.ncentroids, self.nsubvects, self.subvect_len) def name(self): return "PQ_{}x{}b".format(self.nsubvects, self.code_bits) def params(self): return {'_algo': 'PQ', '_ncodebooks': self.nsubvects, '_code_bits': self.code_bits} def encode_X(self, X, **sink): idxs = pq._encode_X_pq(X, codebooks=self.centroids) return idxs + self.offsets # offsets let us index into raveled dists def encode_q(self, q, **sink): return None # we use fit_query() instead, so fail fast def dists_true(self, X, q): return np.sum(self.elemwise_dist_func(X, q), axis=-1) def fit_query(self, q, **sink): self.q_dists_ = _fit_pq_lut(q, centroids=self.centroids, elemwise_dist_func=self.elemwise_dist_func) def dists_enc(self, X_enc, q_unused=None): # this line has each element of X_enc index into the flattened # version of q's distances to the centroids; we had to add # offsets to each col of X_enc above for this to work centroid_dists = self.q_dists_.T.ravel()[X_enc.ravel()] return np.sum(centroid_dists.reshape(X_enc.shape), axis=-1) def _learn_best_quantization(luts): # luts can be a bunch of vstacked luts best_loss = np.inf best_alpha = None best_floors = None best_scale_by = None for alpha in [.001, .002, .005, .01, .02, .05, .1]: alpha_pct = int(100 * alpha) # compute quantized luts this alpha would yield floors = np.percentile(luts, alpha_pct, axis=0) luts_offset = np.maximum(0, luts - floors) ceil = np.percentile(luts_offset, 100 - alpha_pct) scale_by = 255. / ceil luts_quantized = np.floor(luts_offset * scale_by).astype(np.int) luts_quantized = np.minimum(255, luts_quantized) # compute err luts_ideal = (luts - luts_offset) * scale_by diffs = luts_ideal - luts_quantized loss = np.sum(diffs * diffs) if loss <= best_loss: best_loss = loss best_alpha = alpha best_floors = floors best_scale_by = scale_by return best_floors, best_scale_by, best_alpha class OPQEncoder(PQEncoder): def __init__(self, dataset, code_bits=-1, bits_per_subvect=-1, nsubvects=-1, elemwise_dist_func=dists_elemwise_sq, opq_iters=20, quantize_lut=False, algo='OPQ', **opq_kwargs): X = dataset.X_train self.elemwise_dist_func = elemwise_dist_func self.quantize_lut = quantize_lut self.opq_iters = opq_iters self.algo = algo tmp = _parse_codebook_params(X.shape[1], code_bits=code_bits, bits_per_subvect=bits_per_subvect, nsubvects=nsubvects) self.nsubvects, self.ncentroids, self.subvect_len = tmp self.code_bits = int(np.log2(self.ncentroids)) # for fast lookups via indexing into flattened array self.offsets = np.arange(self.nsubvects, dtype=np.int) * self.ncentroids if self.algo == 'Bolt': # Note: we always pass in 0 iters in the reported experiments, # so it never rotates anything self.centroids, _, self.rotations = pq.learn_bopq( X, ncodebooks=nsubvects, codebook_bits=bits_per_subvect, niters=opq_iters, **opq_kwargs) elif self.algo == 'OPQ': self.centroids, _, self.R = pq.learn_opq( X, ncodebooks=nsubvects, codebook_bits=bits_per_subvect, niters=opq_iters, **opq_kwargs) else: raise ValueError("argument algo must be one of {OPQ, Bolt}") # learn appropriate offsets and shared scale factor for quantization self.lut_offsets = np.zeros(self.nsubvects) self.order_idxs = np.arange(self.nsubvects, dtype=np.int) if self.quantize_lut: # TODO put this logic in separate function print("learning quantization...") num_rows = min(10*1000, len(X) // 2) _, queries = dsets.extract_random_rows( X[num_rows:], how_many=1000, remove_from_X=False) X = X[:num_rows] # limit to first 10k rows of X # compute luts for all the queries luts = [self._fit_query(q, quantize=False) for q in queries] luts = np.vstack(luts) assert luts.shape == (self.ncentroids * len(queries), self.nsubvects) self.lut_offsets, self.scale_by, _ = _learn_best_quantization(luts) def name(self): return "{}_{}x{}b_iters={}_quantize={}".format( self.algo, self.nsubvects, self.code_bits, self.opq_iters, int(self.quantize_lut)) def params(self): return {'_algo': self.algo, '_ncodebooks': self.nsubvects, '_code_bits': self.code_bits, 'opq_iters': self.opq_iters, '_quantize': self.quantize_lut} def _fit_query(self, q, quantize=False): if self.algo == 'OPQ': qR = pq.opq_rotate(q, self.R).ravel() elif self.algo == 'Bolt': qR = pq.bopq_rotate(q, self.rotations).ravel() lut = _fit_pq_lut(qR, centroids=self.centroids, elemwise_dist_func=self.elemwise_dist_func) if quantize: if False: # roughly laplace distro, reaching all the way to 0 ax = sb.distplot(lut.ravel(), hist=False, rug=True) ax.set_xlabel('Query dist to centroids (lut dist histogram)') ax.set_ylabel('Fraction of queries') plt.show() lut = np.maximum(0, lut - self.lut_offsets) lut = np.floor(lut * self.scale_by).astype(np.int) return np.minimum(lut, 255) return lut def encode_X(self, X, **sink): if self.algo == 'OPQ': X = pq.opq_rotate(X, self.R) elif self.algo == 'Bolt': X = pq.bopq_rotate(X, self.rotations) idxs = pq._encode_X_pq(X, codebooks=self.centroids) return idxs + self.offsets # offsets let us index into raveled dists def fit_query(self, q, quantize=True, **sink): quantize = quantize and self.quantize_lut self.q_dists_ = self._fit_query(q, quantize=quantize) if quantize: # print "min, max lut values: {}, {}".format(np.min(self.q_dists_), # np.max(self.q_dists_)) assert np.min(self.q_dists_) >= 0 assert np.max(self.q_dists_) <= 255 if False: _, axes = plt.subplots(3, figsize=(9, 11)) sb.violinplot(data=self.q_dists_, inner="box", cut=0, ax=axes[0]) axes[0].set_xlabel('Codebook') axes[0].set_ylabel('Distance to query') axes[0].set_ylim([0, np.max(self.q_dists_)]) sb.heatmap(data=self.q_dists_, ax=axes[1], cbar=False, vmin=0) axes[1].set_xlabel('Codebook') axes[1].set_ylabel('Centroid') sb.distplot(self.q_dists_.ravel(), hist=False, rug=True, vertical=False, ax=axes[2]) axes[2].set_xlabel('Centroid dist to query') axes[2].set_ylabel('Fraction of centroids') axes[2].set_xlim([0, np.max(self.q_dists_) + .5]) # plot where the mean is mean_dist = np.mean(self.q_dists_) ylim = axes[2].get_ylim() axes[2].plot([mean_dist, mean_dist], ylim, 'r--') axes[2].set_ylim(ylim) plt.show() # ================================================================ Main def eval_encoder(dataset, encoder, dist_func_true=None, dist_func_enc=None, eval_dists=True, verbosity=1, plot=False, smaller_better=True): X = dataset.X_test queries = dataset.Q true_nn = dataset.true_nn if true_nn is not None: print("eval encoder(): got true_nn with shape: ", true_nn.shape) queries = queries[:1000] # TODO rm for tables; fine for plots print("queries.shape", queries.shape) need_true_dists = eval_dists or plot or true_nn is None if len(queries.shape) == 1: queries = [queries] if dist_func_true is None: dist_func_true = encoder.dists_true if dist_func_enc is None: dist_func_enc = encoder.dists_enc t0 = time.time() # performance metrics RECALL_Rs = [1, 5, 10, 50, 100, 500, 1000] recall_counts = np.zeros(len(RECALL_Rs)) fracs_below_max = [] if eval_dists: all_corrs = [] all_rel_errs = [] all_errs = [] total_dist = 0. if need_true_dists: X = X[:10000] # limit to 10k points because otherwise it takes forever queries = queries[:256, :] print("encoding X...") X_enc = encoder.encode_X(X) print("trying queries...") for i, q in enumerate(queries): if i % 100 == 0: print("trying query {}...".format(i)) q_enc = encoder.encode_q(q) encoder.fit_query(q) if need_true_dists: all_true_dists = dist_func_true(X, q) all_enc_dists = dist_func_enc(X_enc, q_enc) # ------------------------ begin analysis / reporting code # find true knn if need_true_dists: knn_idxs = top_k_idxs(all_true_dists, 10, smaller_better=smaller_better) else: knn_idxs = true_nn[i, :10] # compute fraction of points with enc dists as close as 10th nn knn_enc_dists = all_enc_dists[knn_idxs] if smaller_better: max_enc_dist = np.max(knn_enc_dists) num_below_max = np.sum(all_enc_dists <= max_enc_dist) else: max_enc_dist = np.min(knn_enc_dists) num_below_max = np.sum(all_enc_dists >= max_enc_dist) frac_below_max = float(num_below_max) / len(all_enc_dists) fracs_below_max.append(frac_below_max) # compute recall@R stats top_1000 = top_k_idxs(all_enc_dists, 1000, smaller_better=smaller_better) nn_idx = knn_idxs[0] for i, r in enumerate(RECALL_Rs): recall_counts[i] += nn_idx in top_1000[:r] # compute distortion in distances, quantified by corr and rel err if eval_dists: total_dist += np.sum(all_true_dists) corr, _ = pearsonr(all_enc_dists, all_true_dists) all_corrs.append(corr) rel_errs = (all_enc_dists - all_true_dists) / all_true_dists all_rel_errs.append(rel_errs) all_errs.append(all_enc_dists - all_true_dists) assert not np.any(np.isinf(all_enc_dists)) assert not np.any(np.isnan(all_enc_dists)) assert not np.any(np.isinf(all_true_dists)) assert not np.any(np.isnan(all_true_dists)) if plot and i < 3: # at most 3 plots num_nn = min(10000, len(all_true_dists) - 1) xlim = [0, np.partition(all_true_dists, num_nn)[num_nn]] ylim = [0, np.partition(all_enc_dists, num_nn)[num_nn]] grid = sb.jointplot(x=all_true_dists, y=all_enc_dists, xlim=xlim, ylim=ylim, joint_kws=dict(s=10)) # hack to bully the sb JointGrid into plotting a vert line cutoff = all_true_dists[knn_idxs[-1]] grid.x = [cutoff, cutoff] grid.y = ylim grid.plot_joint(plt.plot, color='r', linestyle='--') # also make it plot cutoff in terms of quantized dist grid.x = xlim grid.y = [max_enc_dist, max_enc_dist] grid.plot_joint(plt.plot, color='k', linestyle='--') if plot: plt.show() t = time.time() - t0 # log a lot of performance metrics / experimental params detailed_stats = [] # list of dicts stats = {} stats['X_rows'] = X.shape[0] stats['X_cols'] = X.shape[1] stats['nqueries'] = len(queries) stats['eval_time_secs'] = t stats['fracs_below_max_mean'] = np.mean(fracs_below_max) stats['fracs_below_max_std'] = np.std(fracs_below_max) stats['fracs_below_max_50th'] = np.median(fracs_below_max) stats['fracs_below_max_90th'] = np.percentile(fracs_below_max, q=90) for i, r in enumerate(RECALL_Rs): key = 'recall@{}'.format(r) val = float(recall_counts[i]) / len(queries) stats[key] = val if eval_dists: corrs = np.hstack(all_corrs) rel_errs = np.hstack(all_rel_errs) rel_errs = rel_errs[~(np.isnan(rel_errs) + np.isinf(rel_errs))] errs = np.hstack(all_errs) stats['corr_mean'] = np.mean(all_corrs) stats['corr_std'] = np.std(all_corrs) stats['mse_mean'] = np.mean(errs * errs) stats['mse_std'] = np.std(errs * errs) stats['rel_err_mean'] = np.mean(rel_errs) stats['rel_err_std'] = np.std(rel_errs) stats['rel_err_sq_mean'] = np.mean(rel_errs * rel_errs) stats['rel_err_sq_std'] = np.std(rel_errs * rel_errs) # sample some relative errs cuz all we need them for is plotting # confidence intervals np.random.shuffle(rel_errs) np.random.shuffle(errs) detailed_stats = [{'corr': all_corrs[i], 'rel_err': rel_errs[i], 'err': errs[i]} for i in range(len(corrs))] for d in detailed_stats: d.update(encoder_params(encoder)) if verbosity > 0: print("------------------------ {}".format(name_for_encoder(encoder))) keys = sorted(stats.keys()) lines = ["{}: {}".format(k, stats[k]) for k in keys if isinstance(stats[k], str)] lines += ["{}: {:.4g}".format(k, stats[k]) for k in keys if not isinstance(stats[k], str)] print("\n".join(lines)) stats.update(encoder_params(encoder)) return stats, detailed_stats # detailed_stats empty unless `eval_dists` def name_for_encoder(encoder): try: return encoder.name() except AttributeError: return str(type(encoder)) def encoder_params(encoder): try: return encoder.params() except AttributeError: return {'algo': name_for_encoder(encoder)} # @_memory.cache def _experiment_one_dataset(which_dataset, eval_dists=False, dotprods=False, save_dir=None): SAVE_DIR = save_dir if save_dir else '../results/acc/' elemwise_dist_func = dists_elemwise_dot if dotprods else dists_elemwise_sq smaller_better = not dotprods N, D = -1, -1 num_queries = -1 # no effect for "real" datasets if isinstance(which_dataset, str): print("WARNING: sampling queries from data file") num_queries = 128 # if just loading one file, need to sample queries norm_len = False # set to true for cosine similarity norm_mean = True max_ncodebooks = 64 # 32B bolt has 64 codebooks dataset_func = functools.partial(load_dataset_object, N=N, D=D, num_queries=num_queries, norm_len=norm_len, norm_mean=norm_mean, D_multiple_of=max_ncodebooks) dataset = dataset_func(which_dataset) print("=== Using Dataset: {} ({}x{})".format(dataset.name, N, D)) dicts = [] detailed_dicts = [] nbytes_list = [8, 16, 32] # max_opq_iters = 5 # TODO uncomment below max_opq_iters = 20 # ------------------------------------------------ Bolt # Note: we considered having learned rotations like OPQ but constrained # to be block diagonal; this is why you'll see mentions of rotations # in some of the Bolt code. However, it ended up not helping at all # and also slows down Bolt considerably. All the reported results are # without any rotation. # rotation_sizes = [8, 16, 32] rotation_sizes = [32] # rotation_sizes = [16] for nbytes in nbytes_list: for opq_iters in [0]: # 0 opq iters -> no rotations rot_sizes = rotation_sizes if opq_iters > 0 else [16] for rot_sz in rot_sizes: nsubvects = nbytes * 2 encoder = OPQEncoder(dataset, nsubvects=nsubvects, bits_per_subvect=4, opq_iters=opq_iters, R_sz=rot_sz, elemwise_dist_func=elemwise_dist_func, algo='Bolt', quantize_lut=True) stats, detailed_stats = eval_encoder( dataset, encoder, eval_dists=eval_dists, smaller_better=smaller_better) stats['rot_sz'] = rot_sz for d in detailed_dicts: d['rot_sz'] = rot_sz dicts.append(stats) detailed_dicts += detailed_stats # ------------------------------------------------ PQ # for codebook_bits in [4, 8]: for codebook_bits in [8]: for nbytes in nbytes_list: nsubvects = nbytes * (8 // codebook_bits) encoder = PQEncoder(dataset, nsubvects=nsubvects, bits_per_subvect=codebook_bits, elemwise_dist_func=elemwise_dist_func) stats, detailed_stats = eval_encoder( dataset, encoder, eval_dists=eval_dists, smaller_better=smaller_better) dicts.append(stats) detailed_dicts += detailed_stats # ------------------------------------------------ OPQ init = 'identity' opq_iters = max_opq_iters for codebook_bits in [8]: for nbytes in nbytes_list: nsubvects = nbytes * (8 // codebook_bits) encoder = OPQEncoder(dataset, nsubvects=nsubvects, bits_per_subvect=codebook_bits, opq_iters=opq_iters, init=init, elemwise_dist_func=elemwise_dist_func) stats, detailed_stats = eval_encoder( dataset, encoder, eval_dists=eval_dists, smaller_better=smaller_better) dicts.append(stats) detailed_dicts += detailed_stats for d in dicts: d['dataset'] = dataset.name d['norm_mean'] = norm_mean for d in detailed_dicts: d['dataset'] = dataset.name d['norm_mean'] = norm_mean savedir = os.path.join(SAVE_DIR, dataset.name) pyn.save_dicts_as_data_frame(dicts, savedir, name='summary') # also just save versions with timestamps to recover from clobbering pyn.save_dicts_as_data_frame(dicts, savedir, name='summary', timestamp=True) if eval_dists: pyn.save_dicts_as_data_frame(detailed_dicts, savedir, name='all_results') pyn.save_dicts_as_data_frame(detailed_dicts, savedir, name='all_results', timestamp=True) return dicts, detailed_dicts def experiment(eval_dists=False, dotprods=False): # which_datasets = [dsets.Mnist] which_datasets = [dsets.Mnist, dsets.Sift1M, dsets.LabelMe, dsets.Convnet1M] # which_datasets = [dsets.Glove] # which_datasets = [dsets.Deep1M, dsets.Gist] if eval_dists: save_dir = '../results/acc_dotprods/' if dotprods else '../results/acc_l2' else: save_dir = '../results/recall_at_r/' for which_dataset in which_datasets: _dicts, _details = _experiment_one_dataset( which_dataset, eval_dists=eval_dists, dotprods=dotprods, save_dir=save_dir) def main(): import doctest doctest.testmod() np.set_printoptions(precision=3) opts = pyn.parse_cmd_line() opts.setdefault('eval_l2_dists', False) opts.setdefault('eval_dotprods', False) opts.setdefault('eval_recall@R', False) if opts['eval_l2_dists']: print(">>>>>>>> evaluating l2 dists") experiment(eval_dists=True, dotprods=False) if opts['eval_dotprods']: print(">>>>>>>> evaluating dot prods") experiment(eval_dists=True, dotprods=True) if opts['eval_recall@R']: print(">>>>>>>> evaluating recall@R") experiment(eval_dists=False, dotprods=False) return if __name__ == '__main__': main()
#!/bin/env/python import copy import numpy as np from functools import reduce import numba from sklearn.decomposition import PCA from sklearn import linear_model from . import subspaces as subs from joblib import Memory _memory = Memory('.', verbose=0) # def bucket_id_to_new_bucket_ids(old_id): # i = 2 * old_id # return i, i + 1 class Bucket(object): __slots__ = 'N D id sumX sumX2 point_ids support_add_and_remove'.split() def __init__(self, D=None, N=0, sumX=None, sumX2=None, point_ids=None, bucket_id=0, support_add_and_remove=False): # self.reset(D=D, sumX=sumX, sumX2=sumX2) # assert point_ids is not None if point_ids is None: assert N == 0 point_ids = (set() if support_add_and_remove else np.array([], dtype=np.int)) self.N = len(point_ids) self.id = bucket_id # this is just so that we can store the point ids as array instead of # set, while still retaining option to run our old code that needs # them to be a set for efficient inserts and deletes self.support_add_and_remove = support_add_and_remove if support_add_and_remove: self.point_ids = set(point_ids) else: self.point_ids = np.asarray(point_ids) # figure out D if (D is None or D < 1) and (sumX is not None): D = len(sumX) elif (D is None or D < 1) and (sumX2 is not None): D = len(sumX2) assert D is not None self.D = D # figure out + sanity check stats arrays self.sumX = np.zeros(D, dtype=np.float32) if (sumX is None) else sumX self.sumX2 = np.zeros(D, dtype=np.float32) if (sumX2 is None) else sumX2 # noqa # print("D: ", D) # print("sumX type: ", type(sumX)) assert len(self.sumX) == D assert len(self.sumX2) == D self.sumX = np.asarray(self.sumX).astype(np.float32) self.sumX2 = np.asarray(self.sumX2).astype(np.float32) def add_point(self, point, point_id=None): assert self.support_add_and_remove # TODO replace with more numerically stable updates if necessary self.N += 1 self.sumX += point self.sumX2 += point * point if point_id is not None: self.point_ids.add(point_id) def remove_point(self, point, point_id=None): assert self.support_add_and_remove self.N -= 1 self.sumX -= point self.sumX2 -= point * point if point_id is not None: self.point_ids.remove(point_id) def deepcopy(self, bucket_id=None): # deep copy bucket_id = self.id if bucket_id is None else bucket_id return Bucket( sumX=np.copy(self.sumX), sumX2=np.copy(self.sumX2), point_ids=copy.deepcopy(self.point_ids), bucket_id=bucket_id) def split(self, X=None, dim=None, val=None, X_orig=None): id0 = 2 * self.id id1 = id0 + 1 if X is None or self.N < 2: # copy of this bucket + an empty bucket return (self.deepcopy(bucket_id=id0), Bucket(D=self.D, bucket_id=id1)) assert dim is not None assert val is not None assert self.point_ids is not None my_idxs = np.asarray(self.point_ids) # print("my_idxs shape, dtype", my_idxs.shape, my_idxs.dtype) X = X[my_idxs] X_orig = X if X_orig is None else X_orig[my_idxs] mask = X_orig[:, dim] < val not_mask = ~mask X0 = X[mask] X1 = X[not_mask] ids0 = my_idxs[mask] ids1 = my_idxs[not_mask] def create_bucket(points, ids, bucket_id): sumX = points.sum(axis=0) if len(ids) else None sumX2 = (points * points).sum(axis=0) if len(ids) else None # return Bucket(N=len(ids), D=self.D, point_ids=ids, return Bucket(D=self.D, point_ids=ids, sumX=sumX, sumX2=sumX2, bucket_id=bucket_id) return create_bucket(X0, ids0, id0), create_bucket(X1, ids1, id1) def optimal_split_val(self, X, dim, possible_vals=None, X_orig=None, return_possible_vals_losses=False): if self.N < 2 or self.point_ids is None: if return_possible_vals_losses: return 0, 0, np.zeros(len(possible_vals), dtype=X.dtype) return 0, 0 # my_idxs = np.array(list(self.point_ids)) my_idxs = np.asarray(self.point_ids) if X_orig is not None: X_orig = X_orig[my_idxs] return optimal_split_val( X[my_idxs], dim, possible_vals=possible_vals, X_orig=X_orig, return_possible_vals_losses=return_possible_vals_losses) def col_means(self): return self.sumX.astype(np.float64) / max(1, self.N) def col_variances(self, safe=False): if self.N < 1: return np.zeros(self.D, dtype=np.float32) E_X2 = self.sumX2 / self.N E_X = self.sumX / self.N ret = E_X2 - (E_X * E_X) return np.maximum(0, ret) if safe else ret def col_sum_sqs(self): return self.col_variances() * self.N @property def loss(self): # if self.N < 1: # return 0 # # less stable version with one less divide and mul # return max(0, np.sum(self.sumX2 - (self.sumX * (self.sumX / self.N)))) # more stable version, that also clamps variance at 0 return max(0, np.sum(self.col_sum_sqs())) # expected_X = self.sumX / self.N # expected_X2 = self.sumX2 / self.N # return max(0, np.sum(expected_X2 - (expected_X * expected_X)) * self.N) # @numba.jit(nopython=True) # numpy cumsum in insanely slow # def _cumsum_cols(X): # X = np.copy(X) # for i in range(1, X.shape[0]): # X[i] += X[i - 1] # return X # numpy cumsum in insanely slow; also, having the nested loops is twice # as fast as assigning rows (ie, X[i] += X[i-1]) @numba.njit(fastmath=True) def _cumsum_cols(X): out = np.empty(X.shape, X.dtype) for j in range(X.shape[1]): out[0, j] = X[0, j] for i in range(1, X.shape[0]): for j in range(X.shape[1]): out[i, j] = X[i, j] + out[i - 1, j] return out @numba.njit(fastmath=True, cache=True) # njit = no python, cache binary def _cumsse_cols(X): N, D = X.shape cumsses = np.empty((N, D), X.dtype) cumX_row = np.empty(D, X.dtype) cumX2_row = np.empty(D, X.dtype) for j in range(D): cumX_row[j] = X[0, j] cumX2_row[j] = X[0, j] * X[0, j] cumsses[0, j] = 0 # no err in bucket with 1 element for i in range(1, N): one_over_count = 1. / (i + 1) for j in range(D): cumX_row[j] += X[i, j] cumX2_row[j] += X[i, j] * X[i, j] meanX = cumX_row[j] * one_over_count cumsses[i, j] = cumX2_row[j] - (cumX_row[j] * meanX) return cumsses # def optimal_split_val(X, dim, possible_vals=None, return_val_idx=False): # @_memory.cache def optimal_split_val(X, dim, possible_vals=None, X_orig=None, # return_possible_vals_losses=False, force_val='median'): return_possible_vals_losses=False, force_val=None, # shrink_towards_median=True): shrink_towards_median=False): X_orig = X if X_orig is None else X_orig # X_orig = X # TODO rm if X_orig.shape != X.shape: print("X orig shape: ", X_orig.shape) print("X shape: ", X.shape) assert X_orig.shape == X.shape if force_val in ('mean', 'median'): assert not return_possible_vals_losses x = X_orig[:, dim] val = np.median(x) if force_val == 'median' else np.mean(x) mask = X_orig < val X0 = X[mask] errs0 = X0 - X0.mean(axis=0) loss0 = np.sum(errs0 * errs0) X1 = X[~mask] errs = X1 - X1.mean(axis=0) loss1 = np.sum(errs * errs) return val, loss0 + loss1 N, D = X.shape # sort_idxs = np.argsort(X[:, dim]) sort_idxs = np.argsort(X_orig[:, dim]) X_sort = X[sort_idxs] # use_jit = False use_jit = True if use_jit: # X_sort = X_sort[:100] # TODO rm # X_sort = np.ascontiguousarray(X_sort) # N, D = X_sort.shape # print("about to call jitted func; N, D = ", N, D) sses_head = _cumsse_cols(X_sort) # print("got thru first call...") # X_sort_rev = np.ascontiguousarray(X_sort[::-1]) # sses_tail = _cumsse_cols(X_sort_rev)[::-1] sses_tail = _cumsse_cols(X_sort[::-1])[::-1] # print("returned from jitted func!") else: X_sort_sq = X_sort * X_sort # cumX_head = np.cumsum(X_sort, axis=0) # cumX2_head = np.cumsum(X_sort_sq, axis=0) # cumX_tail = np.cumsum(X_sort[::-1], axis=0)[::-1] # cumX2_tail = np.cumsum(X_sort_sq[::-1], axis=0)[::-1] cumX_head = _cumsum_cols(X_sort) cumX2_head = _cumsum_cols(X_sort_sq) cumX_tail = _cumsum_cols(X_sort[::-1])[::-1] cumX2_tail = _cumsum_cols(X_sort_sq[::-1])[::-1] all_counts = np.arange(1, N + 1).reshape(-1, 1) EX_head = cumX_head / all_counts # E[X], starting from 0 EX_tail = cumX_tail / all_counts[::-1] # E[X], starting from N-1 # EX2_head = cumX2_head / all_counts # E[X^2], starting from 0 # EX2_tail = cumX2_tail / all_counts[::-1] # E[X^2], starting from N-1 # mses_head = EX2_head - (EX_head * EX_head) # mses from 0 # mses_tail = EX2_tail - (EX_tail * EX_tail) # mses from N-1 # sses_head = mses_head * all_counts # # sses_tail = mses_tail * all_counts[::-1] # simpler equivalent of above; mse * N reduces to this sses_head = cumX2_head - (cumX_head * EX_head) sses_tail = cumX2_tail - (cumX_tail * EX_tail) # # TODO rm # mse_head_diffs = sses_head[1:] - sses_head[:-1] # # print("mse_head_diffs[:20]", mse_head_diffs[:20]) # assert np.all(mse_head_diffs > -.1) # should be nondecreasing # mse_tail_diffs = sses_tail[1:] - sses_tail[:-1] # assert np.all(mse_tail_diffs < .1) # should be nonincreasing sses = sses_head sses[:-1] += sses_tail[1:] # sse of X_sort[:i] + sse of X_sort[i:] sses = sses.sum(axis=1) if shrink_towards_median: minsse, maxsse = np.min(sses), np.max(sses) scale = maxsse - minsse # n_over_2 = N // 2 # scale = (maxsse - minsse) / n_over_2 coeffs = np.abs(np.arange(N, dtype=np.float32)) penalties = coeffs * (scale / np.max(coeffs)) sses += penalties # # TODO rm # E_X = X.mean(axis=0) # E_X2 = (X * X).mean(axis=0) # sse_true = np.sum(E_X2 - (E_X * E_X)) * N # print("sses[0], sses[-1], true loss, np.sum(X.var(axis=0)) * N", # sses[0], sses[-1], sse_true, np.sum(X.var(axis=0)) * N) # X_orig_sort = X_orig[sort_idxs] if possible_vals is None or not len(possible_vals): # can split anywhere best_idx = np.argmin(sses) next_idx = min(N - 1, best_idx + 1) # best_val = (X_sort[best_idx, dim] + X_sort[next_idx, dim]) / 2. # X_orig_sort = X_orig[sort_idxs] col = X_orig[:, dim] best_val = (col[sort_idxs[best_idx]] + col[sort_idxs[next_idx]]) / 2 # best_val = (X_orig_sort[best_idx, dim] + X_orig_sort[next_idx, dim]) / 2 else: # have to choose one of the values in possible_vals sorted_col = X_orig[:, dim][sort_idxs] idxs = np.searchsorted(sorted_col, possible_vals) # idxs = np.unique(idxs) idxs = np.maximum(0, idxs - 1) # searchsorted returns first idx larger sses_for_idxs = sses[idxs] which_idx_idx = np.argmin(sses_for_idxs) best_idx = idxs[which_idx_idx] best_val = possible_vals[which_idx_idx] # print("return_possible_vals_losses: ", return_possible_vals_losses) ret = best_val, sses[best_idx] return ret + (sses_for_idxs,) if return_possible_vals_losses else ret def evenly_spaced_quantiles(x, nquantiles, dedup=True): x = np.unique(x) # handle x with fewer unique elements than nquantiles (or same number, or # not that many more; basically just want each returned value to be uniq # and useful for binning the distribution) if len(x) == nquantiles: return x elif len(x) == 1: return np.linspace(-1, 3, num=nquantiles) * x[0] elif len(x) < 2 * nquantiles: return np.linspace(np.min(x), np.max(x), num=nquantiles) n = nquantiles + 1 fracs = np.arange(1, n) / float(n) return np.array([np.quantile(x, frac) for frac in fracs]) class PointInfo(object): __slots__ = 'data bucket_id'.split() def __init__(self, data, bucket_id): self.data = data self.bucket_id = bucket_id class Split(object): __slots__ = 'dim val loss_change'.split() def __init__(self, dim, val, loss_change=None): self.dim = dim self.val = val self.loss_change = loss_change def _sort_and_append_orig_idx(x, ascending=True): sort_idxs = np.argsort(x) if not ascending: sort_idxs = sort_idxs[::-1] x_sort = x[sort_idxs] orig_idxs = np.arange(len(x))[sort_idxs] return list(zip(x_sort, orig_idxs)) def _split_existing_buckets(buckets): return [buck.split() for buck in buckets] # new_buckets = [] # # D = len(buckets[0].sumX) # for buck in buckets: # # buck0 = copy.deepcopy(bucket) # # buck0 = Bucket(N=buck.N, D=D, point_ids=copy.deepcopy(buck.point_ids), # # sumX=np.copy(buck.sumX), sumX2=np.copy(buck.sumX2)) # # buck0 = buck.copy() # # buck1 = Bucket(D=buckets[0].D) # new_buckets.append((buck0, buck1)) # return new_buckets class MultiSplit(object): __slots__ = 'dim vals scaleby offset'.split() def __init__(self, dim, vals, scaleby=None, offset=None): self.dim = dim self.vals = np.asarray(vals) self.scaleby = scaleby self.offset = offset def preprocess_x(self, x): if self.offset is not None: x = x - self.offset if self.scaleby is not None: x = x * self.scaleby return x def learn_multisplits_orig(X, nsplits, log2_max_vals_per_split=4, try_nquantiles=16, return_centroids=True, # learn_quantize_params=False, learn_quantize_params='int16', # learn_quantize_params=True, # verbose=1): verbose=2): # verbose=3): X = X.astype(np.float32) N, D = X.shape max_vals_per_split = 1 << log2_max_vals_per_split X_hat = np.zeros_like(X) # initially, one big bucket with everything buckets = [Bucket(sumX=X.sum(axis=0), sumX2=(X * X).sum(axis=0), point_ids=np.arange(N))] total_loss = sum([bucket.loss for bucket in buckets]) # values to try in each dim, after buckets no longer get to pick optimal # ones; we try values that at evenly spaced quantiles possible_split_vals = np.empty((D, try_nquantiles), dtype=X.dtype) for dim in range(D): # possible_split_vals[dim] = evenly_spaced_quantiles( # X[:, dim], try_nquantiles) # exclude enpoints, so we get appropriate number of points linearly # spaced *between* min and max values minval, maxval = np.min(X[:, dim]), np.max(X[:, dim]) possible_split_vals[dim] = np.linspace( minval, maxval, num=(try_nquantiles + 2))[1:-1] # print("initial possible split vals: ") # print(possible_split_vals[:8]) # print(possible_split_vals[8:16]) # import sys; sys.exit() if verbose > 0: print("================================") print("learn_multisplits(): initial loss: ", total_loss) splits = [] col_losses = np.zeros(D, dtype=np.float32) # TODO rm? for s in range(nsplits): # if s >= 2: # print("exiting after two splits") # import sys; sys.exit() if verbose > 1: print("------------------------ finding split #:", s) for i, buck in enumerate(buckets): # TODO rm sanity check assert buck.id == i nbuckets = len(buckets) # compute number of bucket groups and size of each ngroups = min(nbuckets, max_vals_per_split) nbuckets_per_group = nbuckets // ngroups assert nbuckets_per_group * ngroups == nbuckets # sanity check # try_ndims = 8 # try_ndims = 4 try_ndims = 1 # dim_heuristic = 'eigenvec' # dim_heuristic = 'bucket_eigenvecs' dim_heuristic = 'variance' if dim_heuristic == 'eigenvec': # compute current reconstruction of X, along with errs for buck in buckets: # print("point ids: ", buck.point_ids) if len(buck.point_ids): centroid = buck.col_means() # X_hat[np.array(buck.point_ids)] = centroid X_hat[buck.point_ids] = centroid X_res = X - X_hat # pick dims by looking at top principal component v = subs.top_principal_component(X_res) try_dims = np.argsort(np.abs(v))[-try_ndims:] elif dim_heuristic == 'bucket_eigenvecs': dim_scores = np.zeros(D, dtype=np.float32) for buck in buckets: if buck.N < 2: continue X_buck = X[buck.point_ids] v, lamda = subs.top_principal_component( X_buck, return_eigenval=True) v *= lamda dim_scores += np.abs(v) # X_buck -= X_buck.mean(axis=0) try_dims = np.argsort(dim_scores)[-try_ndims:] elif dim_heuristic == 'variance': # pick out dims to consider splitting on # try_dims = np.arange(D) # TODO restrict to subset? col_losses[:] = 0 for buck in buckets: col_losses += buck.col_sum_sqs() # try_dims = np.argsort(col_losses)[-8:] try_dims = np.argsort(col_losses)[-try_ndims:] # try_dims = np.argsort(col_losses)[-2:] # try_dims = np.arange(2) # try_dims = np.arange(D) # TODO restrict to subset? losses = np.zeros(len(try_dims), dtype=X.dtype) losses_for_vals = np.zeros(try_nquantiles, dtype=X.dtype) all_split_vals = [] # vals chosen by each bucket/group for each dim # determine for this dim what the best split vals are for each # group and what the loss is when using these split vals for d, dim in enumerate(try_dims): if verbose > 2: # print("---------------------- dim = ", dim) print("======== dim = {}, ({:.5f}, {:.5f})".format( dim, np.min(X[:, dim]), np.max(X[:, dim]))) # just let each bucket pick its optimal split val for this dim; # special case of below where each "group" is one bucket, and # instead of having to pick val from fixed set, can be anything if nbuckets_per_group == 1: split_vals = [] # each bucket contributes one split val for buck in buckets: val, loss = buck.optimal_split_val(X, dim) losses[d] += loss split_vals.append(val) all_split_vals.append(split_vals) # buckets have to pick from fixed set of possible values; each # group of buckets (defined by common prefix) have to agree on # one val, so we sum loss for each possible value across all # buckets in the group, and then take val with lowest sum else: split_vals = [] # each group contributes one split val for g in range(ngroups): # print("------------------------ group #", g) start_idx = g * nbuckets_per_group end_idx = start_idx + nbuckets_per_group group_buckets = buckets[start_idx:end_idx] # print("bucket ids, counts: ", # [buck.id for buck in group_buckets], # [buck.N for buck in group_buckets]) # compute loss for each possible split value, summed # across all buckets in this group; then choose best possible_vals = possible_split_vals[dim] # print("possible split vals: ", possible_vals) losses_for_vals[:] = 0 # losses_for_vals = np.zeros_like(losses_for_vals) # print("losses for vals: ", losses_for_vals) for b, buck in enumerate(group_buckets): _, _, val_losses = buck.optimal_split_val( X, dim, possible_vals=possible_vals, return_possible_vals_losses=True) losses_for_vals += val_losses best_val_idx = np.argmin(losses_for_vals) best_val = possible_vals[best_val_idx] best_loss = losses_for_vals[best_val_idx] losses[d] += best_loss # print("best {val idx, val, loss} = ", # best_val_idx, best_val, best_loss) split_vals.append(best_val) all_split_vals.append(split_vals) # determine best dim to split on, and pull out associated split # vals for all buckets best_tried_dim_idx = np.argmin(losses) best_dim = try_dims[best_tried_dim_idx] use_split_vals = all_split_vals[best_tried_dim_idx] split = MultiSplit(dim=best_dim, vals=use_split_vals) if learn_quantize_params: # if len(use_split_vals) > 1: # after 1st split # minsplitval = np.min(use_split_vals) # maxsplitval = np.max(use_split_vals) # gap = maxsplitval - minsplitval # offset = minsplitval - .02 * gap # scale = 250. / gap # slightly below 255. / gap # else: # 1st split; only one bucket, so no intersplit range # assert np.min(use_split_vals) == np.max(use_split_vals) # x = X[:, best_dim] # offset = np.min(x) # scale = 255. / np.max(x - offset) # # x -= offset # # scale = 128. / np.max(split.vals - offset) # # scale = 1 # TODO rm # # x = X[:, best_dim].copy() # x = X[:, best_dim] # offset = np.min(x) # # scale = 255. / np.max(x - offset) # scale = 250. / np.max(use_split_vals) # slightly below 255 # simple version, which also handles 1 bucket: just set min # value to be avg of min splitval and xval, and max value to # be avg of max splitval and xval x = X[:, best_dim] offset = (np.min(x) + np.min(use_split_vals)) / 2 upper_val = (np.max(x) + np.max(use_split_vals)) / 2 - offset scale = 254. / upper_val if learn_quantize_params == 'int16': scale = 2. ** int(np.log2(scale)) split.offset = offset split.scaleby = scale split.vals = (split.vals - split.offset) * split.scaleby split.vals = np.clip(split.vals, 0, 255).astype(np.int32) splits.append(split) # apply this split to get next round of buckets new_buckets = [] for i, buck in enumerate(buckets): group_idx = i // nbuckets_per_group val = use_split_vals[group_idx] new_buckets += list(buck.split(X, dim=best_dim, val=val)) buckets = new_buckets if verbose > 1: print("bucket counts: ", [buck.N for buck in buckets]) print("loss from buckets: ", sum([bucket.loss for bucket in buckets])) print("dim losses: ", losses) if verbose > 2: print("loss from sse computation: ", losses[best_tried_dim_idx]) print("using dim, split_vals:", best_dim, use_split_vals) # maybe return centroids in addition to set of MultiSplits and loss loss = sum([bucket.loss for bucket in buckets]) if verbose > 0: print("learn_multisplits(): returning loss: ", loss) if return_centroids: centroids = np.vstack([buck.col_means() for buck in buckets]) assert centroids.shape == (len(buckets), X.shape[1]) return splits, loss, centroids return splits, loss @_memory.cache def learn_multisplits( X, nsplits=4, return_centroids=True, return_buckets=False, # learn_quantize_params=False, # learn_quantize_params='int16', X_orig=None, try_ndims=1, # learn_quantize_params='int16', X_orig=None, try_ndims=2, learn_quantize_params='int16', X_orig=None, try_ndims=4, # learn_quantize_params='int16', X_orig=None, try_ndims=8, # learn_quantize_params='int16', X_orig=None, try_ndims=16, # learn_quantize_params=True, # verbose=3): # verbose=2): verbose=1): assert nsplits <= 4 # >4 splits means >16 split_vals for this func's impl X = X.astype(np.float32) N, D = X.shape X_orig = X if X_orig is None else X_orig X_hat = np.zeros_like(X) # initially, one big bucket with everything buckets = [Bucket(sumX=X.sum(axis=0), sumX2=(X * X).sum(axis=0), point_ids=np.arange(N))] total_loss = sum([bucket.loss for bucket in buckets]) if verbose > 0: print("================================") # print("learn_multisplits(): initial loss: ", total_loss) print("learn_multisplits(): initial loss: ", total_loss) # print("learn_multisplits(): trying ndims: ", min(D, try_ndims)) splits = [] col_losses = np.zeros(D, dtype=np.float32) # TODO rm? for s in range(nsplits): if verbose > 1: print("------------------------ finding split #:", s) # dim_heuristic = 'eigenvec' # dim_heuristic = 'bucket_eigenvecs' dim_heuristic = 'bucket_sse' # dim_heuristic = 'kurtosis' if dim_heuristic == 'eigenvec': # compute current reconstruction of X, along with errs if s > 0: for buck in buckets: # print("point ids: ", buck.point_ids) if len(buck.point_ids): centroid = buck.col_means() # X_hat[np.array(buck.point_ids)] = centroid X_hat[buck.point_ids] = centroid X_res = X - X_hat else: X_res = X # pick dims by looking at top principal component v = subs.top_principal_component(X_res) try_dims = np.argsort(np.abs(v))[-try_ndims:] elif dim_heuristic == 'bucket_eigenvecs': dim_scores = np.zeros(D, dtype=np.float32) for buck in buckets: if buck.N < 2: continue X_buck = X[buck.point_ids] v, lamda = subs.top_principal_component( X_buck, return_eigenval=True) v *= lamda dim_scores += np.abs(v) # X_buck -= X_buck.mean(axis=0) try_dims = np.argsort(dim_scores)[-try_ndims:] elif dim_heuristic == 'bucket_sse': col_losses[:] = 0 for buck in buckets: col_losses += buck.col_sum_sqs() try_dims = np.argsort(col_losses)[-try_ndims:] elif dim_heuristic == 'kurtosis': # compute X_res if s > 0: for buck in buckets: # print("point ids: ", buck.point_ids) if len(buck.point_ids): centroid = buck.col_means() # X_hat[np.array(buck.point_ids)] = centroid X_hat[buck.point_ids] = centroid X_res = X - X_hat else: X_res = X col_losses[:] = 0 for buck in buckets: col_losses += buck.col_sum_sqs() try_dims = np.argsort(col_losses)[-try_ndims:] from scipy import stats col_losses *= col_losses # just 4th central moment col_losses *= stats.kurtosis(X_res, axis=0) try_dims = np.argsort(col_losses)[-try_ndims:] losses = np.zeros(len(try_dims), dtype=X.dtype) all_split_vals = [] # vals chosen by each bucket/group for each dim # determine for this dim what the best split vals are for each # group and what the loss is when using these split vals # print("try_dims: ", try_dims) for d, dim in enumerate(try_dims): # print("s, d, dim = ", s, d, dim) if verbose > 2: # print("---------------------- dim = ", dim) print("======== dim = {}, ({:.5f}, {:.5f})".format( dim, np.min(X[:, dim]), np.max(X[:, dim]))) split_vals = [] # each bucket contributes one split val for b, buck in enumerate(buckets): val, loss = buck.optimal_split_val(X, dim, X_orig=X_orig) losses[d] += loss if d > 0 and losses[d] >= np.min(losses[:d]): if verbose > 2: print("early abandoning after bucket {}!".format(b)) break # this dim already can't be the best split_vals.append(val) all_split_vals.append(split_vals) # determine best dim to split on, and pull out associated split # vals for all buckets best_tried_dim_idx = np.argmin(losses) best_dim = try_dims[best_tried_dim_idx] use_split_vals = all_split_vals[best_tried_dim_idx] split = MultiSplit(dim=best_dim, vals=use_split_vals) if learn_quantize_params: # simple version, which also handles 1 bucket: just set min # value to be avg of min splitval and xval, and max value to # be avg of max splitval and xval x = X[:, best_dim] offset = (np.min(x) + np.min(use_split_vals)) / 2 upper_val = (np.max(x) + np.max(use_split_vals)) / 2 - offset scale = 254. / upper_val if learn_quantize_params == 'int16': scale = 2. ** int(np.log2(scale)) split.offset = offset split.scaleby = scale split.vals = (split.vals - split.offset) * split.scaleby split.vals = np.clip(split.vals, 0, 255).astype(np.int32) splits.append(split) # apply this split to get next round of buckets new_buckets = [] for i, buck in enumerate(buckets): group_idx = i val = use_split_vals[group_idx] new_buckets += list(buck.split(X, dim=best_dim, val=val, X_orig=X_orig)) buckets = new_buckets if verbose > 1: print("bucket counts: ", [buck.N for buck in buckets]) # print("loss from buckets: ", # sum([bucket.loss for bucket in buckets])) print("dim losses: ", losses) if verbose > 2: print("loss from sse computation: ", losses[best_tried_dim_idx]) print("using dim, split_vals:", best_dim, use_split_vals) # maybe return centroids in addition to set of MultiSplits and loss loss = sum([bucket.loss for bucket in buckets]) if verbose > 0: print("learn_multisplits(): returning loss: ", loss) ret = [splits, loss] if return_centroids: centroids = np.vstack([buck.col_means() for buck in buckets]) assert centroids.shape == (len(buckets), X.shape[1]) ret.append(centroids) # return splits, loss, centroids if return_buckets: # print("returning buckets!") ret.append(buckets) return tuple(ret) @numba.njit(fastmath=True, cache=True) def _XtX_encoded(X_enc, K=16): N, C = X_enc.shape D = C * K # note that this is total number of centroids, not orig D out = np.zeros((D, D), np.int32) # out = np.zeros((D, D), np.float32) # D = int(C * K) # note that this is total number of centroids, not orig D # out = np.zeros((D, D), np.int8) for n in range(N): for c in range(C): code_left = X_enc[n, c] dim_left = (K * c) + code_left out[dim_left, dim_left] += 1 for cc in range(c + 1, C): code_right = X_enc[n, cc] dim_right = (K * cc) + code_right out[dim_left, dim_right] += 1 # populate lower triangle for d in range(D): for dd in range(d + 1, D): out[dd, d] = out[d, dd] return out @numba.njit(fastmath=True, cache=True) def _XtY_encoded(X_enc, Y, K=16): N, C = X_enc.shape N, M = Y.shape D = int(C * K) # note that this is total number of centroids, not orig D out = np.zeros((D, M), Y.dtype) for n in range(N): for c in range(C): code_left = X_enc[n, c] dim_left = (K * c) + code_left for m in range(M): out[dim_left, m] += Y[n, m] return out @numba.njit(fastmath=True, cache=True) def _XW_encoded(X_enc, W, K=16): N, C = X_enc.shape D, M = W.shape out = np.zeros((N, M), W.dtype) for n in range(N): for c in range(C): code_left = X_enc[n, c] dim_left = (K * c) + code_left for m in range(M): out[n, m] += W[dim_left, m] return out @numba.njit(fastmath=True, cache=True) def _densify_X_enc(X_enc, K=16): N, C = X_enc.shape D = C * K out = np.zeros((N, D), np.int8) for n in range(N): for c in range(C): code_left = X_enc[n, c] dim_left = (K * c) + code_left out[n, dim_left] = 1 return out def _fit_ridge_enc(X_enc=None, Y=None, K=16, lamda=1, X_bin=None): if X_bin is None: X_bin = _densify_X_enc(X_enc, K=K) est = linear_model.Ridge(fit_intercept=False, alpha=lamda) est.fit(X_bin, Y) return est.coef_.T def encoded_lstsq(X_enc=None, X_bin=None, Y=None, K=16, XtX=None, XtY=None, precondition=True, stable_ridge=True): if stable_ridge: return _fit_ridge_enc(X_enc=X_enc, Y=Y, X_bin=X_bin, K=K, lamda=1) if XtX is None: XtX = _XtX_encoded(X_enc, K=K).astype(np.float32) lamda = 1 # TODO cross-validate to get lamda # N = X_enc.shape[0] # # lamda = N / (K * K) # Y_bar = Y - Y.mean(axis=0) # lamda = N * np.var(Y - Y.mean(axis=0)) / (K * K) # # lamda = N * np.var(Y - Y.mean(axis=0)) / K # lamda = N * np.var(Y) / K # lamda = N * np.var(Y) / (K * K) # # lamda = N * 1e4 # should shrink coeffs to almost 0 # # alpha = unscaled_alpha * np.var(X - X.mean(axis=0)) * N / D # lamda = N / (1e5) # sorta works # lamda = N / (1e4) # sorta works lamda = max(1, lamda) print("using lamda = ", lamda) # lamda = max(1, len(X_enc) / 1e6) # lamda = max(1, len(X_enc) / 1e5) # lamda = max(1, len(X_enc) / 1e4) # lamda = max(1, len(X_enc) / float(K * K)) # lamda = len(X_enc) / float(K) # print("computing and regularizing XtX using lambda = ", lamda) XtX += np.diag(np.ones(XtX.shape[0]) * lamda).astype(np.float32) # ridge if XtY is None: XtY = _XtY_encoded(X_enc, Y, K=K) XtX = XtX.astype(np.float64) XtY = XtY.astype(np.float64) # preconditioning to avoid numerical issues (seemingly unnecessary, but # might as well do it) # scale = 1. / np.std(XtX) if precondition: # # pretend cols of X were scaled differently # xscales = np.linalg.norm(XtX, axis=0) + 1e-20 # mulby = (1. / xscales) # XtX *= mulby * mulby # XtY *= mulby.reshape(-1, 1) # yscales = np.linalg.norm(XtY, axis=1) + 1e-20 # yscales = np.linalg.norm(XtY, axis=0) + 1e-20 # yscales = yscales.reshape(-1, 1) # xscales = np.mean(np.linalg.norm(XtX, axis=0)) # xscales = 7 # xscales = 1 # XtY *= (1. / yscales) # XtY *= (1. / yscales.reshape(-1, 1)) # scale = 1. / len(X_enc) scale = 1. / np.linalg.norm(XtX, axis=0).max() XtX = XtX * scale XtY = XtY * scale # W = np.linalg.solve(XtX, XtY) W, _, _, _ = np.linalg.lstsq(XtX, XtY, rcond=None) # doesn't fix it # W, _, _, _ = np.linalg.lstsq(X_bin, Y, rcond=None) # import torch # import torch.nn.functional as F # import torch.optim as optim # def _to_np(A): # return A.cpu().detach().numpy() # niters = 10 # for it in range(niters): # if precondition: # pass # # W *= xscales # # W *= xscales.reshape(-1, 1) # # W /= xscales.reshape(-1, 1) # # W *= yscales.ravel() # # W *= yscales # W *= yscales # undo preconditioning # import matplotlib.pyplot as plt # _, axes = plt.subplots(2, 2, figsize=(13, 10)) # axes[0, 0].imshow(_densify_X_enc(X_enc[:1000]), interpolation='nearest') # axes[0, 1].imshow(XtX, interpolation='nearest') # axes[1, 0].imshow(XtY, interpolation='nearest', cmap='RdBu') # axes[1, 1].imshow(W, interpolation='nearest', cmap='RdBu') # # plt.colorbar() # plt.tight_layout() # plt.show() # import sys; sys.exit() return W def _sparse_encoded_lstsq_gomp(X_enc, Y, nnz_blocks, K=16): assert nnz_blocks >= 1 ncodebooks = X_enc.shape[1] M = Y.shape[1] # precompute XtX and XtY and create initial dense W XtX = _XtX_encoded(X_enc, K=K).astype(np.float32) XtX += np.diag(np.ones(XtX.shape[0])).astype(np.float32) # ridge XtY = _XtY_encoded(X_enc, Y, K=K) W = encoded_lstsq(X_enc, Y, XtX=XtX, XtY=XtY) XtX = np.asfarray(XtX) # since we'll be slicing columns keep_codebook_idxs = np.empty((M, nnz_blocks), dtype=np.int) XtX_G = np.empty((ncodebooks, K * ncodebooks, K), dtype=np.float32) for c in range(ncodebooks): start_idx = c * K end_idx = start_idx + K # use_XtX = XtX[start_idx:end_idx][:, start_idx:end_idx] use_XtX = XtX[:, start_idx:end_idx] XtX_G[c], _ = np.linalg.qr(use_XtX) # KC x K codebook_scores = np.zeros(ncodebooks) for m in range(M): # fully solve one output col at a time # xty = np.ascontiguousarray(XtY[:, m]) targets = np.copy(XtY[:, m]) residuals = targets keep_codebooks = set() w = np.copy(W[:, m]) pq_codebook_idx = int(m / float(M) * ncodebooks) # print("---- m = ", m) for b in range(nnz_blocks): # targets_normed = targets # score each codebook to pick new one to add if b > 0: for c in range(ncodebooks): if c in keep_codebooks: codebook_scores[c] = -np.inf continue X_G = XtX_G[c] codebook_scores[c] = np.linalg.norm(X_G.T @ residuals) keep_codebooks.add(np.argmax(codebook_scores)) else: keep_codebooks.add(pq_codebook_idx) # seed with pq idx # refit model using all the groups selected so far keep_idxs = [np.arange(i * K, (i + 1) * K) for i in sorted(list(keep_codebooks))] keep_idxs = np.hstack(keep_idxs) XtX_subs = XtX[keep_idxs][:, keep_idxs] targets_subs = targets[keep_idxs] w_subs = np.linalg.solve(XtX_subs, targets_subs) # XtX_subs = XtX[:, keep_idxs] # targets_subs = targets[keep_idxs] # w_subs = np.linalg.solve(XtX_subs, targets) # w_subs, resid, _, _ = np.linalg.lstsq(XtX_subs, targets) w[:] = 0 w[keep_idxs] = w_subs # resid_norm_sq = np.linalg.norm(residuals)**2 # print("resid norm sq: ", resid_norm_sq) # print("lstsq mse: ", resid / resid_norm_sq) # residuals = targets - (XtX_subs @ w_subs) residuals = targets - (XtX[:, keep_idxs] @ w_subs) # resid_norm_sq = np.linalg.norm(residuals)**2 # print("new resid norm sq: ", resid_norm_sq) # targets = np.copy(XtY[:, m]) - (XtX @ w) # update return arrays keep_codebook_idxs[m] = np.array(list(keep_codebooks)) W[:, m] = w return W, keep_codebook_idxs # each codebook has const number of nonzero idxs def _sparse_encoded_lstsq_elim_v2(X_enc, Y, nnz_per_centroid, K=16, # uniform_sparsity=False): # never better uniform_sparsity=True, pq_perm_algo='start', stable_ridge=True): ncodebooks = X_enc.shape[1] M = Y.shape[1] nnz_per_centroid = min(M, int(nnz_per_centroid)) nnz_per_centroid = max(1, nnz_per_centroid) assert nnz_per_centroid >= int(np.ceil(M / ncodebooks)) assert nnz_per_centroid <= M X_bin = _densify_X_enc(X_enc, K=K) if not stable_ridge: # precompute XtX and XtY and create initial dense W XtX = _XtX_encoded(X_enc, K=K).astype(np.float32) lamda = 1 # # alpha = unscaled_alpha * np.var(X - X.mean(axis=0)) * N / D # # lamda = np.sqrt(ncodebooks) # N = XtX.shape[0] # lamda = N / (K * K) # lamda = max(1, lamda) # print("using lamda = ", lamda) # lamda = max(1, len(X_enc) / 1e4) # lamda = max(1, len(X_enc) / float(K * K)) XtX += np.diag(np.ones(XtX.shape[0]) * lamda).astype(np.float32) # ridge # XtX += np.diag(np.ones(XtX.shape[0])).astype(np.float32) # ridge XtY = _XtY_encoded(X_enc, Y, K=K) # scale = 1. / len(X_enc) scale = 1. / np.linalg.norm(XtX, axis=0).max() XtX = XtX * scale XtY = XtY * scale W = encoded_lstsq(X_bin=X_bin, Y=Y, XtX=XtX, XtY=XtY, precondition=False, stable_ridge=stable_ridge) # KC x M XtX = np.asfarray(XtX) # since we'll be slicing columns else: # stable_ridge is True W = encoded_lstsq(X_bin=X_bin, Y=Y, stable_ridge=stable_ridge) # score all blocks of W all_scores = np.empty((ncodebooks, M), dtype=np.float) # C x M for c in range(ncodebooks): Xc = X_enc[:, c].reshape(-1, 1) start_idx = c * K end_idx = start_idx + K Wc = W[start_idx:end_idx] Yc = _XtY_encoded(Xc, Wc, K=K) # N x M all_scores[c] = np.linalg.norm(Yc, axis=0) # pq_idxs = _pq_codebook_start_end_idxs(M, ncodebooks) pq_idxs = _pq_codebook_start_end_idxs(Y, ncodebooks, algo=pq_perm_algo) # now pick which cols to keep in each codebook keep_mask = np.zeros((ncodebooks, M), dtype=np.bool) # subvec_len = int(np.ceil(M / ncodebooks)) for c in range(ncodebooks): # initialize with PQ start_idx, end_idx = pq_idxs[c] keep_mask[c, start_idx:end_idx] = 1 subvec_len = end_idx - start_idx assert subvec_len >= 1 keep_nidxs_extra = nnz_per_centroid - subvec_len scores = all_scores[c] scores[start_idx:end_idx] = 0 if uniform_sparsity and keep_nidxs_extra > 0: # take as many other (best) nonzero idxs as we we're allowed to assert len(scores) >= keep_nidxs_extra best_idxs = np.argsort(scores)[-keep_nidxs_extra:] if len(best_idxs) != keep_nidxs_extra: print("len(best_idxs)", len(best_idxs)) print("keep_nidxs_extra", keep_nidxs_extra) assert len(best_idxs) == keep_nidxs_extra keep_mask[c, best_idxs] = True if not uniform_sparsity: scores = all_scores.ravel() nkept_idxs = M # number of nonzeros used already keep_nidxs_total = nnz_per_centroid * ncodebooks keep_nidxs_extra = keep_nidxs_total - nkept_idxs keep_idxs = np.argsort(scores)[-keep_nidxs_extra:] flat_mask = keep_mask.ravel() flat_mask[keep_idxs] = 1 keep_mask = flat_mask.reshape(keep_mask.shape) # at this point, we have the mask for which cols of each centroid to keep; # now we just need to go from a mask to a set of indices and a sparse # matrix of centroids W_sparse = np.empty((ncodebooks * K, M), dtype=np.float32) if uniform_sparsity: ret_idxs = np.empty((ncodebooks, nnz_per_centroid), dtype=np.int) else: ret_idxs = [] # else: # ret_idxs = np.zeros((ncodebooks, M), dtype=np.int) - 1 for c in range(ncodebooks): idxs = np.where(keep_mask[c] != 0)[0] if uniform_sparsity: if len(idxs) != nnz_per_centroid: print("c: ", c) print("len(idxs): ", len(idxs)) print("nnz_per_centroid: ", nnz_per_centroid) print("keep_mask counts:", keep_mask.sum(axis=1)) assert len(idxs) == nnz_per_centroid ret_idxs[c] = idxs else: ret_idxs.append(idxs) zero_idxs = np.where(keep_mask[c] == 0)[0] start_idx = c * K end_idx = start_idx + K Wc = W[start_idx:end_idx] Wc[:, zero_idxs] = 0 W_sparse[start_idx:end_idx] = Wc # now refit W_sparse to each output col; right now it's just the original # W with a bunch of entries zeroed for m in range(M): w = W_sparse[:, m] keep_idxs = np.where(w != 0)[0] if stable_ridge: X_bin_subs = X_bin[:, keep_idxs] w_subs = _fit_ridge_enc(X_bin=X_bin_subs, Y=Y[:, m]) else: xty = XtY[:, m] use_XtX = XtX[keep_idxs][:, keep_idxs] use_xty = xty[keep_idxs] w_subs = np.linalg.solve(use_XtX, use_xty) w[:] = 0 w[keep_idxs] = w_subs W_sparse[:, m] = w # nnzs = [len(idxs) for idxs in ret_idxs] # print("nnzs: ", nnzs) # print(f"returning {ret_idxs.shape[1]} nonzeros per centroid...") return W_sparse, ret_idxs def _sparse_encoded_lstsq_backward_elim(X_enc, Y, nnz_blocks, K=16): ncodebooks = X_enc.shape[1] eliminate_nblocks = ncodebooks - nnz_blocks M = Y.shape[1] # precompute XtX and XtY and create initial dense W XtX = _XtX_encoded(X_enc, K=K).astype(np.float32) XtX += np.diag(np.ones(XtX.shape[0])).astype(np.float32) # ridge XtY = _XtY_encoded(X_enc, Y, K=K) W = encoded_lstsq(X_enc, Y, XtX=XtX, XtY=XtY) XtX = np.asfarray(XtX) # since we'll be slicing columns keep_codebook_idxs = np.empty((M, nnz_blocks), dtype=np.int) codebook_scores = np.zeros(ncodebooks) for m in range(M): # fully solve one output col at a time xty = np.ascontiguousarray(XtY[:, m]) rm_codebook_idxs = set() w = np.copy(W[:, m]) for b in range(eliminate_nblocks): # evaluate contribution of each codebook for c in range(ncodebooks): # if c in rm_codebook_idxs or c == pq_codebook_idx: if c in rm_codebook_idxs: codebook_scores[c] = np.inf continue start_idx = c * K end_idx = start_idx + K # XtX_subs = XtX[:, start_idx:end_idx] # CK x K # w_subs = w[start_idx:end_idx] # K # xtyhat_subs = XtX_subs @ w_subs # CK x 1 # codebook_scores[c] = np.linalg.norm(xtyhat_subs) XtX_subs = XtX[start_idx:end_idx][:, start_idx:end_idx] w_subs = w[start_idx:end_idx] # K xtyhat_subs = XtX_subs @ w_subs # K x 1 codebook_scores[c] = np.linalg.norm(xtyhat_subs) # rm least helpful codebook and refit the least squares rm_codebook_idxs.add(np.argmin(codebook_scores)) keep_codebooks = [i for i in range(ncodebooks) if i not in rm_codebook_idxs] keep_idxs = [np.arange(i * K, (i + 1) * K) for i in keep_codebooks] keep_idxs = np.hstack(keep_idxs) use_XtX = XtX[keep_idxs][:, keep_idxs] use_xty = xty[keep_idxs] w_subs = np.linalg.solve(use_XtX, use_xty) # print("w shape: ", w.shape) # print("rm codebooks: ", rm_codebook_idxs) # print("keep codebooks: ", keep_codebooks) # print("keep idxs: ", keep_idxs) # print("type(keep idxs): ", type(keep_idxs)) # print("w[keep idxs]: ", w[keep_idxs]) # print("resid: ", resid) w[:] = 0 w[keep_idxs] = w_subs # update return arrays keep_idxs = [i for i in range(ncodebooks) if i not in rm_codebook_idxs] keep_codebook_idxs[m] = np.array(keep_codebooks) W[:, m] = w return W, keep_codebook_idxs # CK x M, M x nnz def sparse_encoded_lstsq(X_enc, Y, K=16, nnz_blocks=-1, **kwargs): ncodebooks = X_enc.shape[1] if nnz_blocks < 1: # nnz_per_centroid = Y.shape[1] # default to returning dense centroids W = encoded_lstsq(X_enc, Y, K=16) ncodebooks = X_enc.shape[1] M = Y.shape[1] keep_codebook_idxs = np.empty((ncodebooks, M), dtype=np.int) all_idxs = np.arange(M) for c in range(ncodebooks): keep_codebook_idxs[c] = all_idxs return W, keep_codebook_idxs else: nnz_per_centroid = int(nnz_blocks * Y.shape[1] / ncodebooks) # nnz_blocks = int(np.sqrt(ncodebooks) + .5) # return _sparse_encoded_lstsq_backward_elim( # X_enc, Y, nnz_blocks=nnz_blocks, K=K) # return _sparse_encoded_lstsq_gomp(X_enc, Y, nnz_blocks=nnz_blocks, K=K) # print("nnz_per_centroid: ", nnz_per_centroid) return _sparse_encoded_lstsq_elim_v2( X_enc, Y, nnz_per_centroid=nnz_per_centroid, K=K, **kwargs) # def _pq_codebook_start_end_idxs(D, ncodebooks): def _pq_codebook_start_end_idxs(X, ncodebooks, algo='start'): assert algo in ('start', 'end') # TODO do something smarter here # D = int(D) _, D = X.shape ncodebooks = int(ncodebooks) assert D >= ncodebooks idxs = np.empty((ncodebooks, 2), dtype=np.int) full_subvec_len = D // ncodebooks start_idx = 0 for c in range(ncodebooks): subvec_len = full_subvec_len if algo == 'start': # wider codebooks at the start if c < (D % ncodebooks): subvec_len += 1 elif algo == 'end': # wider codebooks at the end if (ncodebooks - c - 1) < (D % ncodebooks): subvec_len += 1 end_idx = min(D, start_idx + subvec_len) # print("c, start_idx, end_idx: ", c, start_idx, end_idx) # print("start_idx, end_idx: ", c, start_idx, end_idx) idxs[c, 0] = start_idx idxs[c, 1] = end_idx start_idx = end_idx assert idxs[0, 0] == 0 assert idxs[-1, -1] == D return idxs @_memory.cache def _learn_mithral_initialization(X, ncodebooks, pq_perm_algo='start', **kwargs): N, D = X.shape ncentroids_per_codebook = 16 X = X.astype(np.float32) X_res = X.copy() X_orig = X all_centroids = np.zeros( (ncodebooks, ncentroids_per_codebook, D), dtype=np.float32) all_splits = [] pq_idxs = _pq_codebook_start_end_idxs(X, ncodebooks, algo=pq_perm_algo) subvec_len = int(np.ceil(D / ncodebooks)) # for non-pq heuristics nonzeros_heuristic = 'pq' # ------------------------ 0th iteration; initialize all codebooks all_splits = [] all_buckets = [] for c in range(ncodebooks): if nonzeros_heuristic == 'pq': start_idx, end_idx = pq_idxs[c] idxs = np.arange(start_idx, end_idx) elif nonzeros_heuristic == 'pca': v = subs.top_principal_component(X_res) idxs = np.argsort(np.abs(v))[:-subvec_len] elif nonzeros_heuristic == 'disjoint_pca': use_X_res = X_res.copy() if c > 0: # not the first codebook use_X_res[:, idxs] = 0 # can't use same subspace v = subs.top_principal_component(use_X_res) idxs = np.argsort(np.abs(v))[:-subvec_len] use_X_res = X_res[:, idxs] use_X_orig = X_orig[:, idxs] # learn codebook to soak current residuals multisplits, _, buckets = learn_multisplits( use_X_res, X_orig=use_X_orig, return_centroids=False, return_buckets=True, **kwargs) for split in multisplits: split.dim = idxs[split.dim] all_splits.append(multisplits) all_buckets.append(buckets) # update residuals and store centroids centroid = np.zeros(D, dtype=np.float32) for b, buck in enumerate(buckets): if len(buck.point_ids): centroid[:] = 0 centroid[idxs] = buck.col_means() X_res[buck.point_ids] -= centroid # update centroid here in case we want to regularize it somehow all_centroids[c, b] = centroid # print("X_res mse / X mse: ", # (X_res * X_res).mean() / (X_orig * X_orig).mean()) return X_res, all_splits, all_centroids, all_buckets @_memory.cache def learn_mithral(X, ncodebooks, return_buckets=False, lut_work_const=-1, **kwargs): N, D = X.shape ncentroids_per_codebook = 16 X_orig = X.astype(np.float32) X_res0, all_splits0, all_centroids0, all_buckets0 = \ _learn_mithral_initialization(X, ncodebooks, pq_perm_algo='start') mse_orig = (X_orig * X_orig).mean() mse0 = (X_res0 * X_res0).mean() print("X_res mse / X mse: ", mse0 / mse_orig) used_perm_algo = 'start' if False: # choose between having wider codebooks at the start vs the end (if # there might be a meaningful difference) X_res1, all_splits1, all_centroids1, all_buckets1 = \ _learn_mithral_initialization(X, ncodebooks, pq_perm_algo='end') mse1 = (X_res1 * X_res1).mean() if mse0 <= mse1: X_res, all_splits, all_centroids, all_buckets = ( X_res0, all_splits0, all_centroids0, all_buckets0) else: X_res, all_splits, all_centroids, all_buckets = ( X_res1, all_splits1, all_centroids1, all_buckets1) used_perm_algo = 'end' print("X_res1 mse / X mse: ", mse1 / mse_orig) else: X_res, all_splits, all_centroids, all_buckets = ( X_res0, all_splits0, all_centroids0, all_buckets0) # optimize centroids discriminatively conditioned on assignments X_enc = mithral_encode(X, all_splits) if lut_work_const != 1: # if it's 1, equivalent to just doing PQ # # shrink W towards 0 # # if lut_work_const < 0: # W = encoded_lstsq(X_enc, X) # else: # W, nonzero_blocks = sparse_encoded_lstsq( # X_enc, X, nnz_blocks=lut_work_const) # # shrink W towards initial centroids # if lut_work_const < 0: print("fitting dense lstsq to X_res") W = encoded_lstsq(X_enc=X_enc, Y=X_res) else: W, _ = sparse_encoded_lstsq( X_enc, X_res, nnz_blocks=lut_work_const, pq_perm_algo=used_perm_algo) all_centroids_delta = W.reshape(ncodebooks, ncentroids_per_codebook, D) all_centroids += all_centroids_delta # check how much improvement we got X_res -= _XW_encoded(X_enc, W) # if we fit to X_res mse_res = (X_res * X_res).mean() print("X_res mse / X mse after lstsq: ", mse_res / mse_orig) # print("min, median, max, std, of all centroids after lstsq:\n", # all_centroids.min(), np.median(all_centroids), # all_centroids.max(), all_centroids.std()) if return_buckets: return all_splits, all_centroids, all_buckets return all_splits, all_centroids def learn_mithral_v1(X, ncodebooks, niters=1, return_buckets=False, **kwargs): # print("called learn_mithral!"); import sys; sys.exit() N, D = X.shape ncentroids_per_codebook = 16 X = X.astype(np.float32) X_res = X.copy() X_orig = X X_hat = np.zeros_like(X) all_centroids = np.zeros( (ncodebooks, ncentroids_per_codebook, D), dtype=np.float32) all_splits = [] subvec_len = int(np.ceil(D / ncodebooks)) # use_X_res = np.zeros_like(X_res) # TODO multiple iters; also store assignments from each codebook, so # that we can undo effect of its X_hat (can't store X_hat directly for # large data, so recompute on the fly using assignments and centroids) nonzeros_heuristic = 'pq' # nonzeros_heuristic = 'pca' # nonzeros_heuristic = 'disjoint_pca' # TODO store assignments (or maybe just buckets directly) # TODO update just centroids (not assignments) at iter end # ------------------------ 0th iteration; initialize all codebooks all_splits = [] all_buckets = [] for c in range(ncodebooks): if nonzeros_heuristic == 'pq': start_idx = c * subvec_len end_idx = min(D, start_idx + subvec_len) idxs = np.arange(start_idx, end_idx) elif nonzeros_heuristic == 'pca': v = subs.top_principal_component(X_res) idxs = np.argsort(np.abs(v))[:-subvec_len] elif nonzeros_heuristic == 'disjoint_pca': use_X_res = X_res.copy() if c > 0: # not the first codebook use_X_res[:, idxs] = 0 # can't use same subspace v = subs.top_principal_component(use_X_res) idxs = np.argsort(np.abs(v))[:-subvec_len] use_X_res = X_res[:, idxs] use_X_orig = X_orig[:, idxs] # learn codebook to soak current residuals multisplits, _, buckets = learn_multisplits( use_X_res, X_orig=use_X_orig, return_centroids=False, return_buckets=True, **kwargs) for split in multisplits: split.dim = idxs[split.dim] all_splits.append(multisplits) all_buckets.append(buckets) # use_X_res[:, start_idx:end_idx] = 0 # use_X_res[:] = 0 # update residuals and store centroids centroid = np.zeros(D, dtype=np.float32) for b, buck in enumerate(buckets): if len(buck.point_ids): centroid[:] = 0 centroid[idxs] = buck.col_means() # centroid /= 2 # TODO rm X_hat[buck.point_ids] = centroid # update centroid here in case we want to regularize it somehow all_centroids[c, b] = centroid X_res -= X_hat print("X res var / X var: ", X_res.var() / X_orig.var()) # ------------------------ remaining iters for t in range(niters): # now update centroids given assignments and all other centroids # for _ in range(5): # for _ in range(20): for _ in range(10): for c in range(ncodebooks): # print("c: ", c) # undo effect of this codebook buckets = all_buckets[c] for b, buck in enumerate(buckets): if len(buck.point_ids): X_hat[buck.point_ids] = all_centroids[c, b] X_res += X_hat # update centroids based on residuals given all other codebooks for b, buck in enumerate(buckets): if len(buck.point_ids): centroid = X_res[buck.point_ids].mean(axis=0) # keep_ndims = D // 2 # zero_idxs = np.argsort(np.abs(centroid))[:-keep_ndims] # centroid[zero_idxs] = 0 # true_centroid = X_res[buck.point_ids].mean(axis=0) # old_centroid = all_centroids[c, b] # centroid = (true_centroid + old_centroid) / 2 X_hat[buck.point_ids] = centroid all_centroids[c, b] = centroid X_res -= X_hat print("X res var / X var after centroid updates: ", X_res.var() / X_orig.var()) # now update assignments if t == niters - 1: break # end after updating centroids, not assignments for c in range(ncodebooks): # print("c: ", c) # undo effect of this codebook buckets = all_buckets[c] # orig_loss = sum([buck.loss for buck in buckets]) orig_loss = np.sum(X_res * X_res) for b, buck in enumerate(buckets): if len(buck.point_ids): X_hat[buck.point_ids] = all_centroids[c, b] X_res += X_hat multisplits, loss, buckets = learn_multisplits( X_res, X_orig=X_orig, return_centroids=False, return_buckets=True, **kwargs) print("orig loss, loss: ", orig_loss, loss) if loss > orig_loss: X_res -= X_hat continue all_splits[c] = multisplits all_buckets[c] = buckets # update residuals and store centroids # centroid = np.zeros(D, dtype=np.float32) for b, buck in enumerate(buckets): if len(buck.point_ids): centroid = buck.col_means() # centroid /= 2 # TODO rm X_hat[buck.point_ids] = centroid # update centroid here in case we want to regularize it somehow all_centroids[c, b] = centroid X_res -= X_hat print("new X res var / X var: ", X_res.var() / X_orig.var()) if return_buckets: return all_splits, all_centroids, all_buckets return all_splits, all_centroids def mithral_encode(X, multisplits_lists): N, D = X.shape ncodebooks = len(multisplits_lists) X_enc = np.empty((N, ncodebooks), dtype=np.int, order='f') for c in range(ncodebooks): X_enc[:, c] = assignments_from_multisplits(X, multisplits_lists[c]) return np.ascontiguousarray(X_enc) def mithral_lut(q, all_centroids): q = q.reshape(1, 1, -1) # all_centroids is shape ncodebooks, ncentroids, D return (q * all_centroids).sum(axis=2) # ncodebooks, ncentroids def learn_splits_greedy(X, nsplits, verbose=2): N, D = X.shape assert nsplits <= D # # improve numerical stability # scale = np.std(X) # X *= (1. / scale) # precompute sorted lists of values within each dimension, # along with which row they originally were so look can look # up the whole vector (and bucket) associated with each value dim2sorted = [] for dim in range(D): sorted_with_idx = _sort_and_append_orig_idx(X[:, dim]) dim2sorted.append(sorted_with_idx) splits = [] # buckets = [Bucket(N=N, sumX=X.sum(axis=0), sumX2=(X * X).sum(axis=0), buckets = [Bucket(sumX=X.sum(axis=0), sumX2=(X * X).sum(axis=0), point_ids=np.arange(N))] # all_point_infos = [PointInfo(data=row, bucket_id=0) for row in X] bucket_assignments = np.zeros(N, dtype=np.int) # Z = X - X.mean(axis=0) # total_loss = np.sum(Z * Z) # print("initial SSE: ", total_loss) total_loss = sum([bucket.loss for bucket in buckets]) if verbose > 0: print("learn_splits(): initial loss: ", total_loss) # unused_dims = set(np.arange(X.shape[1])) # all_dims = np.arange(D) col_losses = np.zeros(D, dtype=np.float32) # TODO rm? for s in range(nsplits): if verbose > 1: print("================================ finding split #:", s) best_split = Split(dim=-1, val=-np.inf, loss_change=0) # for d in unused_dims: # for d in all_dims: # for d in all_dims[:2]: # TODO rm col_losses[:] = 0 for buck in buckets: col_losses += buck.col_sum_sqs() # try_dims = [np.argmax(col_losses)] # try_dims = np.argsort(col_losses)[-nsplits:] try_dims = np.argsort(col_losses)[-4:] # for d in [dim]: # TODO multiple dim options? if verbose > 1: print("trying dims: ", try_dims) print("with losses: ", col_losses[try_dims]) for d in try_dims: vals_and_point_ids = dim2sorted[d] new_buckets = _split_existing_buckets(buckets) new_total_loss = total_loss if verbose > 2: print("---------------------- dim = ", d) # for i, (val, point_id) in enumerate(vals_and_point_ids): # skip last point since that just puts everything in one bucket, # which is the same as what we're starting out with for val, point_id in vals_and_point_ids[:-1]: # if verbose > 1: # print("i: {}/{}".format(i, len(vals_and_point_ids) - 1)) # info = all_point_infos[point_id] # point, bucket_id = info.data, info.bucket_id point = X[point_id] bucket_id = bucket_assignments[point_id] bucket0 = new_buckets[bucket_id][0] bucket1 = new_buckets[bucket_id][1] old_loss = bucket0.loss + bucket1.loss bucket0.remove_point(point, point_id=point_id) bucket1.add_point(point, point_id=point_id) new_loss = bucket0.loss + bucket1.loss new_total_loss -= old_loss # sub old loss from these buckets new_total_loss += new_loss # add new loss from these buckets loss_change = new_total_loss - total_loss # if loss_change > .1: # should be nonincreasing # print("got loss change: ", loss_change) # print("old total loss:", total_loss) # print("new total loss:", new_total_loss) # assert loss_change <= .1 # should be nonincreasing # # loss should be no worse than having new buckets unused # assert loss_change <= .1 # if verbose > 2: # print("-------- split point_id, val = ", point_id, val) # print("bucket0 point ids, loss after update: ", # bucket0.point_ids, bucket0.loss) # print("bucket1 point ids, loss after update: ", # bucket1.point_ids, bucket1.loss) # print("loss change = {:.3f};\tnew_loss = {:.3f} ".format( # loss_change, new_total_loss)) if loss_change < best_split.loss_change: best_split.dim = d best_split.val = val best_split.loss_change = loss_change if verbose > 2: print("---------------------- split on dim={}, val={:.3f} ".format( best_split.dim, best_split.val)) buckets = [buck.split(X, dim=best_split.dim, val=best_split.val) for buck in buckets] buckets = reduce(lambda b1, b2: b1 + b2, buckets) # flatten pairs for i, buck in enumerate(buckets): ids = np.asarray(list(buck.point_ids), dtype=np.int) bucket_assignments[ids] = i total_loss = sum([bucket.loss for bucket in buckets]) # unused_dims.remove(best_split.dim) splits.append(best_split) if verbose > 3: print('learn_splits(): new loss: {:.3f} from split at dim {}, ' 'value {:.3f}'.format( total_loss, best_split.dim, best_split.val)) if verbose > 2: print('bucket losses: ') print([bucket.loss for bucket in buckets]) print('bucket N, sumX, sumX2') print([bucket.N for bucket in buckets]) print([list(bucket.sumX) for bucket in buckets]) print([list(bucket.sumX2) for bucket in buckets]) # for split in splits: # split.val *= scale # undo preconditioning # total_loss *= scale * scale return splits, total_loss def learn_splits_conditional(X, nsplits, dim_algo='greedy_var', split_algo='mean', **sink): N, D = X.shape assert nsplits <= D # unused_dims = set(np.arange(X.shape[1])) col_means = X.mean(axis=0) # dims = np.arange(D) used_mask = np.ones(D, dtype=np.float32) splits = [] buckets = [Bucket(sumX=X.sum(axis=0), sumX2=(X * X).sum(axis=0), point_ids=np.arange(N))] col_losses = np.zeros(D, dtype=np.float32) for s in range(nsplits): print("---- learning split {}/{}...".format(s + 1, nsplits)) print("current number of buckets: ", len(buckets)) # col_vars = X.var(axis=0) col_losses[:] = 0 for buck in buckets: col_losses += buck.col_sum_sqs() col_losses *= used_mask if dim_algo == 'greedy_var': dim = np.argmax(col_losses) used_mask[dim] = 0 if split_algo == 'mean': val = col_means[dim] new_buckets = [] for buck in buckets: new_buckets += list(buck.split(X=X, dim=dim, val=val)) buckets = new_buckets splits.append(Split(dim=dim, val=val)) return splits, -1 # def learn_splits_simple(X, nsplits, dim_algo='randunif', split_algo='mean', # def learn_splits_simple(X, nsplits, dim_algo='greedy_var', split_algo='median', def learn_splits_simple(X, nsplits, dim_algo='greedy_var', split_algo='mean', **sink): # unused_dims = set(np.arange(X.shape[1])) unused_dims = list(np.arange(X.shape[1])) # random.choice can't use set col_means = X.mean(axis=0) col_vars = X.var(axis=0) col_medians = np.median(X, axis=0) # overall_mean = np.mean(col_means) # overall_median = np.median(col_medians) # overall_var = X.var() var_idxs_descending = np.argsort(col_vars)[::-1] splits = [] for s in range(nsplits): if dim_algo == 'randunif': dim = np.random.choice(unused_dims) unused_dims.remove(dim) elif dim_algo == 'greedy_var': dim = var_idxs_descending[s] if split_algo == 'mean': val = col_means[dim] elif split_algo == 'median': val = col_medians[dim] splits.append(Split(dim=dim, val=val)) return splits, -1 def learn_splits(X, nsplits, return_centroids=True, algo='multisplits', **kwargs): # indirect to particular func; will likely need to try something simpler # for debugging and/or as experimental control # return learn_splits_greedy(X, nsplits, **kwargs) # return learn_splits_simple(X, nsplits, **kwargs) # return learn_splits_conditional(X, nsplits, **kwargs) # return learn_splits_greedy(X, nsplits) # TODO fwd kwargs if algo == 'multisplits': return learn_multisplits( X, nsplits, return_centroids=return_centroids) if algo == 'splits': splits, loss = learn_splits_greedy(X, nsplits) if return_centroids: centroids = centroids_from_splits(X, splits) return splits, loss, centroids return splits, loss def assignments_from_splits(X, splits): nsplits = len(splits) indicators = np.empty((nsplits, len(X)), dtype=np.int) for i, split in enumerate(splits): indicators[i] = X[:, split.dim] > split.val # compute assignments by treating indicators in a row as a binary num # scales = (2 ** np.arange(nsplits)).astype(np.int) scales = (1 << np.arange(nsplits)).astype(np.int) return (indicators.T * scales).sum(axis=1).astype(np.int) def assignments_from_multisplits(X, splits): N, _ = X.shape nsplits = len(splits) # indicators = np.zeros((nsplits, len(X)), dtype=np.int) assert len(splits) >= 1 # dim0 = splits[0].dim # assert len(splits[0].vals) == 1 # only 1 initial split # indicators[0] = X > splits[0].vals[0] max_ngroups = len(splits[-1].vals) nsplits_affecting_group_id = int(np.log2(max_ngroups)) assert 1 << nsplits_affecting_group_id == max_ngroups # power of 2 # np.log2(max_nsplits) # determine group ids for each point; this is the one that's annoying # because the number of bits changes after split group_ids = np.zeros(N, dtype=np.int) for i in range(min(nsplits, nsplits_affecting_group_id)): split = splits[i] vals = split.vals[group_ids] # x = X[:, split.dim] # if split.offset is not None: # x = x - split.offset # if split.scaleby is not None: # x = x * split.scaleby # indicators = x > vals indicators = split.preprocess_x(X[:, split.dim]) > vals group_ids = (group_ids * 2) + indicators if nsplits <= nsplits_affecting_group_id: return group_ids # compute remaining bits assignments = np.copy(group_ids) for i in range(nsplits_affecting_group_id, nsplits): split = splits[i] vals = split.vals[group_ids] # x = X[:, split.dim] # if split.offset is not None: # x = x - split.offset # if split.scaleby is not None: # x = x * split.scaleby # indicators = x > vals indicators = split.preprocess_x(X[:, split.dim]) > vals assignments = (assignments * 2) + indicators return assignments def _centroids_from_assignments(X, assignments, ncentroids): centroids = np.empty((ncentroids, X.shape[1]), dtype=X.dtype) for c in range(ncentroids): centroids[c] = X[assignments == c].mean(axis=0) return centroids def centroids_from_splits(X, splits): ncentroids = int(1 << len(splits)) assignments = assignments_from_splits(X, splits) return _centroids_from_assignments(X, assignments, ncentroids=ncentroids) @_memory.cache def learn_splits_in_subspaces(X, subvect_len, nsplits_per_subs, return_centroids=True, algo='multisplits', verbose=2): N, D = X.shape # N /= 100 # TODO rm after debug splits_lists = [] nsubs = int(np.ceil(D) / subvect_len) # stuff for sse stats tot_sse = 0 X_bar = X - np.mean(X, axis=0) col_sses = np.sum(X_bar * X_bar, axis=0) + 1e-14 tot_sse_using_mean = np.sum(col_sses) if verbose > 1: print("original sum of sses within each col: ", tot_sse_using_mean) if return_centroids: ncentroids = int(2 ** nsplits_per_subs) # this order seems weird, but matches _learn_centroids, etc; helps with # eventual vectorized lookups centroids = np.empty((ncentroids, nsubs, subvect_len), dtype=X.dtype) for m in range(nsubs): start_col = m * subvect_len end_col = start_col + subvect_len X_subs = X[:, start_col:end_col] splits, sse, subs_centroids = learn_splits( X_subs, nsplits=nsplits_per_subs, verbose=(verbose - 1), return_centroids=True, algo=algo) centroids[:, m, :] = subs_centroids splits_lists.append(splits) tot_sse += sse if verbose > 1: # print("col sses in subspace: ", col_sses[start_col:end_col]) # print("sum col sses in subspace: ", col_sses[start_col:end_col].sum()) # print("buckets claim sse:", sse) # print("N: ", N) # print("(sse / N)", (sse / N)) # print("np.var(X_subs)", np.var(X_subs)) orig_sse_in_subs = col_sses[start_col:end_col].sum() # print("learning splits: mse / var(X) in subs {}/{} = {:3g}".format( # m + 1, nsubs, (sse / N) / np.var(X_subs))) print("learning splits: sse / orig sse in subs {}/{} = {:3g}".format( m + 1, nsubs, sse / orig_sse_in_subs)) # import sys; sys.exit() # print("exiting after one subspace") # import sys; sys.exit() if verbose > 0: print("-- learn_splits_in_subspaces: new / orig mse: {:.3g}".format( tot_sse / tot_sse_using_mean)) # print("tot_sse_using_mean: ", tot_sse_using_mean) if return_centroids: return splits_lists, centroids return splits_lists def encode_using_splits(X, subvect_len, splits_lists, split_type='single'): N, D = X.shape nsubs = int(np.ceil(D) / subvect_len) X_enc = np.empty((X.shape[0], nsubs), dtype=np.int, order='f') for m in range(nsubs): start_col = m * subvect_len end_col = start_col + subvect_len X_subs = X[:, start_col:end_col] if split_type == 'single': X_enc[:, m] = assignments_from_splits(X_subs, splits_lists[m]) elif split_type == 'multi': X_enc[:, m] = assignments_from_multisplits(X_subs, splits_lists[m]) return np.ascontiguousarray(X_enc) def _plot_stuff_on_trace(): import matplotlib as mpl import matplotlib.pyplot as plt from joblib import Memory _memory = Memory('.', verbose=0) mpl.rcParams['lines.linewidth'] = .5 @_memory.cache def _load_trace(): return np.loadtxt('assets/debug/Trace/Trace_TRAIN.txt') # try_ndims = 128 # try_ndims = 64 try_ndims = 4 # limit_n = 20 # limit_n = 50 # limit_n = 200 limit_n = 500 # X = np.loadtxt('assets/debug/Trace/Trace_TRAIN.txt')[:limit_n] X = _load_trace()[:limit_n] y = (X[:, 0] - 1).astype(np.int) X = X[:, 1:] _, axes = plt.subplots(3, 4, figsize=(13, 9), sharey=True) colors = ('blue', 'red', 'green', 'black') axes[0, 0].set_title('Trace Dataset\n(colored by class)') for lbl in np.unique(y): X_subset = X[y == lbl] axes[0, 0].plot(X_subset.T, color=colors[lbl]) # visualize output with only 1 codebook (no need for updates) ncodebooks = 1 splits, centroids, buckets = learn_mithral( X, ncodebooks, return_buckets=True, try_ndims=try_ndims, niters=1) centroids = centroids[0] # only one codebook axes[0, 1].set_title('centroids') axes[0, 1].plot(centroids.T) X_hat = np.zeros_like(X) for c, splitlist in enumerate(splits): for s, split in enumerate(splitlist): assert len(splitlist) == 4 vals = (split.vals / split.scaleby) + split.offset for val in vals: axes[0, c].scatter(split.dim, val, color=colors[s], marker='o', zorder=5) for b in buckets[0]: # only one codebook, so use first list if b.N > 0: X_hat[b.point_ids] = b.col_means() X_res = X - X_hat axes[0, 2].set_title('reconstructions') axes[0, 2].plot(X_hat.T) # axes[0, 3].set_title('residuals (mean={:.2f})'.format(X_res.mean())) axes[0, 3].set_title('residuals (var={:.2f})'.format(X_res.var())) axes[0, 3].plot(X_res.T) # visualize output with only 2 codebooks, no updates ncodebooks = 2 splits, centroids, buckets = learn_mithral( X, ncodebooks, return_buckets=True, try_ndims=try_ndims, niters=1) # centroids = centroids[0] # only one codebook axes[1, 0].set_title('centroids[0]') axes[1, 0].plot(centroids[0].T) axes[1, 1].set_title('centroids[1]') axes[1, 1].plot(centroids[1].T) X_hat = np.zeros_like(X) # print("splits: ", splits) for c, splitlist in enumerate(splits): for s, split in enumerate(splitlist): assert len(splitlist) == 4 vals = (split.vals / split.scaleby) + split.offset for val in vals: axes[1, c].scatter(split.dim, val, color=colors[s]) for c in range(len(buckets)): # for each codebook for b, buck in enumerate(buckets[c]): if buck.N > 0: X_hat[buck.point_ids] += centroids[c, b] X_res = X - X_hat axes[1, 2].set_title('reconstructions') axes[1, 2].plot(X_hat.T) # axes[1, 3].set_title('residuals (mean={:.2f})'.format(X_res.mean())) axes[1, 3].set_title('residuals (var={:.2f})'.format(X_res.var())) axes[1, 3].plot(X_res.T) # visualize output with only 2 codebooks, with centroid updates ncodebooks = 2 splits, centroids, buckets = learn_mithral( X, ncodebooks, return_buckets=True, try_ndims=try_ndims, niters=1) axes[2, 0].set_title('centroids[0]') axes[2, 0].plot(centroids[0].T) axes[2, 1].set_title('centroids[1]') axes[2, 1].plot(centroids[1].T) X_hat = np.zeros_like(X) for c in range(len(buckets)): # for each codebook for b, buck in enumerate(buckets[c]): if buck.N > 0: X_hat[buck.point_ids] += centroids[c, b] X_res = X - X_hat axes[2, 2].set_title('reconstructions') axes[2, 2].plot(X_hat.T) # axes[2, 3].set_title('residuals (mean={:.2f})'.format(X_res.mean())) axes[2, 3].set_title('residuals (var={:.2f})'.format(X_res.var())) axes[2, 3].plot(X_res.T) plt.tight_layout() plt.show() def test_encoded_ops(): N, C, K = 100, 8, 16 X_enc = np.random.randint(K, size=(N, C)) # print(X_enc) X_bin = _densify_X_enc(X_enc) # print(X_enc_binary) assert np.all(X_bin.sum(axis=1) == C) XtX = _XtX_encoded(X_enc) XtX2 = X_bin.T @ X_bin assert np.all(XtX == XtX2) M = 17 Y = np.random.randn(N, M).astype(np.float32) XtY = _XtY_encoded(X_enc, Y) XtY2 = X_bin.T @ Y # print(XtY[:2]) # print(XtY2[:2]) assert np.all(XtY == XtY2) D = C * K W = np.random.randn(D, M).astype(np.float32) XW = _XW_encoded(X_enc, W) XW2 = X_bin @ W assert np.all(XW == XW2) def main(): test_encoded_ops() # print(_pq_codebook_start_end_idxs(6, 3)) # print(_pq_codebook_start_end_idxs(8, 3)) # print(_pq_codebook_start_end_idxs(9, 3)) # print(_pq_codebook_start_end_idxs(10, 3)) if __name__ == '__main__': main()
#!/usr/bin/env python from __future__ import division, absolute_import import abc import matplotlib.pyplot as plt import numpy as np import seaborn as sb from . import product_quantize as pq from . import subspaces as subs from . import clusterize from .utils import kmeans # ================================================================ misc funcs def dists_elemwise_sq(x, q): diffs = x - q return diffs * diffs def dists_elemwise_l1(x, q): return np.abs(x - q) def dists_elemwise_dot(x, q): return x * q def extract_random_rows(X, how_many, remove_from_X=True): split_start = np.random.randint(len(X) - how_many - 1) split_end = split_start + how_many rows = np.copy(X[split_start:split_end]) if remove_from_X: return np.vstack((X[:split_start], X[split_end:])), rows return X, rows # XXX: not clear whether this function is correct in general, but # does always pass the asserts (which capture the invariants we want) def _insert_zeros(X, nzeros): N, D = X.shape D_new = D + nzeros X_new = np.zeros((N, D_new), dtype=X.dtype) # print("attempting to insert {} zeros into X of shape {}".format(nzeros, X.shape)) step = int(D / (nzeros + 1)) - 1 step = max(1, step) # print("using step: ", step) for i in range(nzeros): in_start = step * i in_end = in_start + step # out_start = in_start + i + 1 out_start = (step + 1) * i out_end = out_start + step X_new[:, out_start:out_end] = X[:, in_start:in_end] # out_start = out_end # out_end += step out_end += 1 # account for the last 0 remaining_len = D - in_end out_remaining_len = D_new - out_end # print "step", step # print "in_start, in_end", in_start, in_end # print "out_start, out_end", out_start, out_end # print "D, D_new", D, D_new # print "remaining_len, out_remaining_len", remaining_len, out_remaining_len assert remaining_len == out_remaining_len assert remaining_len >= 0 if remaining_len: # X_new[:, out_end:out_end+remaining_len] = X[:, in_end:D] X_new[:, out_end:] = X[:, in_end:] # print("first cols of old and new X:") # print(X[:, 0]) # print(X_new[:, 0]) # print(X_new.shape) # print((X_new.sum(axis=0) != 0).sum()) assert X.shape[0] == X_new.shape[0] cols_nonzero = X_new.sum(axis=0) != 0 orig_cols_nonzero = X.sum(axis=0) != 0 # new_cols_nonzero = cols_nonzero & (~orig_cols_nonzero) # print("zero cols: ", np.where(~cols_nonzero)[0]) assert cols_nonzero.sum() == orig_cols_nonzero.sum() nzeros_added = (~cols_nonzero).sum() - (~orig_cols_nonzero).sum() assert nzeros_added == nzeros # assert np.array_equal(X[:, 0], X_new[:, 0]) # assert np.array_equal(X[:, -1], X_new[:, -1]) return X_new # def ensure_num_cols_multiple_of(X, multiple_of, min_ncols=-1): def ensure_num_cols_multiple_of(X, multiple_of): remainder = X.shape[1] % multiple_of if remainder > 0: return _insert_zeros(X, multiple_of - remainder) # # TODO rm and uncomment above after debug # add_ncols = multiple_of - remainder # new_ncols = X.shape[1] + add_ncols # new_X = np.zeros((X.shape[0], new_ncols), dtype=X.dtype) # new_X[:, :X.shape[1]] = X # return new_X return X def _learn_best_quantization(luts): assert luts.ndim == 2 # luts can be a bunch of vstacked luts, but not 3D best_loss = np.inf best_alpha = None best_floors = None best_scale_by = None for alpha in [.001, .002, .005, .01, .02, .05, .1]: # alpha_pct = int(100 * alpha) alpha_pct = 100 * alpha # compute quantized luts this alpha would yield floors = np.percentile(luts, alpha_pct, axis=0) luts_offset = np.maximum(0, luts - floors) ceil = np.percentile(luts_offset, 100 - alpha_pct) scale_by = 255. / ceil # if only_shift: # scale_by = 1 << int(np.log2(scale_by)) luts_quantized = np.floor(luts_offset * scale_by).astype(np.int) luts_quantized = np.minimum(255, luts_quantized) # compute err luts_ideal = (luts - luts_offset) * scale_by diffs = luts_ideal - luts_quantized loss = np.sum(diffs * diffs) if loss <= best_loss: best_loss = loss best_alpha = alpha best_floors = floors best_scale_by = scale_by return best_floors, best_scale_by, best_alpha # ================================================================ Quantizers # ------------------------------------------------ Abstract Base Class class MultiCodebookEncoder(abc.ABC): def __init__(self, ncodebooks, ncentroids=256, quantize_lut=False, upcast_every=-1, accumulate_how='sum'): self.ncodebooks = ncodebooks self.ncentroids = ncentroids self.quantize_lut = quantize_lut self.upcast_every = upcast_every if upcast_every >= 1 else 1 self.upcast_every = min(self.ncodebooks, self.upcast_every) assert self.upcast_every in (1, 2, 4, 8, 16, 32, 64, 128, 256) self.accumulate_how = accumulate_how self.code_bits = int(np.log2(self.ncentroids)) # for fast lookups via indexing into flattened array self.offsets = (np.arange(self.ncodebooks, dtype=np.int) * self.ncentroids) def name(self): return "{}_{}x{}b_quantize={}".format( self.preproc, self.ncodebooks, self.code_bits, int(self.quantize_lut)) def params(self): return {'ncodebooks': self.ncodebooks, 'code_bits': self.code_bits, 'quantize': self.quantize_lut} def _learn_lut_quantization(self, X, Q=None): if self.quantize_lut: # TODO put this logic in separate function print("learning quantization...") # print("initial Q: ", Q) if Q is None: # num_rows = min(10 * 1000, len(X) // 2) # _, queries = extract_random_rows( # X[num_rows:], how_many=1000, remove_from_X=False) # X = X[:num_rows] # limit to first 10k rows of X _, Q = extract_random_rows( X, how_many=1000, remove_from_X=False) Q = Q.T # want each row to be one query, not each col # Q = self._pad_ncols(Q) # if self.preproc == 'OPQ': # Q = pq.opq_rotate(Q, self.R) # elif self.preproc == 'BOPQ': # Q = pq.bopq_rotate(Q, self.rotations) # elif self.preproc == 'GEHT': # Q = Q[:, self.perm] # print("Q shape: ", Q.shape) # compute luts for all the queries # luts = [self.encode_Q(q, quantize=False) for q in Q] luts = self.encode_Q(Q, quantize=False) # luts = np.vstack(luts) # print("ncodebooks: ", self.ncodebooks) # print("luts shape: ", luts.shape) assert luts.shape == (len(Q), self.ncodebooks, self.ncentroids) luts = np.moveaxis(luts, 2, 1) assert luts.shape == (len(Q), self.ncentroids, self.ncodebooks) luts = luts.reshape(len(Q) * self.ncentroids, self.ncodebooks) self.lut_offsets, self.scale_by, _ = _learn_best_quantization(luts) # print("self.lut_offsets.shape", self.lut_offsets.shape) # print("self.scale_by.shape", self.scale_by.shape) # print("self.scale_by", self.scale_by) assert self.lut_offsets.shape == (self.ncodebooks,) # self.lut_offsets = self.lut_offsets[:, np.newaxis] self.total_lut_offset = np.sum(self.lut_offsets) # print("lut offsets: ", self.lut_offsets) def dists_enc(self, X_enc, Q_luts, unquantize=True, offset=None, scale=None): X_enc = np.ascontiguousarray(X_enc) if unquantize: offset = self.total_lut_offset if offset is None else offset scale = self.scale_by if scale is None else scale all_dists = np.empty((len(Q_luts), len(X_enc)), dtype=np.float32) for i, lut in enumerate(Q_luts): centroid_dists = lut.ravel()[X_enc.ravel()] dists = centroid_dists.reshape(X_enc.shape) if self.upcast_every < 2 or not self.quantize_lut: dists = dists.sum(axis=-1) else: dists = dists.reshape(dists.shape[0], -1, self.upcast_every) if self.accumulate_how == 'sum': # sum upcast_every vals, then clip to mirror saturating # unsigned addition, then sum without saturation (like u16) dists = dists.sum(2) dists = np.clip(dists, 0, 255).sum(axis=-1) elif self.accumulate_how == 'mean': # mirror hierarchical avg_epu8 # print("reducing using mean!") # print("fraction of low bits that are 1: ", # np.mean(dists % 2 == 1)) # ya, ~.5, or maybe ~.495 while dists.shape[-1] > 2: dists = (dists[:, :, ::2] + dists[:, :, 1::2] + 1) // 2 dists = (dists[:, :, 0] + dists[:, :, 1] + 1) // 2 dists = dists.sum(axis=-1) # clipping not needed # undo biasing; if low bits are {0,0} or {1,1}, no bias # from the averaging; but if {0,1}, then rounds up by # .5; happens with prob ~=~ .5, so each avg op adds .25; # the other tricky thing here is that rounding up when # you're averaging averages biases it even farther # base_bias = .5 * .5 # assert self.upcast_every >= 2 # bias_per_upcast = 0 # nlevels = int(np.log2(self.upcast_every)) # for level in range(nlevels): # num_avg_ops = self.upcast_every / (2 << level) # print("num_avg_ops: ", num_avg_ops) # bias_per_op = (1 << level) * base_bias # print("level multiplier: ", 1 << level) # bias_per_upcast += num_avg_ops * bias_per_op # bias = bias_per_upcast * (self.ncodebooks / self.upcast_every) # num_avg_ops = (self.upcast_every - 1) * ( # self.ncodebooks / self.upcast_every) # num_avg_ops = (self.upcast_every - 1) * np.sqrt( # self.ncodebooks / self.upcast_every) # num_avg_ops = (self.upcast_every - 1) # bias = num_avg_ops * base_bias # bias = (self.ncodebooks / 2) * int(np.log2(self.upcast_every)) # bias = (self.ncodebooks / 2) * int(np.log2(self.upcast_every)) # bias = 0 # dists -= int(bias * self.upcast_every) dists *= self.upcast_every # convert mean to sum # I honestly don't know why this is the formula, but wow # does it work well bias = self.ncodebooks / 4 * np.log2(self.upcast_every) dists -= int(bias) else: raise ValueError("accumulate_how must be 'sum' or 'mean'") if self.quantize_lut and unquantize: # dists = (dists / self.scale_by) + self.total_lut_offset dists = (dists / scale) + offset all_dists[i] = dists return all_dists.T # ------------------------------------------------ Product Quantization def _learn_centroids(X, ncentroids, ncodebooks, subvect_len): ret = np.empty((ncentroids, ncodebooks, subvect_len)) # print("_learn_centroids(): running kmeans...") tot_sse = 0 X_bar = X - np.mean(X, axis=0) col_sses = np.sum(X_bar * X_bar, axis=0) + 1e-14 tot_sse_using_mean = np.sum(col_sses) for i in range(ncodebooks): print("running kmeans in subspace {}/{}...".format( i + 1, ncodebooks), end=" ") start_col = i * subvect_len end_col = start_col + subvect_len X_in = X[:, start_col:end_col] # centroids, labels = kmeans(X_in, ncentroids) centroids, labels, sse = kmeans(X_in, ncentroids, return_sse=True) # X_bar = X_in - np.mean(X_in, axis=0) # sse_using_mean = np.sum(X_bar * X_bar) + 1e-14 subspace_sse = np.sum(col_sses[start_col:end_col]) print("mse / {{var(X_subs), var(X)}}: {:.3g}, {:.3g}".format( sse / subspace_sse, sse * ncodebooks / tot_sse_using_mean)) tot_sse += sse # print("centroids shape: ", centroids.shape) # print("ret shape: ", ret.shape) ret[:, i, :] = centroids print("--- total mse / var(X): {:.3g}".format(tot_sse / tot_sse_using_mean)) return ret def _parse_codebook_params(D, code_bits=-1, bits_per_subvect=-1, ncodebooks=-1): if ncodebooks < 0: ncodebooks = code_bits // bits_per_subvect elif code_bits < 1: code_bits = bits_per_subvect * ncodebooks elif bits_per_subvect < 1: bits_per_subvect = code_bits // ncodebooks ncentroids = int(2 ** bits_per_subvect) subvect_len = D // ncodebooks assert code_bits % bits_per_subvect == 0 if D % subvect_len: print("D, ncodebooks, subvect_len = ", D, ncodebooks, subvect_len) assert D % subvect_len == 0 # TODO rm this constraint return ncodebooks, ncentroids, subvect_len def _fit_pq_lut(q, centroids, elemwise_dist_func): _, ncodebooks, subvect_len = centroids.shape q = q.reshape((1, ncodebooks, subvect_len)) q_dists = np.sum(centroids * q, axis=-1) return q_dists # ncentroids, ncodebooks, row-major class PQEncoder(MultiCodebookEncoder): def __init__(self, ncodebooks, ncentroids=256, elemwise_dist_func=dists_elemwise_dot, preproc='PQ', encode_algo=None, quantize_lut=False, upcast_every=-1, accumulate_how='sum', **preproc_kwargs): super().__init__( ncodebooks=ncodebooks, ncentroids=ncentroids, quantize_lut=quantize_lut, upcast_every=upcast_every, accumulate_how=accumulate_how) self.elemwise_dist_func = elemwise_dist_func self.preproc = preproc self.encode_algo = encode_algo self.preproc_kwargs = preproc_kwargs def _pad_ncols(self, X): return ensure_num_cols_multiple_of(X, self.ncodebooks) def fit(self, X, Q=None): self.subvect_len = int(np.ceil(X.shape[1] / self.ncodebooks)) X = self._pad_ncols(X) self.centroids = None if self.preproc == 'BOPQ': self.centroids, _, self.rotations = pq.learn_bopq( X, ncodebooks=self.ncodebooks, codebook_bits=self.code_bits, **self.preproc_kwargs) elif self.preproc == 'OPQ': self.centroids, _, self.R = pq.learn_opq( X, ncodebooks=self.ncodebooks, codebook_bits=self.code_bits, **self.preproc_kwargs) elif self.preproc == 'GEHT': self.perm = subs.greedy_eigenvector_threshold( X, subspace_len=self.subvect_len, **self.preproc_kwargs) assert X.shape[1] == len(set(self.perm)) X = X[:, self.perm] if self.centroids is None: if self.encode_algo in ('splits', 'multisplits'): assert self.encode_algo != 'splits' # TODO rm self.splits_lists, self.centroids = \ clusterize.learn_splits_in_subspaces( X, subvect_len=self.subvect_len, nsplits_per_subs=self.code_bits, algo=self.encode_algo) # print("centroids shape: ", self.centroids.shape) # # TODO rm # # yep, yields identical errs as mithral with pq_perm_algo='end' # self.splits_lists, self.centroids = clusterize.learn_mithral( # X, ncodebooks=self.ncodebooks) # print("centroids shape: ", self.centroids.shape) else: self.centroids = _learn_centroids( X, self.ncentroids, self.ncodebooks, self.subvect_len) self._learn_lut_quantization(X, Q) def name(self): return "{}_{}".format(self.preproc, super().name()) def params(self): d = super().params() d['_preproc'] = self.preproc return d def encode_Q(self, Q, quantize=True): # quantize param enables quantization if set in init; separate since # quantization learning needs to call this func, but vars like # lut_offsets aren't set when this function calls it Q = np.atleast_2d(Q) Q = self._pad_ncols(Q) if self.preproc == 'OPQ': Q = pq.opq_rotate(Q, self.R) elif self.preproc == 'BOPQ': Q = pq.bopq_rotate(Q, self.rotations) elif self.preproc == 'GEHT': Q = Q[:, self.perm] luts = np.zeros((Q.shape[0], self.ncodebooks, self.ncentroids)) # print("Q shape: ", Q.shape) for i, q in enumerate(Q): lut = _fit_pq_lut(q, centroids=self.centroids, elemwise_dist_func=self.elemwise_dist_func) if self.quantize_lut and quantize: lut = np.maximum(0, lut - self.lut_offsets) lut = np.floor(lut * self.scale_by).astype(np.int) lut = np.minimum(lut, 255) luts[i] = lut.T return luts def encode_X(self, X, **sink): X = self._pad_ncols(X) if self.preproc == 'OPQ': X = pq.opq_rotate(X, self.R) elif self.preproc == 'BOPQ': X = pq.bopq_rotate(X, self.rotations) elif self.preproc == 'GEHT': X = X[:, self.perm] if self.encode_algo in ('splits', 'multisplits'): split_type = ('multi' if self.encode_algo == 'multisplits' else 'single') idxs = clusterize.encode_using_splits( X, self.subvect_len, self.splits_lists, split_type=split_type) else: idxs = pq._encode_X_pq(X, codebooks=self.centroids) return idxs + self.offsets # offsets let us index into raveled dists # ------------------------------------------------ Mithral # def _mithral_quantize_luts(luts, lut_work_const, force_power_of_2=False): def _mithral_quantize_luts(luts, lut_work_const, force_power_of_2=True): nqueries, ncodebooks, ncentroids = luts.shape # if lut_work_const < 0: # not time constrained # assert luts.shape == (nqueries, ncodebooks, ncentroids) # luts2d = np.moveaxis(luts, 2, 1) # assert luts2d.shape == (nqueries, ncentroids, ncodebooks) # luts2d = luts2d.reshape(nqueries * ncentroids, ncodebooks) # # if True: # if False: # # ax = sb.distplot(luts.ravel(), hist=False, rug=True) # _, ax = plt.subplots(1, figsize=(13, 5)) # # sb.violinplot(data=luts2d, inner='point', ax=ax) # # sb.boxenplot(data=luts2d, ax=ax) # means = luts2d.mean(axis=0) # # # rm largest and smallest entry in each col # # argmaxs = np.argmax(luts2d, axis=0) # # argmins = np.argmax(luts2d, axis=0) # # for c in range(luts.shape[1]): # # luts2d[argmins[c], c] = means[c] # # luts2d[argmaxs[c], c] = means[c] # maxs = luts2d.max(axis=0) # mins = luts2d.min(axis=0) # gaps = maxs - mins # max_idx = np.argmax(gaps) # print(f"biggest gap = {np.max(gaps)} at idx {max_idx}") # gaps[max_idx] = 0 # max_idx = np.argmax(gaps) # print(f"2nd biggest gap = {np.max(gaps)} at idx {max_idx}") # gaps[max_idx] = 0 # max_idx = np.argmax(gaps) # print(f"3rd biggest gap = {np.max(gaps)} at idx {max_idx}") # gaps[max_idx] = 0 # max_idx = np.argmax(gaps) # print(f"4th biggest gap = {np.max(gaps)} at idx {max_idx}") # gaps[max_idx] = 0 # max_idx = np.argmax(gaps) # print(f"5th biggest gap = {np.max(gaps)} at idx {max_idx}") # # for i in range(len(luts2d)): # # row = luts2d[i] # # luts2d[i, row == mins] = means # # luts2d[i, row == maxs] = means # luts2d -= mins # # luts2d -= means # # luts2d *= 255 / (maxs - mins).max() # luts2d *= 255 / gaps.max() # luts2d = np.minimum(luts2d, 255) # sb.stripplot(data=luts2d, ax=ax, size=4) # ax.set_xlabel('Query dist to centroids (lut dist histogram)') # ax.set_ylabel('Fraction of queries') # plt.show() # import sys; sys.exit() # offsets, scale, _ = _learn_best_quantization(luts2d) # offsets = offsets[np.newaxis, :, np.newaxis] # luts = np.maximum(0, luts - offsets) * scale # luts = np.floor(luts).astype(np.int) # luts = np.minimum(255, luts) # return luts, offsets.sum(), scale # luts = np.zeros((Q.shape[0], self.ncodebooks, self.ncentroids)) mins = luts.min(axis=(0, 2)) maxs = luts.max(axis=(0, 2)) gaps = maxs - mins # gaps[np.argmax(gaps)] = 0 # use 2nd highest gap = np.max(gaps) if force_power_of_2: exponent = np.ceil(np.log2(gap)) scale = 2 ** int(-exponent) # scale is a power of 2, so can just shift scale *= (255.5 - 1e-10) # so max val is at most 255 else: scale = (255.5 - 1e-10) / gap offsets = mins[np.newaxis, :, np.newaxis] luts_quantized = (luts - offsets) * scale luts_quantized = (luts_quantized + .5).astype(np.int) # luts_quantized = np.minimum(luts_quantized, 255) assert np.min(luts_quantized) >= 0 assert np.max(luts_quantized) <= 255. # print("total offset: ", mins.sum()) return luts_quantized, offsets.sum(), scale # # compute offset taking into account stuff getting rounded down # luts_hat = (luts / scale) + offsets # diffs = luts - luts_hat # print("mean of diffs: ", diffs.mean()) # offset = diffs.mean() + offsets.sum() # return luts_quantized, offset, scale class MithralEncoder(MultiCodebookEncoder): def __init__(self, ncodebooks, lut_work_const=-1): super().__init__( ncodebooks=ncodebooks, ncentroids=16, # quantize_lut=True, upcast_every=64, # quantize_lut=True, upcast_every=32, quantize_lut=True, upcast_every=16, # quantize_lut=True, upcast_every=8, # quantize_lut=True, upcast_every=4, # quantize_lut=True, upcast_every=2, # quantize_lut=True, upcast_every=1, accumulate_how='mean') self.lut_work_const = lut_work_const def name(self): return "{}_{}".format('mithral', super().name()) def params(self): return {'ncodebooks': self.ncodebooks, 'lut_work_const': self.lut_work_const} def fit(self, X, Q=None): self.splits_lists, self.centroids = clusterize.learn_mithral( X, self.ncodebooks, lut_work_const=self.lut_work_const) # self._learn_lut_quantization(X, Q) def encode_X(self, X): idxs = clusterize.mithral_encode(X, self.splits_lists) return idxs + self.offsets def encode_Q(self, Q, quantize=True): Q = np.atleast_2d(Q) luts = np.zeros((Q.shape[0], self.ncodebooks, self.ncentroids)) for i, q in enumerate(Q): luts[i] = clusterize.mithral_lut(q, self.centroids) if self.quantize_lut: luts, offset, scale = _mithral_quantize_luts(luts, self.lut_work_const) return luts, offset, scale return luts, 0, 1 def main(): X = np.ones((3, 75), dtype=np.int) _insert_zeros(X, 53) if __name__ == '__main__': main()
#!/bin/env/python import os import numpy as np import matplotlib.pyplot as plt import seaborn as sb import pandas as pd import pathlib as pl # from . import files from . import amm_results as res from . import amm_methods as ameth sb.set_context('poster') # sb.set_context('talk') # sb.set_cmap('tab10') RESULTS_DIR = pl.Path('results/amm') FIGS_SAVE_DIR = pl.Path('../figs/amm') if not os.path.exists(FIGS_SAVE_DIR): FIGS_SAVE_DIR.mkdir(parents=True) def save_fig(name): plt.savefig(os.path.join(FIGS_SAVE_DIR, name + '.png'), dpi=300, bbox_inches='tight') def _xlabel_for_xmetric(x_metric): return {'d': 'Sketch Size', 'secs': 'Time (s)', 'muls': 'Number of Multiplies', 'nlookups': 'Number of Lookups', 'ops': 'Number of Operations', 'Latency': 'Latency (ms)', 'Throughput': 'Throughput (elements/s)'}[x_metric] # if x_metric == 'd': # return 'Log2(Sketch Size)' # elif x_metric == 'secs': # return 'Time (s)' # elif x_metric == 'muls': # # return 'Log10(# of Multiplies)' # return 'Number of Multiplies' # elif x_metric == 'nlookups': # # return 'Log10(# of Table Lookups)' # return 'Number of Table Lookups' # elif x_metric == 'ops': # # return 'Log10(# of Operations)' # return 'Number of Operations' # elif x_metric == 'Latency': # return 'Latency (ms)' def _clean_results_df(df, default_D=None): # for Exact, set d = D if default_D is not None and ('d' in df): mask = df['d'].isna() df.loc[mask, 'd'] = default_D # clean up column names + other strings for old, new in [('method', 'Method'), ('acc_amm', 'Accuracy'), ('r_sq', 'R-Squared'), ('nmultiplies', 'muls')]: try: df.rename({old: new}, axis=1, inplace=True) except KeyError: pass # replace_dict = {'Bolt+MultiSplits': 'Ours', # replace_dict = {'Mithral': 'Ours', replace_dict = {'Mithral': 'Ours', 'MithralPQ': 'OursPQ', 'Exact': 'Brute Force', 'CooccurSketch': 'CD'} # def _replace_method_name(name): # return replace_dict.get(name, name) df['Method'] = df['Method'].apply(lambda s: replace_dict.get(s, s)) # create ops column that sums number of multiplies + lookups df['muls'] = df['muls'].fillna(0) mask = ~df['nlookups'].isna() df['ops'] = df['muls'] df['ops'].loc[mask] += df['nlookups'].loc[mask] # df['muls'] = np.log10(df['muls']) # df['ops'] = np.log10(df['ops']) # join with cpp timing results matmul_latencies, matmul_thruputs = res.load_matmul_times_for_n_d_m() sketch_latencies, sketch_thruputs = res.load_sketch_times_for_n_d_m() # multisplit_latencies, multisplit_thruputs = \ # res.load_multisplit_times_for_n_d_m() mithral_latencies, mithral_thruputs = res.load_mithral_times_for_n_d_m() bolt_latencies, bolt_thruputs = res.load_bolt_times_for_n_d_m() # row_dicts = [] all_latencies = [] all_thruputs = [] # for _, row in df.itertuples(): # print("d col: ") # print(df['d']) fast_sketch_methods = set([m.lower() for m in ameth.FAST_SKETCH_METHODS]) slow_sketch_methods = set([m.lower() for m in ameth.SLOW_SKETCH_METHODS]) for _, row in df.iterrows(): # row = dict(*row) N, D, M = [int(row[k]) for k in ('N', 'D', 'M')] method = row['Method'].lower() # if 'split' in method.lower(): # print("using method: ", method) if method in ('bolt', 'ours', 'ourspq'): # TODO check if in vq methods, instead of hardcoding ncodebooks = int(row['ncodebooks']) key = (N, D, M, ncodebooks) if method in ('ours', 'ourspq'): # latencies = multisplit_latencies[key] # thruputs = multisplit_thruputs[key] latencies = mithral_latencies[key] thruputs = mithral_thruputs[key] elif method == 'bolt': latencies = bolt_latencies[key] thruputs = bolt_thruputs[key] # all_latencies.append(np.median(latencies)) # all_thruputs.append(np.median(thruputs)) elif method == 'brute force': key = (N, D, M) latencies = matmul_latencies[key] thruputs = matmul_thruputs[key] elif method in fast_sketch_methods: d = int(row['d']) key = (N, D, M, d) latencies = sketch_latencies[key] thruputs = sketch_thruputs[key] else: # slow sketch-based methods # print("method: ", method) # assert method in slow_sketch_methods # print("method: ", method) # print("fast sketch methods: ", fast_sketch_methods) # assert False # TODO rm secs = row['secs'] lat = secs * 1000 thruput = N * M / secs latencies = [lat] thruputs = [thruput] # print("d: ", d) # print("key:", key) # print("sketch_latencies:") # import pprint # pprint.pprint(sketch_latencies) # secs = row['secs'] # lat = secs * 1000 # thruput = N * M / secs # # # version where we pretend same efficiency as matmul # # nmuls = int(row['muls']) # # exact_nmuls = N * D * M # # scale = nmuls / exact_nmuls # # lat *= scale # # thruput /= scale # all_latencies.append(lat) # all_thruputs.append(thruput) all_latencies.append(np.mean(latencies)) all_thruputs.append(np.mean(thruputs)) # print("len latencies: ", len(all_latencies)) # print("len thruputs: ", len(all_thruputs)) # print("df len: ", df.shape[0]) df['Latency'] = all_latencies df['Throughput'] = all_thruputs print("cleaned df:\n", df) # print(df) # print(df.loc[:11]) # print(df.loc[10:]) # for row in df.iterrows(): # print(row) # import sys; sys.exit() # make stuff log scale # if 'd' in df: # df['d'] = np.log2(df['d']).astype(np.int32) df['Log10(MSE)'] = np.log10(1. - df['R-Squared'] + 1e-10) df = df.sort_values('Method', axis=0) return df def make_cifar_fig(x_metric='d', y_metric='Accuracy'): # fig, axes = plt.subplots(2, 1, figsize=(6, 9), sharex=True) fig, axes = plt.subplots(2, 1, figsize=(11, 13.5), sharex=True) df10 = pd.read_csv(RESULTS_DIR / 'cifar10.csv') df100 = pd.read_csv(RESULTS_DIR / 'cifar100.csv') # dfs = (df10, df100) # for df in dfs: df10 = df10.loc[~(df10['ncodebooks'] < 4)] df100 = df100.loc[~(df100['ncodebooks'] < 4)] # if x_metric in ('Latency', 'Throughput'): # # TODO get results for PQ + Bolt # # df10 = df10.loc[~df10['method'].isin(['PQ', 'Bolt'])] # # include_methods = ('Bolt+MultiSplits', 'Bolt', 'Exact') # include_methods = ['Bolt+MultiSplits', 'Bolt', 'Exact'] # include_methods += 'PQ SVD FD-AMM CooccurSketch'.split() # TODO rm # # print("uniq methods: ", df10['method'].unique()) # # df10 = df10.loc[~df10['method'].isin(['PQ'])] # df10 = df10.loc[df10['method'].isin(include_methods)] # # df100 = df100.loc[~df100['method'].isin(['PQ', 'Bolt'])] # # df100 = df100.loc[~df100['method'].isin(['PQ'])] # df100 = df100.loc[df100['method'].isin(include_methods)] df10 = _clean_results_df(df10, default_D=512) df100 = _clean_results_df(df100, default_D=512) def lineplot(data, ax): # order = 'Ours Bolt Exact PQ SVD FD-AMM CD'.split() # order = [m for m in order if m in data['Method'].unique()] order = list(data['Method'].unique()) move_methods_to_front = ['Ours', 'OursPQ', 'Brute Force'] for elem in move_methods_to_front[:]: if elem in order: order.remove(elem) else: move_methods_to_front.remove(elem) order = move_methods_to_front + order # order = None # print("uniq methods:\n", data['Method'].unique()) # print("using order:\n", order) # cmap = plt.get_cmap('tab10') # palette = {'Ours': 'red', 'Bolt': cmap(0), 'Exact': cmap(1), # 'PQ': cmap(2), 'SVD': cmap(4), 'FD-AMM': cmap(5), # 'CD': cmap(6)} palette = None # have to specify markers or seaborn freaks out because it doesn't # have enough of them filled_markers = ('o', 'v', '^', '<', '>', '8', 's', 'p', '*', 'h', 'H', 'D', 'd', 'P', 'X') sb.lineplot(data=data, x=x_metric, y=y_metric, hue='Method', style='Method', style_order=order, hue_order=order, # markers=True, dashes=False, ax=ax, palette=palette) markers=filled_markers, dashes=False, ax=ax, palette=palette) # palette='tab10') lineplot(df10, axes[0]) lineplot(df100, axes[1]) # plt.suptitle('Sketch size vs Classification Accuracy') xlbl = _xlabel_for_xmetric(x_metric) # plt.suptitle('{} vs {}'.format(xlbl, y_metric)) plt.suptitle('Approximating Softmax Layers') axes[0].set_title('CIFAR-10') for ax in axes: ax.set_ylabel(y_metric) axes[0].set_xlabel(None) axes[1].set_xlabel(xlbl) axes[1].set_title('CIFAR-100') handles, labels = axes[0].get_legend_handles_labels() handles, labels = handles[1:], labels[1:] # rm 'Method' title axes[0].legend(handles, labels, fontsize='small') # axes[1].legend(handles, labels, fontsize='small') # plt.figlegend(handles, labels, loc='lower center', ncol=1) # plt.figlegend(handles, labels, loc='center right', ncol=1) axes[1].get_legend().remove() # axes[1].get_legend().remove() if x_metric in ('muls', 'ops', 'nlookups', 'Latency', 'Throughput'): axes[0].semilogx() plt.tight_layout() # plt.subplots_adjust(top=.92, bottom=.2) plt.subplots_adjust(top=.92, bottom=.22) save_fig('cifar_{}_{}'.format(x_metric, y_metric)) # def make_ecg_fig(y_metric='R-Squared'): def make_ecg_fig(x_metric='d'): fig, axes = plt.subplots(2, 1, figsize=(6, 9)) df = pd.read_csv(RESULTS_DIR / 'ecg.csv') df = _clean_results_df(df, default_D=24) # D = 24 # if 'd' in df: # mask = df['d'].isna() # df.loc[mask, 'd'] = D # df['d'] = np.log2(df['d']) # df.rename({'method': 'Method', 'acc_amm': 'Accuracy', # 'r_sq': 'R-Squared', 'nmultiplies': 'muls'}, # axis=1, inplace=True) # df['Log10(MSE)'] = np.log10(1. - df['R-Squared'] + 1e-10) # avoid log10(0) # df['muls'] = df['muls'].fillna(0) # df['nlookups'] = df['nlookups'].fillna(0) # # mask = ~df['nlookups'].isna() # # print("mask: ", mask) # # print('muls, nlookups') # # print(df[['muls', 'nlookups']]) # # add_to_muls = df['nlookups'].loc[mask] # equivalent_muls = df['muls'].add(df['nlookups']) # # df['muls'] = equivalent_muls # df['muls'] = equivalent_muls # # import sys; sys.exit() # df['muls'] = np.log10(df['muls']) df['Compression Ratio'] = df['nbytes_orig'] / df['nbytes_blosc_byteshuf'] def lineplot(data, ycol, ax): sb.lineplot(data=data, hue='Method', x=x_metric, y=ycol, style='Method', markers=True, dashes=False, ax=ax) lineplot(df, ycol='R-Squared', ax=axes[0]) lineplot(df, ycol='Compression Ratio', ax=axes[1]) xlbl = _xlabel_for_xmetric(x_metric) axes[0].set_title('ECG: {} vs R-Squared'.format(xlbl)) axes[1].set_title('ECG: {} vs Compression Ratio'.format(xlbl)) axes[0].set_ylim([0, 1]) axes[0].set_ylabel('R-Squared') axes[1].set_ylabel('Compression Ratio') axes[1].set_xlabel(xlbl) if x_metric in ('muls', 'ops', 'nlookups'): axes[0].semilogx() # axes[0].semilogx() plt.tight_layout() plt.subplots_adjust(top=.92, bottom=.2) save_fig('ecg_{}'.format(x_metric)) def make_caltech_fig(x_metric='d'): """x_metric should be in {'d', 'secs', 'muls'}""" fig, ax = plt.subplots(1, 1, figsize=(6, 6)) df = pd.read_csv(RESULTS_DIR / 'caltech.csv') df = _clean_results_df(df, default_D=27) sb.lineplot(data=df, hue='Method', x=x_metric, y='Log10(MSE)', style='Method', markers=True, dashes=False, ax=ax) ax.set_ylabel('Log10(MSE + 1e-10)') if x_metric == 'd': ax.set_title('Caltech: Sketch Size vs Log Squared Error') ax.set_xlabel('Log2(Sketch Size)') elif x_metric == 'secs': ax.set_title('Caltech: Computation Time vs Log Squared Error') ax.set_xlabel('Time (s)') elif x_metric == 'muls': ax.set_title('Caltech: # of Multiplies vs Log Squared Error') ax.set_xlabel('Log10(# of Multiplies)') plt.tight_layout() plt.subplots_adjust(top=.92, bottom=.2) save_fig('caltech_{}'.format(x_metric)) def main(): # for x_metric in 'd secs muls'.split(): # for x_metric in ['muls']: # for y_metric in ('Accuracy', 'R-Squared'): # make_cifar_fig(x_metric, y_metric) # make_cifar_fig('d', 'Accuracy') # make_cifar_fig('muls', 'Accuracy') make_cifar_fig('ops', 'Accuracy') make_cifar_fig('Latency', 'Accuracy') make_cifar_fig('Throughput', 'Accuracy') # make_cifar_fig('Accuracy') # make_cifar_fig('Accuracy') # make_cifar_fig('R-Squared') # make_ecg_fig(x_metric='d') # make_ecg_fig(x_metric='secs') # make_ecg_fig(x_metric='muls') # make_caltech_fig(x_metric='d') # make_caltech_fig(x_metric='secs') # make_caltech_fig(x_metric='muls') print("done") if __name__ == '__main__': main()
#!/usr/bin/env python import time import numpy as np from .utils import kmeans, orthonormalize_rows, random_rotation from joblib import Memory _memory = Memory('.', verbose=0) # ================================================================ PQ @_memory.cache def learn_pq(X, ncentroids, nsubvects, subvect_len, max_kmeans_iters=16): codebooks = np.empty((ncentroids, nsubvects, subvect_len)) assignments = np.empty((X.shape[0], nsubvects), dtype=np.int) # print "codebooks shape: ", codebooks.shape for i in range(nsubvects): start_col = i * subvect_len end_col = start_col + subvect_len X_in = X[:, start_col:end_col] centroids, labels = kmeans(X_in, ncentroids, max_iter=max_kmeans_iters) codebooks[:, i, :] = centroids assignments[:, i] = labels return codebooks, assignments # [2**nbits x M x D/M], [N x M] def reconstruct_X_pq(assignments, codebooks): """assignments: N x M ints; codebooks: 2**nbits x M x D/M floats""" _, M = assignments.shape subvect_len = codebooks.shape[2] assert assignments.shape[1] == codebooks.shape[1] D = M * subvect_len pointsCount = assignments.shape[0] points = np.zeros((pointsCount, D), dtype=np.float32) for i in range(M): subspace_start = subvect_len * i subspace_end = subspace_start + subvect_len subspace_codes = assignments[:, i] points[:, subspace_start:subspace_end] = codebooks[subspace_codes, i, :] return points def _dists_elemwise_sq(x, q): diffs = x - q return diffs * diffs def _dists_elemwise_l1(x, q): return np.abs(x - q) def _encode_X_pq(X, codebooks, elemwise_dist_func=_dists_elemwise_sq): ncentroids, nsubvects, subvect_len = codebooks.shape assert X.shape[1] == (nsubvects * subvect_len) idxs = np.empty((X.shape[0], nsubvects), dtype=np.int) X = X.reshape((X.shape[0], nsubvects, subvect_len)) for i, row in enumerate(X): row = row.reshape((1, nsubvects, subvect_len)) dists = elemwise_dist_func(codebooks, row) dists = np.sum(dists, axis=2) idxs[i, :] = np.argmin(dists, axis=0) return idxs # [N x nsubvects] def compute_reconstruction_error(X, X_hat, subvect_len=-1): diffs = X - X_hat diffs_sq = diffs * diffs if subvect_len > 0: errs = [] for i in range(0, diffs_sq.shape[1], subvect_len): errs_block = diffs_sq[:, i:i+subvect_len] errs.append(np.mean(errs_block)) print(" errors in each block: {} ({})".format( np.array(errs), np.sum(errs))) X_bar = X - np.mean(X, axis=0) col_sses = np.sum(X_bar * X_bar, axis=0) + 1e-14 tot_sse_using_mean = np.sum(col_sses) errors = np.mean(diffs_sq, axis=1) # variances = np.var(X, axis=1) # return np.mean(errors) / np.mean(variances) return np.mean(errors) / (tot_sse_using_mean / X_bar.size) # ================================================================ Gaussian OPQ # https://github.com/yahoo/lopq/blob/master/python/lopq/model.py; see # https://github.com/yahoo/lopq/blob/master/LICENSE. For this function only: # # Copyright 2015, Yahoo Inc. # Licensed under the terms of the Apache License, Version 2.0. # See the LICENSE file associated with the project for terms. # @_memory.cache def eigenvalue_allocation(num_buckets, eigenvalues, shuffle=False): """ Compute a permutation of eigenvalues to balance variance accross buckets of dimensions. Described in section 3.2.4 in http://research.microsoft.com/pubs/187499/cvpr13opq.pdf Note, the following slides indicate this function will break when fed eigenvalues < 1 without the scaling trick implemented below: https://www.robots.ox.ac.uk/~vgg/rg/slides/ge__cvpr2013__optimizedpq.pdf :param int num_buckets: the number of dimension buckets over which to allocate eigenvalues :param ndarray eigenvalues: a vector of eigenvalues :param bool shuffle: whether to randomly shuffle the order of resulting buckets :returns ndarray: a vector of indices by which to permute the eigenvectors """ D = len(eigenvalues) dims_per_bucket = D // num_buckets eigenvalue_product = np.zeros(num_buckets, dtype=float) bucket_size = np.zeros(num_buckets, dtype=int) permutation = np.zeros((num_buckets, dims_per_bucket), dtype=int) # We first must scale the eigenvalues by dividing by their # smallets non-zero value to avoid problems with the algorithm # when eigenvalues are less than 1. min_non_zero_eigenvalue = np.min(np.abs(eigenvalues[np.nonzero(eigenvalues)])) eigenvalues = eigenvalues / min_non_zero_eigenvalue # this is not actually a requirement, but I'm curious about whether this # condition is ever violated if not np.all(eigenvalues > 0): print("WARNING: some eigenvalues were nonpositive") # Iterate eigenvalues in descending order sorted_inds = np.argsort(eigenvalues)[::-1] log_eigs = np.log2(abs(eigenvalues)) for ind in sorted_inds: # Find eligible (not full) buckets eligible = (bucket_size < dims_per_bucket).nonzero() # Find eligible bucket with least eigenvalue product i = eigenvalue_product[eligible].argmin(0) bucket = eligible[0][i] # Update eigenvalue product for this bucket eigenvalue_product[bucket] = eigenvalue_product[bucket] + log_eigs[ind] # Store bucket assignment and update size permutation[bucket, bucket_size[bucket]] = ind bucket_size[bucket] += 1 if shuffle: shuffle_idxs = np.arange(num_buckets, dtype=np.int) np.random.shuffle(shuffle_idxs) permutation = permutation[shuffle_idxs] # wow, these are within <1% of each other # print "opq eigenvalue log prods: ", eigenvalue_product return np.reshape(permutation, D) def learn_opq_gaussian_rotation(X_train, ncodebooks, shuffle=False): means = np.mean(X_train, axis=0) cov = np.dot(X_train.T, X_train) - np.outer(means, means) eigenvals, eigenvects = np.linalg.eigh(cov) order_idxs = eigenvalue_allocation(ncodebooks, eigenvals, shuffle=shuffle) assert len(order_idxs) == X_train.shape[1] return eigenvects[:, order_idxs].T # rows are projections # ================================================================ OPQ def _update_centroids_opq(X, assignments, ncentroids): # [N x D], [N x M] nsubvects = assignments.shape[1] subvect_len = X.shape[1] // nsubvects assert X.shape[0] == assignments.shape[0] assert X.shape[1] % nsubvects == 0 codebooks = np.zeros((ncentroids, nsubvects, subvect_len), dtype=np.float32) for i, row in enumerate(X): for m in range(nsubvects): start_col = m * subvect_len end_col = start_col + subvect_len codebooks[assignments[i, m], m, :] += row[start_col:end_col] for m in range(nsubvects): code_counts = np.bincount(assignments[:, m], minlength=ncentroids) codebooks[:, m] /= np.maximum(code_counts, 1).reshape((-1, 1)) # no div by 0 return codebooks class NumericalException(Exception): pass def _debug_rotation(R): D = np.max(R.shape) identity = np.identity(D, dtype=np.float32) RtR = np.dot(R.T, R) R_det = np.linalg.det(RtR) print("determinant of R*R: ", R_det) R_trace = np.trace(RtR) print("trace of R*R, trace divided by D: {}, {}".format(R_trace, R_trace / D)) off_diagonal_abs_mean = np.mean(np.abs(RtR - identity)) print("mean(abs(off diagonals of R*R)): ", off_diagonal_abs_mean) if R_det < .999 or R_det > 1.001: raise NumericalException("Bad determinant") if R_trace < .999 * D or R_trace > 1.001 * D: raise NumericalException("Bad trace") if off_diagonal_abs_mean > .001: raise NumericalException("Bad off-diagonals") def opq_rotate(X, R): # so other code need not know what to transpose return np.dot(np.atleast_2d(X), R.T) def opq_undo_rotate(X, R): # so other code need not know what to transpose return np.dot(np.atleast_2d(X), R) # @_memory.cache def opq_initialize(X_train, ncodebooks, init='gauss'): X = X_train _, D = X.shape if init == 'gauss' or init == 'gauss_flat' or init == 'gauss_shuffle': permute = (init == 'gauss_shuffle') R = learn_opq_gaussian_rotation(X_train, ncodebooks, shuffle=permute) R = R.astype(np.float32) if init == 'gauss_flat': # assert R.shape[0] == R.shape[1] D = R.shape[1] d = D // ncodebooks assert d * ncodebooks == D # same # of dims in each subspace local_r = random_rotation(int(d)) tiled = np.zeros((D, D)) for c in range(ncodebooks): start = c * d end = start + d tiled[start:end, start:end] = local_r R = np.dot(R, tiled) X_rotated = opq_rotate(X, R) elif init == 'identity': R = np.identity(D, dtype=np.float32) # D x D X_rotated = X elif init == 'random': R = np.random.randn(D, D).astype(np.float32) R = orthonormalize_rows(R) X_rotated = opq_rotate(X, R) else: raise ValueError("Unrecognized initialization method: ".format(init)) return X_rotated, R # loosely based on: # https://github.com/arbabenko/Quantizations/blob/master/opqCoding.py @_memory.cache def learn_opq(X_train, ncodebooks, codebook_bits=8, niters=10, initial_kmeans_iters=1, init='gauss', debug=False): """init in {'gauss', 'identity', 'random'}""" print("OPQ: Using init '{}'".format(init)) t0 = time.time() X = X_train.astype(np.float32) N, D = X.shape ncentroids = int(2**codebook_bits) subvect_len = D // ncodebooks assert D % subvect_len == 0 # equal number of dims for each codebook X_rotated, R = opq_initialize(X_train, ncodebooks=ncodebooks, init=init) # initialize codebooks by running kmeans on each rotated dim; this way, # setting niters=0 corresponds to normal PQ codebooks, assignments = learn_pq(X_rotated, ncentroids=ncentroids, nsubvects=ncodebooks, subvect_len=subvect_len, max_kmeans_iters=1) for it in np.arange(niters): # compute reconstruction errors X_hat = reconstruct_X_pq(assignments, codebooks) # err = compute_reconstruction_error(X_rotated, X_hat, subvect_len=subvect_len) err = compute_reconstruction_error(X_rotated, X_hat) print("---- OPQ {}x{}b iter {}: mse / variance = {:.5f}".format( ncodebooks, codebook_bits, it, err)) # update rotation matrix based on reconstruction errors U, s, V = np.linalg.svd(np.dot(X_hat.T, X), full_matrices=False) R = np.dot(U, V) # update centroids using new rotation matrix X_rotated = opq_rotate(X, R) assignments = _encode_X_pq(X_rotated, codebooks) codebooks = _update_centroids_opq(X_rotated, assignments, ncentroids) X_hat = reconstruct_X_pq(assignments, codebooks) err = compute_reconstruction_error(X_rotated, X_hat) t = time.time() - t0 print("---- OPQ {}x{}b final mse / variance = {:.5f} ({:.3f}s)".format( ncodebooks, codebook_bits, err, t)) return codebooks, assignments, R # ================================================================ Block OPQ def bopq_rotate(X, rotations): X = np.atleast_2d(X) _, D = X.shape R_sz = len(rotations[0]) nrots = int(D / R_sz) assert nrots == len(rotations) rot_starts = R_sz * np.arange(nrots) rot_ends = rot_starts + R_sz X_out = np.copy(X) for i, R in enumerate(rotations): start, end = rot_starts[i], rot_ends[i] X_out[:, start:end] = np.dot(X[:, start:end], R.T) return X_out @_memory.cache # opq with block diagonal rotations def learn_bopq(X_train, ncodebooks, codebook_bits=4, niters=20, initial_kmeans_iters=1, R_sz=16, **sink): t0 = time.time() X = X_train.astype(np.float32) N, D = X.shape ncentroids = int(2**codebook_bits) subvect_len = D // ncodebooks assert D % subvect_len == 0 # equal number of dims for each codebook # compute number of rotations and subspaces associated with each nrots = int(D / R_sz) rot_starts = R_sz * np.arange(nrots) rot_ends = rot_starts + R_sz # X_rotated, R = opq_initialize(X_train, ncodebooks=ncodebooks, init=init) X_rotated = X # hardcode identity init # TODO allow others rotations = [np.eye(R_sz) for i in range(nrots)] # initialize codebooks by running kmeans on each rotated dim; this way, # setting niters=0 corresponds to normal PQ codebooks, assignments = learn_pq(X_rotated, ncentroids=ncentroids, nsubvects=ncodebooks, subvect_len=subvect_len, max_kmeans_iters=1) for it in np.arange(niters): # compute reconstruction errors X_hat = reconstruct_X_pq(assignments, codebooks) # err = compute_reconstruction_error(X_rotated, X_hat, subvect_len=subvect_len) err = compute_reconstruction_error(X_rotated, X_hat) print("---- BOPQ {} {}x{}b iter {}: mse / variance = {:.5f}".format( R_sz, ncodebooks, codebook_bits, it, err)) rotations = [] for i in range(nrots): start, end = rot_starts[i], rot_ends[i] X_sub = X[:, start:end] X_hat_sub = X_hat[:, start:end] # update rotation matrix based on reconstruction errors U, s, V = np.linalg.svd(np.dot(X_hat_sub.T, X_sub), full_matrices=False) R = np.dot(U, V) rotations.append(R) X_rotated[:, start:end] = np.dot(X_sub, R.T) # update assignments and codebooks based on new rotations assignments = _encode_X_pq(X_rotated, codebooks) codebooks = _update_centroids_opq(X_rotated, assignments, ncentroids) X_hat = reconstruct_X_pq(assignments, codebooks) err = compute_reconstruction_error(X_rotated, X_hat) t = time.time() - t0 print("---- BOPQ {} {}x{}b final mse / variance = {:.5f} ({:.3f}s)".format( R_sz, ncodebooks, codebook_bits, err, t)) return codebooks, assignments, rotations
#!/bin/env/python import os import shutil def ls(dir='.'): return os.listdir(dir) def is_hidden(path): return os.path.basename(path).startswith('.') def is_visible(path): return not is_hidden(path) def join_paths(dir, contents): return [os.path.join(dir, f) for f in contents] def files_matching(dir, prefix=None, suffix=None, abs_paths=False, only_files=False, only_dirs=False, recursive=False, only_visible=False, only_hidden=False): files = os.listdir(dir) if recursive: abs_dir = dir paths = join_paths(abs_dir, files) for path in paths: if not os.path.isdir(path): continue matches = files_matching( path, prefix=prefix, suffix=suffix, abs_paths=abs_paths, only_files=only_files, only_dirs=only_dirs, recursive=True) matches = join_paths(path, matches) matches = [os.path.relpath(m, start=dir) for m in matches] files += matches if prefix: files = [f for f in files if f.startswith(prefix)] if suffix: files = [f for f in files if f.endswith(suffix)] if only_files or only_dirs: paths = join_paths(dir, files) if only_files: files = [f for f, p in zip(files, paths) if os.path.isfile(p)] if only_dirs: files = [f for f, p in zip(files, paths) if os.path.isdir(p)] if abs_paths: files = join_paths(os.path.abspath(dir), files) if only_visible: files = [f for f in files if is_visible(f)] if only_hidden: files = [f for f in files if is_hidden(f)] return sorted(files) def list_subdirs(dir, startswith=None, endswith=None, abs_paths=False, recursive=False, only_visible=False): return files_matching(dir, startswith, endswith, abs_paths, only_dirs=True, recursive=recursive, only_visible=only_visible) def list_files(dir, startswith=None, endswith=None, abs_paths=False, recursive=False, only_visible=False): return files_matching(dir, startswith, endswith, abs_paths, only_files=True, recursive=recursive, only_visible=only_visible) def remove(path): if os.path.exists(path): try: os.remove(path) except (OSError): shutil.rmtree(path) def force_create_dir(dir): if os.path.exists(dir): remove(dir) os.makedirs(dir) def ensure_dir_exists(dir_or_file): if '.' in os.path.basename(dir_or_file): # this looks like a file dirname = os.path.dirname(dir_or_file) else: dirname = dir_or_file if not os.path.exists(dirname): os.makedirs(dirname) def basename(f, noext=False): name = os.path.basename(f) if noext: name = name.split('.')[0] return name
#!/bin/env python from __future__ import absolute_import, division, print_function import numpy as np import os import PIL from PIL import Image from PIL import ImageOps # can't just do PIL.ImageOps for some reason from . import files # ================================ TODO rm duplicate code from imagenet.py # adapted from https://github.com/keras-team/keras-preprocessing/blob/master/ # keras_preprocessing/image/utils.py under MIT license def img_to_array(img, layout='nhwc', dtype='float32', mode='RGB'): """Converts a PIL Image instance to a Numpy array. # Arguments img: PIL Image instance. layout: Image data format, either "nchw" or "nhwc". dtype: Dtype to use for the returned array. # Returns A 3D Numpy array. # Raises ValueError: if invalid `img` or `layout` is passed. """ # print("img info:", img.format, img.size, img.mode) # if img.mode == 'L': if img.mode != mode: img = img.convert(mode=mode) if layout not in ('nchw', 'nhwc'): raise ValueError('Unknown layout: %s' % layout) # Numpy array x has format (height, width, channel) # or (channel, height, width) # but original PIL image has format (width, height, channel) x = np.asarray(img, dtype=dtype) if len(x.shape) == 3: if layout == 'nchw': x = x.transpose(2, 0, 1) elif len(x.shape) == 2: # print("x is only rank 2...WTF!?") if layout == 'nchw': x = x.reshape((1, x.shape[0], x.shape[1])) else: x = x.reshape((x.shape[0], x.shape[1], 1)) else: raise ValueError('Unsupported image shape: %s' % (x.shape,)) return x def resize_img(img, ratio_or_size): if ratio_or_size is None or np.max(ratio_or_size) < 0: return img try: nrows = ratio_or_size[0] ncols = ratio_or_size[1] nrows = img.height if nrows < 0 else nrows ncols = img.width if ncols < 0 else ncols except AttributeError: nrows = img.height * ratio_or_size ncols = img.width * ratio_or_size new_size = (nrows, ncols) is_downsampling = (nrows < img.height) or (ncols < img.width) interp = PIL.Image.LANCZOS if is_downsampling else PIL.Image.BICUBIC return img.resize(new_size, resample=interp) def crop_img(img, crop_how=None, new_size=(224, 224), resize_shorter_to=256): if crop_how is None: return img assert crop_how in ('center', 'square') height, width = img.height, img.width if (height == width) and (new_size is None): return img if crop_how == 'center': return center_crop(img, new_size=new_size, resize_shorter_to=resize_shorter_to) if new_size is None: new_width = min(width, height) new_height = new_width else: new_height, new_width = new_size assert new_width <= width assert new_height <= height left = (width - new_width) // 2 top = (height - new_height) // 2 # right = (width + new_width) // 2 # bottom = (height + new_height) // 2 right = left + new_width bottom = top + new_height return img.crop((left, top, right, bottom)) def center_crop(img, new_size=(224, 224), resize_shorter_to=256): height, width = img.height, img.width minsize = min(height, width) new_height = (height * resize_shorter_to) // minsize new_width = (width * resize_shorter_to) // minsize img = img.resize( (new_width, new_height), resample=Image.BICUBIC) assert min(new_width, new_height) == resize_shorter_to return crop_img(img, crop_how='square', new_size=new_size) def pad_img(img, pad_how='square', fill_value=0): if pad_how is None: return img assert pad_how == 'square' # no other kinds of cropping supported height, width = img.height, img.width if height == width: return img new_size = max(height, width) delta_w = new_size - width pad_left = delta_w // 2 pad_right = delta_w - pad_left delta_h = new_size - height pad_top = delta_h // 2 pad_bottom = delta_h - pad_top padding = pad_left, pad_top, pad_right, pad_bottom return ImageOps.expand(img, border=padding, fill=fill_value) def load_jpg(path, layout='nhwc', dtype=None, resample=None, crop=None, pad=None): img = PIL.Image.open(path) img = pad_img(img, pad) img = crop_img(img, crop) img = resize_img(img, ratio_or_size=resample) return img_to_array(img, layout=layout, dtype=dtype) # assumes one subdir for each class, with class name equal to subdir name # @_memory.cache def load_jpegs_from_dir(dirpath, remove_classes=None, require_suffix=None, layout='nhwc', dtype=None, resample=(224, 224), crop='center', pad=None, verbose=1, limit_per_class=None, only_return_path=False): subdirs = sorted(files.list_subdirs(dirpath, only_visible=True)) if remove_classes is not None: if isinstance(remove_classes, str): remove_classes = [remove_classes] for classname in remove_classes: subdirs.remove(classname) if verbose > 0: print("found {} classes in directory: {}".format(len(subdirs), dirpath)) classname_to_label = {name: i for i, name in enumerate(subdirs)} # noqa label_to_classname = {i: name for name, i in classname_to_label.items()} all_imgs = [] all_labels = [] for subdir in subdirs: subdir_path = os.path.join(dirpath, subdir) img_paths = files.list_files( subdir_path, endswith=require_suffix, abs_paths=True, only_visible=True) if limit_per_class is not None and limit_per_class > 0: img_paths = img_paths[:limit_per_class] if verbose > 1: print("loading {:4d} images for class '{}'".format( len(img_paths), subdir)) # not certain += [...] version was working label = classname_to_label[subdir] for i in range(len(img_paths)): all_labels.append(label) # all_labels += [] * len(img_paths) if only_return_path: imgs = img_paths else: imgs = [load_jpg(f, layout=layout, dtype=dtype, resample=resample, crop=crop, pad=pad)[np.newaxis, :, :, :] for f in img_paths] all_imgs += imgs if only_return_path: X = all_imgs else: try: # this works iff resampled/padded/cropped to same size X = np.concatenate(all_imgs, axis=0) except ValueError: # otherwise strip batch dim (so each image is 3D) X = [img.reshape(img.shape[1:]) for img in all_imgs] y = np.array(all_labels, dtype=np.int32) return (X, y), label_to_classname
#!/bin/env python from __future__ import print_function import numpy as np import os import warnings import h5py from sklearn.datasets import load_digits import keras from keras.preprocessing import image # from python import imagenet, svhn, caltech # from python.datasets import caltech from . import imagenet from . import svhn from .data_utils import stratified_split_train_test from joblib import Memory _memory = Memory('.', verbose=1) # DATA_DIR = os.path.expanduser('~/Desktop/datasets/nn-search') DATA_DIR = os.path.expanduser('data') join = os.path.join DEFAULT_AUG_KWARGS = { 'shear_range': 0.2, 'zoom_range': 0.2, 'horizontal_flip': True } class LabeledDataset(object): __slots__ = 'name X_train y_train X_test y_test _collection'.split() def __init__(self, name, X_train, y_train, X_test=None, y_test=None): self.name = name self.X_train = X_train self.y_train = y_train self.X_test = X_test self.y_test = y_test def generators(self, batch_size, augment=True, preprocessing_function=None, aug_kwargs=None): _aug_kwargs = DEFAULT_AUG_KWARGS if aug_kwargs is not None: _aug_kwargs.update(aug_kwargs) if not augment: _aug_kwargs = {} nclasses = len(np.unique(self.y_train)) y_train = keras.utils.to_categorical(self.y_train, num_classes=nclasses) y_test = keras.utils.to_categorical(self.y_test, num_classes=nclasses) train_datagen = image.ImageDataGenerator( preprocessing_function=preprocessing_function, **_aug_kwargs) train_generator = train_datagen.flow( self.X_train, y_train, batch_size=batch_size) test_datagen = image.ImageDataGenerator( preprocessing_function=preprocessing_function) test_generator = test_datagen.flow( self.X_test, y_test, batch_size=batch_size) return train_generator, test_generator class HugeLabeledDataset(object): def __init__(self, name, train_dir, test_dir, train_nsamples=None, test_nsamples=None): self.name = name # self.train_dir = os.path.abspath(train_dir) # self.test_dir = os.path.abspath(test_dir) self.train_dir = train_dir self.test_dir = test_dir self.train_nsamples = int(train_nsamples or -1) self.test_nsamples = int(test_nsamples or -1) def generators(self, batch_size=None, augment=True, preprocessing_function=None, aug_kwargs=None, train_batch_size=None, test_batch_size=None, **flow_kwargs): _aug_kwargs = DEFAULT_AUG_KWARGS if aug_kwargs is not None: _aug_kwargs.update(aug_kwargs) if not augment: _aug_kwargs = {} flow_kwargs = flow_kwargs or {} flow_kwargs.setdefault('target_size', (224, 224)) flow_kwargs.setdefault('class_mode', 'categorical') train_generator = None test_generator = None if self.train_dir: train_batch_size = int(train_batch_size or batch_size) flow_kwargs['batch_size'] = train_batch_size print("HugeLabeledDataset: creating flow from train dir: ", self.train_dir) train_datagen = image.ImageDataGenerator( preprocessing_function=preprocessing_function, **_aug_kwargs) train_generator = train_datagen.flow_from_directory( self.train_dir, **flow_kwargs) if self.test_dir: test_batch_size = int(test_batch_size or batch_size) flow_kwargs['batch_size'] = test_batch_size print("HugeLabeledDataset: creating flow from test dir: ", self.test_dir) test_datagen = image.ImageDataGenerator( preprocessing_function=preprocessing_function) test_generator = test_datagen.flow_from_directory( self.test_dir, **flow_kwargs) return train_generator, test_generator class Random: UNIFORM = 'uniform' GAUSS = 'gauss' WALK = 'walk' BLOBS = 'blobs' DIGITS = 'Digits' MNIST = 'MNIST' FASHION_MNIST = 'FashionMNIST' CIFAR10 = 'Cifar10' CIFAR100 = 'Cifar100' SVHN = 'SVHN' CALTECH101 = 'Caltech101' CALTECH256 = 'Caltech256' CUB200 = 'CUB200' FLOWERS102 = 'Flowers102' INDOOR67 = 'Indoor67' IMAGENET_TINY = 'TinyImagenet' # 64x64, 200? classes IMAGENET_10_CLASSES = 'ImageNet-10-Classes' # full res, 10cls, 1k/cls IMAGENET_100_CLASSES = 'ImageNet-100-Classes' # full res, 100cls, 1k/cls IMAGENET_1_EXAMPLE = 'ImageNet-1-Example' # full res, 1k cls, 1/cls IMAGENET_10_EXAMPLES = 'ImageNet-10-Examples' # full res, 1k cls, 10/cls IMAGENET_25_EXAMPLES = 'ImageNet-25-Examples' # full res, 1k cls, 25/cls IMAGENET_50_EXAMPLES = 'ImageNet-50-Examples' # full res, 1k cls, 50/cls IMAGENET_100_EXAMPLES = 'ImageNet-100-Examples' # full res, 1k cls, 100/cls IMAGENET_64PX = 'ImageNet64' # 64x64, all examples IMAGENET = 'ImageNet' IMAGENET_ONE_OF_EACH = 'ImagenetOneOfEach' MINIPLACES = 'Miniplaces' ALL_IMAGENET_DATASETS = [ IMAGENET, IMAGENET_64PX, IMAGENET_TINY, IMAGENET_ONE_OF_EACH, IMAGENET_10_CLASSES, IMAGENET_100_CLASSES, IMAGENET_1_EXAMPLE, IMAGENET_10_EXAMPLES, IMAGENET_100_EXAMPLES] ALL_KERAS_DATASETS = [MNIST, CIFAR10, CIFAR100, FASHION_MNIST] def _load_file(fname, *args, **kwargs): fname = os.path.join(DATA_DIR, fname) print("trying to load file at path: {}".format(fname)) if fname.split('.')[-1] == 'txt': return np.loadtxt(fname, *args, **kwargs) return np.load(fname, *args, **kwargs) def _load_digits_X_y(ntrain=1000): X, y = load_digits(return_X_y=True) X_train, X_test = X[:ntrain], X[ntrain:] y_train, y_test = y[:ntrain], y[ntrain:] return LabeledDataset('Digits', X_train, y_train, X_test, y_test) # return X[:-nqueries], X[-nqueries:] # X, Q def _load_keras_dset(which_dataset): from keras import datasets as kd dataClass = {CIFAR10: kd.cifar10, CIFAR100: kd.cifar100, MNIST: kd.mnist, FASHION_MNIST: kd.fashion_mnist}[which_dataset] (X_train, y_train), (X_test, y_test) = dataClass.load_data() pretty_name = str(which_dataset).split('.')[-1].split("'")[0] return LabeledDataset(pretty_name, X_train, y_train, X_test, y_test) def load_imagenet_64(limit_ntrain=-1): # if we're not going to use the whole training set, don't even load in all # the files it's split into (necessary unless you have >18GB of free RAM) which_file_idxs = None if limit_ntrain > 0: nchunks = int(np.ceil( limit_ntrain / imagenet.IMAGENET_64_TRAIN_CHUNK_NSAMPLES)) which_file_idxs = np.arange(1, nchunks + 1) X_train, y_train = imagenet.load_train_data_64x64( which_file_idxs=which_file_idxs) X_test, y_test = imagenet.load_test_data_64x64() return LabeledDataset(IMAGENET_64PX, X_train, y_train, X_test, y_test, train_nsamples=1e6) def load_imagenet_tiny(): X_train, y_train = imagenet.load_train_data_tiny() X_test, y_test = imagenet.load_test_data_tiny() return LabeledDataset(IMAGENET_TINY, X_train, y_train, X_test, y_test) def load_imagenet_one_of_each(): X, y = imagenet.load_data_one_of_each() return LabeledDataset(IMAGENET_ONE_OF_EACH, X, y, X, y, train_nsamples=1e3) def load_imagenet(load_train=True, load_val=True): train_path = imagenet.IMAGENET_TRAIN_PATH if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset( IMAGENET, train_path, test_path, train_nsamples=1281167, test_nsamples=50e3) def load_imagenet_10_classes(load_train=True, load_val=True): train_path = imagenet.IMAGENET_10_CLASSES_TRAIN_PATH if load_train else None test_path = imagenet.IMAGENET_10_CLASSES_TEST_PATH if load_val else None return HugeLabeledDataset( IMAGENET_10_CLASSES, train_path, test_path, train_nsamples=13000, test_nsamples=500) def load_imagenet_100_classes(load_train=True, load_val=True): train_path = imagenet.IMAGENET_100_CLASSES_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_100_CLASSES_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_100_CLASSES, train_path, test_path, train_nsamples=129395, test_nsamples=5000) def load_imagenet_1_example(load_train=True, load_val=True): train_path = imagenet.IMAGENET_1_EXAMPLE_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_10_EXAMPLES, train_path, test_path, train_nsamples=1e3, test_nsamples=50e3) def load_imagenet_10_examples(load_train=True, load_val=True): train_path = imagenet.IMAGENET_10_EXAMPLES_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_10_EXAMPLES, train_path, test_path, train_nsamples=10e3, test_nsamples=50e3) def load_imagenet_25_examples(load_train=True, load_val=True): train_path = imagenet.IMAGENET_25_EXAMPLES_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_25_EXAMPLES, train_path, test_path, train_nsamples=25e3, test_nsamples=50e3) def load_imagenet_50_examples(load_train=True, load_val=True): train_path = imagenet.IMAGENET_50_EXAMPLES_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_50_EXAMPLES, train_path, test_path, train_nsamples=50e3, test_nsamples=50e3) def load_imagenet_100_examples(load_train=True, load_val=True): train_path = imagenet.IMAGENET_100_EXAMPLES_TRAIN_PATH \ if load_train else None test_path = imagenet.IMAGENET_TEST_PATH if load_val else None return HugeLabeledDataset(IMAGENET_10_EXAMPLES, train_path, test_path, train_nsamples=100e3, test_nsamples=50e3) def _load_miniplaces(): path = '/data/ddmg/neuro/datasets/Miniplaces/miniplaces.h5' with h5py.File(path, 'r') as hf: X_train = hf['X_train'][()] Y_train = hf['Y_train'][()] X_val = hf['X_val'][()] Y_val = hf['Y_val'][()] return LabeledDataset(MINIPLACES, X_train, Y_train, X_val, Y_val) def _load_svhn(): (X_train, y_train), (X_test, y_test) = svhn.load_data() return LabeledDataset(SVHN, X_train, y_train, X_test, y_test) def load_caltech101(): data_dir = '../datasets/caltech/101_ObjectCategories' return HugeLabeledDataset(CALTECH101, data_dir, None) # (X, y), _ = caltech.load_caltech101() # return LabeledDataset(IMAGENET_ONE_OF_EACH, X, y, X, y) def load_caltech256(): data_dir = '../datasets/caltech/256_ObjectCategories' return HugeLabeledDataset(CALTECH256, data_dir, None) # (X, y), _ = caltech.load_caltech256() # return LabeledDataset(IMAGENET_ONE_OF_EACH, X, y, X, y) def load_flowers102(): data_dir = '../datasets/flowers102' train_dir = os.path.join(data_dir, 'train') test_dir = os.path.join(data_dir, 'test') return HugeLabeledDataset(FLOWERS102, train_dir, test_dir, train_nsamples=1020, test_nsamples=6149) def load_cub200(): # note that this is 2011 version of CUB200 data_dir = '../datasets/cub200' train_dir = os.path.join(data_dir, 'train') test_dir = os.path.join(data_dir, 'test') return HugeLabeledDataset(CUB200, train_dir, test_dir, train_nsamples=5994, test_nsamples=5794) def load_indoor67(): # this is the subset with predefined train vs test split data_dir = '../datasets/indoor67' train_dir = os.path.join(data_dir, 'train') test_dir = os.path.join(data_dir, 'test') return HugeLabeledDataset(INDOOR67, train_dir, test_dir, train_nsamples=(67 * 80), test_nsamples=(67 * 20)) # @_memory.cache def load_dataset(which_dataset, norm_mean=False, norm_len=False, flatten=False, Ntrain=-1, Ntest=-1, ensure_channels=False, test_frac=None, scale_to_0_1=False): if which_dataset == DIGITS: dset = _load_digits_X_y() elif which_dataset in ALL_KERAS_DATASETS: dset = _load_keras_dset(which_dataset) elif which_dataset == IMAGENET_64PX: dset = load_imagenet_64(limit_ntrain=Ntrain) elif which_dataset == IMAGENET_TINY: dset = load_imagenet_tiny() elif which_dataset == IMAGENET_ONE_OF_EACH: dset = load_imagenet_one_of_each() elif which_dataset == MINIPLACES: dset = _load_miniplaces() elif which_dataset == SVHN: dset = _load_svhn() elif which_dataset == CALTECH101: return load_caltech101() elif which_dataset == CALTECH256: return load_caltech256() elif which_dataset == CUB200: return load_cub200() elif which_dataset == FLOWERS102: return load_flowers102() elif which_dataset == IMAGENET: return load_imagenet() elif which_dataset == IMAGENET_10_CLASSES: return load_imagenet_10_classes() elif which_dataset == IMAGENET_100_CLASSES: return load_imagenet_100_classes() elif which_dataset == IMAGENET_1_EXAMPLE: return load_imagenet_1_example() elif which_dataset == IMAGENET_10_EXAMPLES: return load_imagenet_10_examples() elif which_dataset == IMAGENET_25_EXAMPLES: return load_imagenet_25_examples() elif which_dataset == IMAGENET_50_EXAMPLES: return load_imagenet_50_examples() elif which_dataset == IMAGENET_100_EXAMPLES: return load_imagenet_100_examples() else: raise ValueError("unrecognized dataset {}".format(which_dataset)) if isinstance(dset, HugeLabeledDataset): # only has flow_from_directory() generators; no postprocessing # possible, so go ahead and return immediately return dset train_is_test = (dset.X_train.base is dset.X_test) or \ (dset.X_test.base is dset.X_train) train_test_equal = np.array_equal(dset.X_train[:10], dset.X_test[:10]) train_test_same = train_is_test or train_test_equal if train_test_same: if test_frac is None: warnings.warn("WARNING: Training data is also the test data! " "Reversing order of test data. Consider passing " "test_frac > 0 to automatically perform a " "stratified train-test split.") dset.X_test = dset.X_test[::-1] else: X_train, X_test, y_train, y_test = stratified_split_train_test( dset.X_train, dset.y_train, train_frac=(1. - test_frac)) dset = LabeledDataset(dset.name, X_train, y_train, X_test, y_test) train_is_test = False train_test_equal = False train_test_same = False if train_is_test: dset.X_test = np.copy(dset.X_test) dset.y_test = np.copy(dset.y_test) train_is_test = False if flatten: dset.X_train = dset.X_train.reshape(dset.X_train.shape[0], -1) dset.X_test = dset.X_test.reshape(dset.X_test.shape[0], -1) dset.X_train = dset.X_train.astype(np.float32) dset.X_test = dset.X_test.astype(np.float32) X_train = dset.X_train X_test = dset.X_test if Ntrain > 0: dset.X_train = X_train[:Ntrain] dset.y_train = dset.y_train[:Ntrain] if Ntest > 0: dset.X_test = np.copy(X_test[:Ntest]) dset.y_test = np.copy(dset.y_test[:Ntest]) if scale_to_0_1: min_X = min(np.min(dset.X_train), np.min(dset.X_test)) max_X = max(np.max(dset.X_train), np.max(dset.X_test)) dset.X_train = (dset.X_train - min_X) / max_X # if not train_is_test: dset.X_test = (dset.X_test - min_X) / max_X if norm_mean: means = np.mean(dset.X_train, axis=0) dset.X_train -= means # if not train_is_test: # don't subtract means twice from same array dset.X_test -= means if norm_len: dset.X_train /= np.linalg.norm(dset.X_train, axis=1, keepdims=True) # if not train_is_test: # don't divide by norms twice on same array dset.X_test /= np.linalg.norm(dset.X_test, axis=1, keepdims=True) if ensure_channels: import keras.backend as K # don't import keras unless we need it if len(X_train.shape) == 3: # no channels; e.g., MNIST img_rows, img_cols = X_train.shape[-2], X_train.shape[-1] # K.set_image_data_format('channels_last') # for argmax layer if K.image_data_format() == 'channels_first': dset.X_train = dset.X_train.reshape( X_train.shape[0], 1, img_rows, img_cols) dset.X_test = dset.X_test.reshape( X_test.shape[0], 1, img_rows, img_cols) else: dset.X_train = dset.X_train.reshape( X_train.shape[0], img_rows, img_cols, 1) dset.X_test = dset.X_test.reshape( X_test.shape[0], img_rows, img_cols, 1) return dset # if D_multiple_of > 1: # X_train = ensure_num_cols_multiple_of(X_train, D_multiple_of) # X_test = ensure_num_cols_multiple_of(X_test, D_multiple_of) # Q = ensure_num_cols_multiple_of(Q, D_multiple_of) # return X_train, Q, X_test, true_nn
#!/bin/env python from __future__ import division, print_function import numpy as np # import pyedflib as edf # pip install pyedflib # import mne from . import paths from . import files ECG_DIR = paths.UCD_ECG NUM_RECORDINGS = 25 def main(): pass print("ecg dir: ", ECG_DIR) fpaths = files.list_files(ECG_DIR, abs_paths=True) # fpaths = files.list_files(ECG_DIR) assert len(fpaths) == NUM_RECORDINGS # print("fpaths: ", "\n".join(fpaths)) # print("number of fpaths: ", len(fpaths)) for path in fpaths: print("------------------------ ", path) # f = edf.EdfReader(path) # print(f.signals_in_file) magical_start_offset = 1025 # from looking at raw binary # raw = bytes(open(path, 'rb').read())[magical_start_offset:] with open(path, 'rb') as f: raw = f.read() # raw = open(path, 'rb').read() print("length of raw: ", len(raw)) print("type(raw)", type(raw)) a = np.frombuffer(raw, offset=magical_start_offset, dtype=np.uint16) # a = np.frombuffer(raw, dtype=np.uint16) print(len(a)) print(len(a) / 3) # print("number of bytes: ", len(raw)) # with open(path, 'rb') as f: # # f.seek(magical_start_offset) # f.read(magical_start_offset) # a = np.fromfile(f, dtype=np.int16) # print(len(a)) # print(len(a) / 3) if __name__ == '__main__': main()
#!/usr/env/python import os DATASETS_DIR = os.path.expanduser("~/Desktop/datasets/") def to_path(*args): return os.path.join(DATASETS_DIR, *args) # straightforward datasets MSRC_12 = to_path('MSRC-12', 'origData') UCR = to_path('ucr/UCRArchive_2018') UCR_INFO = to_path('ucr/DataSummary.csv') UWAVE = to_path('uWave', 'extracted') PAMAP = to_path('PAMAP_Dataset') PAMAP2 = to_path('PAMAP2_Dataset') WARD = to_path('WARD1.0') UCI_GAS = to_path('uci-gas-sensor') # ampds2 AMPD2_POWER = to_path('ampds2', 'electric') AMPD2_GAS = to_path('ampds2', 'gas') AMPD2_WEATHER = to_path('ampds2', 'weather') AMPD2_WATER = to_path('ampds2', 'water') # caltech-{101,256} CALTECH_101 = to_path('caltech', '101_ObjectCategories') CALTECH_256 = to_path('caltech', '256_ObjectCategories') # ECG data SHAREE_ECG = to_path('sharee-ecg-database') INCART_ECG = to_path('incart-12-lead-ecg')
#!/bin/env python import os import numpy as np from sklearn.datasets.samples_generator import make_blobs from joblib import Memory _memory = Memory('.', verbose=1) DATA_DIR = os.path.expanduser('~/Desktop/datasets/nn-search') join = os.path.join class Random: UNIFORM = 'uniform' GAUSS = 'gauss' WALK = 'walk' BLOBS = 'blobs' class Gist: DIR = join(DATA_DIR, 'gist') TRAIN = join(DIR, 'gist_train.npy') # noqa TEST = join(DIR, 'gist.npy') # noqa TEST_100 = join(DIR, 'gist_100k.npy') # noqa TEST_200 = join(DIR, 'gist_200k.npy') # noqa QUERIES = join(DIR, 'gist_queries.npy') # noqa TRUTH = join(DIR, 'gist_truth.npy') # noqa class Sift1M: DIR = join(DATA_DIR, 'sift1m') TRAIN = join(DIR, 'sift_learn.npy') # noqa TEST = join(DIR, 'sift_base.npy') # noqa TEST_100 = join(DIR, 'sift_100k.txt') # noqa TEST_200 = join(DIR, 'sift_200k.txt') # noqa QUERIES = join(DIR, 'sift_queries.npy') # noqa TRUTH = join(DIR, 'sift_groundtruth.npy') # noqa class Sift10M: DIR = join(DATA_DIR, 'sift1b') # TRAIN = join(DIR, 'big_ann_learn_10M.npy') # noqa TRAIN = join(DIR, 'big_ann_learn_1M.npy') # noqa # TODO use 10M? TRAIN_1M = join(DIR, 'big_ann_learn_1M.npy') # noqa TEST = join(DIR, 'sift_10M.npy') # noqa QUERIES = join(DIR, 'sift_queries.npy') # noqa TRUTH = join(DIR, 'true_nn_idxs_10M.npy') # noqa class Deep1M: """256D PCA of convnet activations; see OTQ paper supporting webiste, http://sites.skoltech.ru/compvision/projects/aqtq/""" DIR = join(DATA_DIR, 'deep1m') # noqa TRAIN = join(DIR, 'deep1M_learn.npy') # noqa TEST = join(DIR, 'deep1M_base.npy') # noqa TEST_100 = join(DIR, 'deep1M_test_100k.npy') # noqa QUERIES = join(DIR, 'deep1M_queries.npy') # noqa TRUTH_TRAIN = join(DIR, 'deep1M_truth_train.npy') # noqa TRUTH = join(DIR, 'deep1M_groundtruth.npy') # noqa class Convnet1M: DIR = join(DATA_DIR, 'convnet1m') # noqa TRAIN = join(DIR, 'convnet_train.npy') # noqa TEST = join(DIR, 'convnet_test.npy') # noqa TEST_100 = join(DIR, 'convnet_test_100k.npy') # noqa QUERIES = join(DIR, 'convnet_queries.npy') # noqa TRUTH_TRAIN = join(DIR, 'truth_train.npy') # noqa TRUTH = join(DIR, 'truth_test.npy') # noqa class Mnist: # following other papers (eg, "revisiting additive quantization"), # use mnist test set as queries and training set as database DIR = join(DATA_DIR, 'mnist') # noqa TEST = join(DIR, 'X_train.npy') # noqa QUERIES = join(DIR, 'X_test.npy') # noqa TRUTH = join(DIR, 'truth_Q=test_X=train.npy') # noqa class LabelMe: DIR = join(DATA_DIR, 'labelme') # noqa TRAIN = join(DIR, 'labelme_train.npy') # noqa TEST = join(DIR, 'labelme_train.npy') # noqa QUERIES = join(DIR, 'labelme_test.npy') # noqa TRUTH = join(DIR, 'labelme_truth.npy') # noqa class Glove: DIR = join(DATA_DIR, 'glove') # noqa TEST = join(DIR, 'glove_test.npy') # noqa TEST_100 = join(DIR, 'glove_100k.txt') # noqa TEST_200 = join(DIR, 'glove_200k.txt') # noqa QUERIES = join(DIR, 'glove_queries.npy') # noqa TRUTH = join(DIR, 'glove_truth.npy') # noqa # note that we've only run the real experiments on the ones reported # in the paper (i.e., no cherrypicking) ALL_REAL_DATASETS = [ Gist, Sift1M, Sift10M, Deep1M, Convnet1M, Mnist, LabelMe, Glove] def load_file(fname, *args, **kwargs): if fname.split('.')[-1] == 'txt': return np.loadtxt(fname, *args, **kwargs) return np.load(fname, *args, **kwargs) def extract_random_rows(X, how_many, remove_from_X=True): split_start = np.random.randint(len(X) - how_many - 1) split_end = split_start + how_many rows = np.copy(X[split_start:split_end]) if remove_from_X: return np.vstack((X[:split_start], X[split_end:])), rows return X, rows def _load_complete_dataset(which_dataset, num_queries=10): X_test = np.load(which_dataset.TEST) try: X_train = np.load(which_dataset.TRAIN) print("using separate test set!") except AttributeError: print("No training set found for dataset {}".format(str(which_dataset))) X_train = np.copy(X_test) try: Q = np.load(which_dataset.QUERIES) except AttributeError: assert num_queries > 1 X_train, Q = extract_random_rows(X_train, how_many=num_queries) try: true_nn = np.load(which_dataset.TRUTH).astype(np.int) except AttributeError: true_nn = None return X_train, Q, X_test, true_nn def _ground_truth_for_dataset(which_dataset): return None # TODO # XXX: not clear whether this function is correct in general, but works for # 784D with the nzeros we get for 32 and 64 codebooks def _insert_zeros(X, nzeros): N, D = X.shape D_new = D + nzeros X_new = np.zeros((N, D_new), dtype=X.dtype) step = int(D / (nzeros + 1)) - 1 for i in range(nzeros): in_start = step * i in_end = in_start + step # out_start = in_start + i + 1 out_start = (step + 1) * i out_end = out_start + step X_new[:, out_start:out_end] = X[:, in_start:in_end] # out_start = out_end # out_end += step out_end += 1 # account for the last 0 remaining_len = D - in_end out_remaining_len = D_new - out_end # print "step", step # print "in_start, in_end", in_start, in_end # print "out_start, out_end", out_start, out_end # print "D, D_new", D, D_new # print "remaining_len, out_remaining_len", remaining_len, out_remaining_len assert remaining_len == out_remaining_len assert remaining_len >= 0 if remaining_len: # X_new[:, out_end:out_end+remaining_len] = X[:, in_end:D] X_new[:, out_end:] = X[:, in_end:] assert np.array_equal(X[:, 0], X_new[:, 0]) assert np.array_equal(X[:, -1], X_new[:, -1]) return X_new def ensure_num_cols_multiple_of(X, multiple_of): remainder = X.shape[1] % multiple_of if remainder > 0: return _insert_zeros(X, multiple_of - remainder) return X # @_memory.cache # uncomment to get same randomness each time def load_dataset(which_dataset, N=-1, D=-1, norm_mean=False, norm_len=False, num_queries=10, Ntrain=-1, D_multiple_of=-1): true_nn = None # randomly generated datasets if which_dataset == Random.UNIFORM: X_test = np.random.rand(N, D) X_train = np.random.rand(Ntrain, D) if Ntrain > 0 else X_test Q = np.random.rand(num_queries, D) elif which_dataset == Random.GAUSS: X_test = np.random.randn(N, D) X_train = np.random.randn(Ntrain, D) if Ntrain > 0 else X_test Q = np.random.randn(num_queries, D) elif which_dataset == Random.WALK: X_test = np.random.randn(N, D) X_test = np.cumsum(X_test, axis=1) X_train = np.copy(X_test) if Ntrain > 0: X_train = np.random.randn(Ntrain, D) X_train = np.cumsum(X_train) Q = np.random.randn(num_queries, D) Q = np.cumsum(Q, axis=-1) elif which_dataset == Random.BLOBS: # centers is D x D, and centers[i, j] = (i + j) centers = np.arange(D) centers = np.sum(np.meshgrid(centers, centers), axis=0) X_test, _ = make_blobs(n_samples=N, centers=centers) X_train = np.copy(X_test) if Ntrain > 0: X_train, _ = make_blobs(n_samples=Ntrain, centers=centers) Q, true_nn = make_blobs(n_samples=num_queries, centers=centers) # datasets that are just one block of a "real" dataset elif isinstance(which_dataset, str): # assert False # TODO rm after real experiments X_test = load_file(which_dataset) X_test, Q = extract_random_rows(X_test, how_many=num_queries) X_train = np.copy(X_test) true_nn = _ground_truth_for_dataset(which_dataset) # "real" datasets with predefined train, test, queries, truth elif which_dataset in ALL_REAL_DATASETS: X_train, Q, X_test, true_nn = _load_complete_dataset( which_dataset, num_queries=num_queries) else: raise ValueError("unrecognized dataset {}".format(which_dataset)) N = X_test.shape[0] if N < 1 else N D = X_test.shape[1] if D < 1 else D X_test, X_train = np.copy(X_test)[:N, :D], X_train[:N, :D] Q = Q[:, :D] if len(Q.shape) > 1 else Q[:D] train_is_test = X_train.base is X_test or X_test.base is X_train train_test_equal = np.array_equal(X_train[:100], X_test[:100]) train_test_same = train_is_test or train_test_equal if train_test_same: print("WARNING: Training data is also the test data!") if train_is_test: X_test = np.copy(X_test) if norm_mean: means = np.mean(X_train, axis=0) X_train -= means X_test -= means Q -= means if norm_len: X_test /= np.linalg.norm(X_test, axis=1, keepdims=True) X_train /= np.linalg.norm(X_train, axis=1, keepdims=True) Q /= np.linalg.norm(Q, axis=-1, keepdims=True) # np.set_printoptions(precision=6) # print "start of Q:", Q[:5, :5] # print "start of X_test:", X_test[:5, :5] # TODO don't convert datasets that are originally uint8s to floats X_train = X_train.astype(np.float32) X_test = X_test.astype(np.float32) # Q = np.squeeze(Q.astype(np.float32)) Q = Q.astype(np.float32) if D_multiple_of > 1: X_train = ensure_num_cols_multiple_of(X_train, D_multiple_of) X_test = ensure_num_cols_multiple_of(X_test, D_multiple_of) Q = ensure_num_cols_multiple_of(Q, D_multiple_of) return X_train, Q, X_test, true_nn def read_yael_vecs(path, c_contiguous=True, limit_rows=-1, dtype=None): dim = np.fromfile(path, dtype=np.int32, count=2)[0] print("vector length = {}".format(dim)) if dtype is None: if 'fvecs' in path: dtype = np.float32 elif 'ivecs' in path: dtype = np.int32 elif 'bvecs' in path: dtype = np.uint8 else: raise ValueError("couldn't infer dtype from path {}".format(path)) itemsize = np.dtype(dtype).itemsize assert dim > 0 assert itemsize in (1, 2, 4) cols_for_dim = 4 // itemsize row_size_bytes = 4 + dim * itemsize row_size_elems = row_size_bytes // itemsize limit = int(limit_rows) * row_size_elems if limit_rows > 0 else -1 fv = np.fromfile(path, dtype=dtype, count=limit) fv = fv.reshape((-1, row_size_elems)) if not all(fv.view(np.int32)[:, 0] == dim): raise IOError("Non-uniform vector sizes in " + path) fv = fv[:, cols_for_dim:] if c_contiguous: fv = fv.copy() return fv
#!/usr/bin/env python import os # import matplotlib as mpl import matplotlib.pyplot as plt import numpy as np import pandas as pd import seaborn as sb from joblib import Memory from . import paths from . import files _memory = Memory('./') def _list_csvs(directory): return files.list_files(directory, endswith='.csv', abs_paths=True) ELECTRIC_PATHS = _list_csvs(paths.AMPD2_POWER) GAS_PATHS = _list_csvs(paths.AMPD2_GAS) WATER_PATHS = _list_csvs(paths.AMPD2_WATER) WEATHER_PATHS = _list_csvs(paths.AMPD2_WEATHER) ELECTRIC_COLS = 'UNIX_TS,WHE,RSE,GRE,MHE,B1E,BME,CWE,DWE,EQE,FRE,HPE,OFE,' \ 'UTE,WOE,B2E,CDE,DNE,EBE,FGE,HTE,OUE,TVE,UNE'.split(',') ELECTRIC_DATA_COLS = ELECTRIC_COLS[1:] # ELECTRIC_DATA_COLS.remove('MHE') # linear combo of other cols # ELECTRIC_DATA_COLS.remove('UNE') # linear combo of other cols GAS_DATA_COLS = ['counter', 'avg_rate', 'inst_rate'] WATER_DATA_COLS = ['counter', 'avg_rate'] WEATHER_TIME_COL = 'Date/Time' WEATHER_DATA_COLS = ['Temp (C)', 'Dew Point Temp (C)', 'Rel Hum (%)', 'Wind Dir (10s deg)', 'Wind Spd (km/h)', 'Visibility (km)', 'Stn Press (kPa)'] WEATHER_ALL_COLS = [WEATHER_TIME_COL] + WEATHER_DATA_COLS FIG_SAVE_DIR = os.path.join('figs', 'ampds') # ================================================================ public class HouseRecording(object): def __init__(self, path, cols=None): data = _read_file(path) self.path = path self.name = os.path.basename(path).split('.')[0] self.col_names = cols self.sampleTimes = data[:, 0] self.data = data[:, 1:] # XXX have to use all cols after the first # if 'power' in self.name: # print "initial sample times: ", self.sampleTimes[:50] # print # hack to deal with DWW water not having inst_rate # self.col_names = self.col_names[:self.data.shape[1]] self.data = self.data[:, :len(self.col_names)] class WeatherRecording(object): def __init__(self): df = _load_weather_data() self.name = 'weather' self.col_names = WEATHER_DATA_COLS self.sampleTimes = _datetime_strs_to_unix_timestamps(df[WEATHER_TIME_COL]) self.data = df[WEATHER_DATA_COLS].values.astype(np.float32) # ------------------------ top-level data loading functions def all_power_recordings(): return [HouseRecording(path, cols=ELECTRIC_DATA_COLS) for path in ELECTRIC_PATHS] def all_gas_recordings(): return [HouseRecording(path, cols=GAS_DATA_COLS) for path in GAS_PATHS] def all_water_recordings(): return [HouseRecording(path, cols=WATER_DATA_COLS) for path in WATER_PATHS] def all_weather_recordings(): return [WeatherRecording()] # just one data file, so just one recording def all_timestamp_recordings(): all_recordings = all_power_recordings() + all_gas_recordings() + \ all_water_recordings() + all_weather_recordings() # all_recordings = all_weather_recordings() # TODO rm for r in all_recordings: r.data = r.sampleTimes.astype(np.float64) r.name += '_timestamps' return all_recordings # ================================================================ private # def _read_file(path, cols=None): @_memory.cache def _read_file(path): df = pd.read_csv(path).fillna(method='backfill') # hold prev val # if cols is not None and len(cols) > 0: # timestamps = df[df.columns[0]] # return df.values.astype(np.int32) return df.values.astype(np.float64) # need f64 to not lose timestamps @_memory.cache def _load_weather_data(): path = WEATHER_PATHS[0] df = pd.read_csv(path, sep=',').fillna(method='backfill') # hold prev val return df[WEATHER_ALL_COLS] def _datetimes_to_unix_timestamps(datetimes): # https://stackoverflow.com/q/34038273 return (datetimes.astype(np.int64) / 1e6).astype(np.uint64) def _datetime_strs_to_unix_timestamps(strs): return _datetimes_to_unix_timestamps(pd.to_datetime(strs)) # ================================================================ main def save_fig_png(path): plt.savefig(path, dpi=300, bbox_inches='tight') def _prev_corrs_stats(corr): assert corr.shape[0] == corr.shape[1] # needs to be a correlation mat abs_corr = np.abs(corr) prev_corrs = np.zeros(len(corr) - 1) best_corrs = np.zeros(len(corr) - 1) for i, row in enumerate(abs_corr[1:]): # each row after the first prev_corrs[i] = row[i] # note that i is row index - 1 try: best_corr_idx = np.nanargmax(row[:i+1]) best_corrs[i] = row[best_corr_idx] except ValueError: # if row all nans best_corrs[i] = prev_corrs[i] assert not (best_corrs[i] < prev_corrs[i]) # double neg for nans # avg corr with prev variable, avg highest corr with any preceding variable return np.nanmean(prev_corrs), np.nanmean(best_corrs) def _plot_corr(data, fig, ax, add_title=True): """assumes data is row-major; ie, each col is one variable over time""" # cov = np.cov(data.T) corr = np.corrcoef(data.T) # im = ax.imshow(corr, interpolation='nearest', # cmap=plt.cm.RdBu, # norm=mpl.colors.Normalize(vmin=-1., vmax=1.)) # fig.colorbar(im, ax=ax) # sb.heatmap(corr, center=0, ax=ax, square=True) sb.heatmap(corr, vmin=-1, vmax=1, center=0, ax=ax, square=True) if add_title: mean_prev_corr, mean_best_corr = _prev_corrs_stats(corr) ax.set_title("|rho| prev, best prev =\n{:.2f}, {:.2f}".format( mean_prev_corr, mean_best_corr)) def plot_recordings(recordings, interval_len=1000, norm_means=False, mins_zero=False, savedir=None): for r in recordings: print(("recording {} has data of shape {}".format(r.name, r.data.shape))) fig, axes = plt.subplots(2, 4, figsize=(13, 7)) start_idxs = [0, len(r.data) - interval_len] end_idxs = [interval_len, len(r.data)] # any_nans_in_row = np.isnan(r.data).sum(axis=1) # print np.where(any_nans_in_row)[0] # continue for i, (start, end) in enumerate(zip(start_idxs, end_idxs)): timestamps = r.sampleTimes[start:end] data = r.data[start:end] if norm_means: data -= np.mean(data, axis=0).astype(data.dtype) elif mins_zero: data -= np.min(data, axis=0).astype(data.dtype) # print "data shape", data.shape # print "data final vals", data[-20:] # continue col = i + 1 axes[0, col].plot(timestamps, data, lw=1) axes[1, col].plot(timestamps[1:], np.diff(data, axis=0), lw=1) axes[0, col].set_title('data') axes[1, col].set_title('first derivs') # plot correlation matrices for orig data and first derivs cor_sample_length = max(10000, len(r.data) // 5) data = r.data[:cor_sample_length] _plot_corr(data, fig, axes[0, 0]) _plot_corr(np.diff(data, axis=0), fig, axes[1, 0]) data = r.data[-cor_sample_length:] _plot_corr(data, fig, axes[0, -1]) _plot_corr(np.diff(data, axis=0), fig, axes[1, -1]) # _plot_corr(r.data[:cor_sample_length], fig, axes[0, 0]) # data = r.data[-cor_sample_length:] # _plot_corr(data, fig, axes[2, 1]) plt.tight_layout() # plt.show() if savedir is not None: files.ensure_dir_exists(savedir) # plt.savefig(os.path.join(savedir, r.name)) save_fig_png(os.path.join(savedir, r.name)) def main(): recordings = [] recordings += all_gas_recordings() recordings += all_water_recordings() recordings += all_power_recordings() recordings += all_weather_recordings() norm_means = False # norm_means = True mins_zero = True plot_recordings(recordings, norm_means=norm_means, mins_zero=mins_zero, savedir=FIG_SAVE_DIR) # plt.show() if __name__ == '__main__': main()
#!/bin/env python from __future__ import absolute_import, division, print_function import numpy as np import os import PIL import pickle import psutil # pip install psutil import shutil import sys # just for stderr for warnings # import warnings from PIL import Image from python import files from python import image_utils from joblib import Memory _memory = Memory('.', verbose=1) IMAGENET_ONE_OF_EACH_PATH = '../datasets/one-of-each-imagenet' IMAGENET_ONE_OF_EACH_FLOW_PATH = '../datasets/one-of-each-imagenet-as-folders' # IMAGENET_64_PATH = os.path.expanduser("~/Desktop/datasets/imagenet64") # IMAGENET_TINY_PATH = os.path.expanduser("~/Desktop/datasets/tiny-imagenet-200") IMAGENET_64_PATH = '../datasets/imagenet64' IMAGENET_TINY_PATH = '../datasets/tiny-imagenet-200' IMAGENET_64_TRAIN_CHUNK_NSAMPLES = 128116 IMAGENET_TRAIN_PATH = '../datasets/ILSVRC2012/ILSVRC2012_img_train' IMAGENET_TEST_PATH = '/home/dblalock/datasets/ILSVRC2012/ILSVRC2012_img_val' if not os.path.exists(IMAGENET_TEST_PATH): # try to load local version IMAGENET_TEST_PATH = '../datasets/ILSVRC2012/ILSVRC2012_img_val' IMAGENET_10_CLASSES_TRAIN_PATH = '../datasets/ILSVRC2012_10/ILSVRC2012_img_train' IMAGENET_10_CLASSES_TEST_PATH = '../datasets/ILSVRC2012_10/ILSVRC2012_img_val' IMAGENET_100_CLASSES_TRAIN_PATH = '../datasets/ILSVRC2012_100/ILSVRC2012_img_train' IMAGENET_100_CLASSES_TEST_PATH = '../datasets/ILSVRC2012_100/ILSVRC2012_img_val' IMAGENET_1_EXAMPLE_TRAIN_PATH = '../datasets/imagenet-001-of-each' IMAGENET_10_EXAMPLES_TRAIN_PATH = '../datasets/imagenet-010-of-each' IMAGENET_25_EXAMPLES_TRAIN_PATH = '../datasets/imagenet-025-of-each' IMAGENET_50_EXAMPLES_TRAIN_PATH = '../datasets/imagenet-050-of-each' IMAGENET_100_EXAMPLES_TRAIN_PATH = '../datasets/imagenet-100-of-each' # ================================================================ Downsampled def _unpickle_file(path): with open(path, 'rb') as f: pydict = pickle.load(f) return pydict # @_memory.cache def _load_downsampled_data_file(path, layout='nhwc', dtype=None, X_out=None, y_out=None, start_row=None): d = _unpickle_file(path) X = d['data'] # NOTE: subtracting 1 so min idx is 0; this breaks synset lookup y = np.array(d['labels'], dtype=np.int32) - 1 y = y.ravel() # shouldn't be necessary, but might as well assert X.shape[0] == y.shape[0] assert len(X.shape) == 2 nchan = 3 npixels = X.shape[1] / nchan assert npixels * nchan == X.shape[1] # each row not one img? side_len = int(np.sqrt(npixels)) assert side_len * side_len == npixels X = X.reshape(X.shape[0], nchan, side_len, side_len) layout = 'nhwc' if layout is None else layout assert layout in ('nhwc', 'nchw') if layout == 'nhwc': X = np.moveaxis(X, 1, -1) # make channels last axis X = np.ascontiguousarray(X) if X_out is not None: assert dtype in (None, X_out.dtype) dtype = X_out.dtype if dtype is not None: X = X.astype(dtype) # print("X shape: ", X.shape) # print("y shape: ", y.shape) if start_row is not None: end_row = start_row + X.shape[0] if X_out is not None: assert start_row is not None X_out[start_row:end_row] = X if y_out is not None: assert start_row is not None y_out[start_row:end_row] = y return X, y def load_train_file_64x64(idx, verbose=0, **kwargs): assert idx in np.arange(1, 11) # valid indices are 1 thru 10 path = os.path.join(IMAGENET_64_PATH, "train_data_batch_{}".format(idx)) if verbose > 1: print("loading train file: ", path) return _load_downsampled_data_file(path, **kwargs) def _clean_which_file_idxs(which_file_idxs=None, dtype=None): if which_file_idxs is None: which_file_idxs = np.arange(1, 11) which_file_idxs = np.asarray(which_file_idxs, dtype=np.int32) # don't try to load more training data then we can actually fit in RAM mem_available = psutil.virtual_memory().available itemsize = dtype.itemsize if dtype is not None else 1 one_img_nbytes = 64 * 64 * 3 * itemsize one_file_nbytes = IMAGENET_64_TRAIN_CHUNK_NSAMPLES * one_img_nbytes max_nfiles = (mem_available // one_file_nbytes) - 1 # print("one_img_nbytes", one_img_nbytes) # print("one_file_nbytes", one_file_nbytes) # print("available mem", mem_available) # print("max_nfiles", max_nfiles) if max_nfiles < 1: raise MemoryError( "Minimum amount of RAM needed to load one chunk of ImageNet64x64 " "is {}B, but only {}B are available".format( one_file_nbytes, mem_available)) requested_nfiles = len(which_file_idxs) if max_nfiles < requested_nfiles: requested_nbytes = (requested_nfiles + 1) * one_file_nbytes requested_MB = requested_nbytes // int(1e6) available_MB = mem_available // int(1e6) print("imagenet.load_train_data_64x64: MemoryWarning: " "Only loading {}/10 chunks of ImageNet64 instead of requested " "{}/10 since not enough memory; would need {:}MB, but only {:}MB " "are available".format( max_nfiles, requested_nfiles, requested_MB, available_MB), file=sys.stderr) # warnings.warn(msg, UserWarning) which_file_idxs = which_file_idxs[:max_nfiles] assert np.min(which_file_idxs) >= 1 assert np.max(which_file_idxs) <= 10 return which_file_idxs # NOTE: total size of training data is around 16GB def load_train_data_64x64(which_file_idxs=None, layout='nhwc', dtype=None, verbose=1): which_file_idxs = _clean_which_file_idxs(which_file_idxs, dtype=dtype) if verbose > 0: print("load_train_data_64x64: loading file numbers: ", which_file_idxs) # import sys; sys.exit() if dtype is None: dtype = np.uint8 # default dtype # preallocate output matrix of appropriate size so that we can just # keep one copy of the data in memory (as opposed to loading all the # data matrices and then concatenating them) assert layout in ('nhwc', 'nchw') nrows_per_file = IMAGENET_64_TRAIN_CHUNK_NSAMPLES img_shape = (64, 64, 3) if layout == 'nhwc' else (3, 64, 64) combined_nrows = nrows_per_file * len(which_file_idxs) combined_shape = (combined_nrows,) + img_shape X_combined = np.zeros(combined_shape, dtype=dtype) y_combined = np.zeros(combined_nrows, dtype=np.int32) for i, idx in enumerate(which_file_idxs): start_row = nrows_per_file * i load_train_file_64x64( idx, layout=layout, X_out=X_combined, y_out=y_combined, start_row=start_row, verbose=verbose) return X_combined, y_combined def load_test_data_64x64(layout='nhwc', dtype=None): path = os.path.join(IMAGENET_64_PATH, "val_data") return _load_downsampled_data_file(path, layout=layout, dtype=dtype) # ================================================================ Tiny # # adapted from https://github.com/keras-team/keras-preprocessing/blob/master/ # # keras_preprocessing/image/utils.py under MIT license # def img_to_array(img, layout='nhwc', dtype='float32', mode='RGB'): # """Converts a PIL Image instance to a Numpy array. # # Arguments # img: PIL Image instance. # layout: Image data format, either "nchw" or "nhwc". # dtype: Dtype to use for the returned array. # # Returns # A 3D Numpy array. # # Raises # ValueError: if invalid `img` or `layout` is passed. # """ # # print("img info:", img.format, img.size, img.mode) # # if img.mode == 'L': # if img.mode != mode: # img = img.convert(mode=mode) # if layout not in ('nchw', 'nhwc'): # raise ValueError('Unknown layout: %s' % layout) # # Numpy array x has format (height, width, channel) # # or (channel, height, width) # # but original PIL image has format (width, height, channel) # x = np.asarray(img, dtype=dtype) # if len(x.shape) == 3: # if layout == 'nchw': # x = x.transpose(2, 0, 1) # elif len(x.shape) == 2: # # print("x is only rank 2...WTF!?") # if layout == 'nchw': # x = x.reshape((1, x.shape[0], x.shape[1])) # else: # x = x.reshape((x.shape[0], x.shape[1], 1)) # else: # raise ValueError('Unsupported image shape: %s' % (x.shape,)) # return x # def _resize_img(img, ratio_or_size): # if ratio_or_size is None or np.min(ratio_or_size) < 0: # return img # try: # nrows = ratio_or_size[0] # ncols = ratio_or_size[1] # except AttributeError: # nrows = img.height * ratio_or_size # ncols = img.width * ratio_or_size # new_size = (nrows, ncols) # is_downsampling = (nrows < img.height) or (ncols < img.width) # interp = PIL.Image.LANCZOS if is_downsampling else PIL.Image.BICUBIC # return img.resize(new_size, resample=interp) # def image_utils.load_jpg(path, layout='nhwc', dtype='float32', resample=None): # img = Image.open(path) # img = _resize_img(img, ratio_or_size=resamp) # return img_to_array(img, layout=layout, dtype=dtype) @_memory.cache def _load_tiny_clsids_to_nums(): wnids_path = os.path.join(IMAGENET_TINY_PATH, 'wnids.txt') with open(wnids_path) as f: lines = f.readlines() lines = [line.strip() for line in lines] return {s: i for i, s in enumerate(lines)} def _imagenet_tiny_cls_to_number(classnames): if isinstance(classnames, str): return _load_tiny_clsids_to_nums()[classnames] return [_load_tiny_clsids_to_nums()[name] for name in classnames] @_memory.cache def load_train_data_tiny(layout='nhwc', dtype=None, verbose=1): train_dir = os.path.join(IMAGENET_TINY_PATH, 'train') subdirs = files.list_subdirs(train_dir) all_classes = subdirs assert len(all_classes) == 200 # wrong number of classes?? subdir_paths = files.list_subdirs(train_dir, abs_paths=True) all_imgs = [] all_labels = [] for i, pth in enumerate(np.sort(subdir_paths)): classname = os.path.basename(pth) if verbose > 0: print("loading images for class {}...".format(classname)) imgs_subdir = os.path.join(pth, 'images') img_paths = files.list_files( imgs_subdir, endswith='.JPEG', abs_paths=True) assert len(img_paths) == 500 # supposed to be 500 examples per class... imgs = [image_utils.load_jpg(f, layout=layout, dtype=dtype)[np.newaxis, :, :, :] for f in img_paths] all_imgs += imgs lbl = _imagenet_tiny_cls_to_number(classname) all_labels += [lbl] * len(img_paths) X = np.concatenate(all_imgs, axis=0) y = np.array(all_labels, dtype=np.int32) return X, y @_memory.cache def load_test_data_tiny(layout='nhwc', dtype=None): # no labels given for "true" test set, so use the "val" subset as the # test set test_dir = os.path.join(IMAGENET_TINY_PATH, 'val') imgs_subdir = os.path.join(test_dir, 'images') img_paths = files.list_files( imgs_subdir, endswith='.JPEG', abs_paths=True) assert len(img_paths) == 10000 # wrong number of val images? # load images imgs = [image_utils.load_jpg(f, layout=layout, dtype=dtype)[np.newaxis, :, :, :] for f in img_paths] X = np.concatenate(imgs, axis=0) # load labels # TODO make sure this computation is correct lbls_path = os.path.join(test_dir, 'val_annotations.txt') with open(lbls_path, 'r') as f: lines = f.readlines() fnames = [line.split()[0] for line in lines] class_ids = [line.split()[1] for line in lines] # complicated way that doesn't rely on annotations being sorted fname_to_class_id = dict(zip(fnames, class_ids)) img_fnames = [os.path.basename(pth) for pth in img_paths] img_class_ids = [fname_to_class_id[fname] for fname in img_fnames] labels = _imagenet_tiny_cls_to_number(img_class_ids) y = np.array(labels, dtype=np.int32) return X, y # ================================================================ K-of-each # def load_data_one_of_each(layout='nhwc', dtype=None, size=None): def load_data_one_of_each(layout='nhwc', dtype=None, size=(224, 224)): # np_save_file = os.path.join(IMAGENET_ONE_OF_EACH_PATH, 'oneOfEach.npy') # cached_exists = os.path.exists(np_save_file) # if cached_exists: # return np.load(np_save_file) img_paths = files.list_files(IMAGENET_ONE_OF_EACH_PATH, endswith='.JPEG', abs_paths=True) assert len(img_paths) == 1000 # should be 1000 images... imgs = [image_utils.load_jpg(f, layout=layout, dtype=dtype, resample=size) for f in img_paths] if size is not None: # can only concat if same size imgs = [img[np.newaxis, :, :, :] for img in imgs] X = np.concatenate(imgs, axis=0) else: X = imgs # XXX this is a total hack that will break if we get >1 img per class, and # already (probably) doesn't match up with the synsets # lbls = [os.path.basename(path).split('_')[0] for path in img_paths] y = np.arange(len(X)) return X, y # ================================================================ example def load_flow_example(**kwargs): IMAGENET_EXAMPLE_PATH = os.path.abspath('../datasets/imagenet-example') # print("files in data dir:") # print(files.list_subdirs(IMAGENET_EXAMPLE_PATH)) # import sys; sys.exit() j = os.path.join kwargs.setdefault('target_size', (224, 224)) kwargs.setdefault('batch_size', 16) kwargs.setdefault('class_mode', 'categorical') import keras from keras.preprocessing import image train_datagen = image.ImageDataGenerator() val_datagen = image.ImageDataGenerator() test_datagen = image.ImageDataGenerator() # print("contents of train dir: ", files.list_subdirs(j(IMAGENET_EXAMPLE_PATH, 'train'))) train_generator = train_datagen.flow_from_directory( j(IMAGENET_EXAMPLE_PATH, 'train'), **kwargs) val_generator = val_datagen.flow_from_directory( j(IMAGENET_EXAMPLE_PATH, 'val'), **kwargs) test_generator = val_datagen.flow_from_directory( j(IMAGENET_EXAMPLE_PATH, 'val'), **kwargs) return train_generator, val_generator, test_generator def example_imagenet_train(): import tensorflow as tf import keras from python import models from python import approx_conv_v2 as aconv # both are necessary to actually get consistent output np.random.seed(123) tf.random.set_random_seed(123) model = models.get_model(models.VGG16, weights=None, input_shape=(224, 224, 3)) # swap out normal conv layer with our custom layer model = models.replace_layer_classes( model, {keras.layers.Conv2D: aconv.MyConv2D}) model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy']) model.summary() train_generator, val_generator, test_generator = load_flow_example() model.fit_generator(train_generator, steps_per_epoch=10, epochs=1, validation_steps=1, validation_data=val_generator) model.evaluate_generator(test_generator, steps=2) def main(): # X, y = load_train_file_64x64(1) # X, y = load_train_data_64x64([1, 2, 3]) # X, y = load_train_data_64x64([1, 2, 3, 4, 5]) # works # X, y = load_train_data_64x64() # correctly yields mem warning # X, y = load_train_data_64x64([1]) # X, y = load_test_data_64x64() # X, y = load_data_one_of_each(size=None) # no resampling # X, y = load_data_one_of_each(size=(224, 224)) # wow, imagenet-tiny looks like crap; lots of aliasing X, y = load_train_data_tiny() # X, y = load_test_data_tiny() print("X, y dtypes and shapes:") print(X.dtype) print(y.dtype) print(X.shape) print(y.shape) import matplotlib.pyplot as plt inp = 'y' count = 0 while inp == 'y': _, axes = plt.subplots(3, 3, figsize=(9, 9)) idxs = np.random.choice(np.arange(len(X)), size=axes.size) # offset = 0 # offset = 10000 # idxs = np.arange(offset + 9*count, offset + 9 + 9*count) for i, ax in enumerate(axes.ravel()): idx = idxs[i] img, classId = X[idx], y[idx] ax.imshow(img, interpolation='nearest') ax.set_title("Idx = {}, class = {}".format(idx, classId)) # plt.imshow(X[100*1000], interpolation='nearest') # plt.imshow(X[300*1000], interpolation='nearest') plt.tight_layout() plt.show() count += 1 inp = input("enter y to plot more random imgs; anything else to stop: ") def _folderize_imagenet_one_of_each(): # one-off script olddir = IMAGENET_ONE_OF_EACH_PATH newdir = IMAGENET_ONE_OF_EACH_FLOW_PATH files.ensure_dir_exists(newdir) old_files = files.list_files(olddir, endswith='.JPEG', abs_paths=True) for f in files.list_files(olddir, endswith='.JPEG', abs_paths=True): basename = os.path.basename(f) label = basename.split('_')[0] subdir = os.path.join(newdir, label) files.ensure_dir_exists(subdir) newpath = os.path.join(subdir, basename) # newpath = os.path.join(newdir, ) # print("oldpath: ", f, os.path.exists(f)) # print("newpath: ", newpath) shutil.copy(f, newpath) def _make_imagenet_k_of_each(k=10): out_path = '../datasets/imagenet-{:03d}-of-each'.format(k) print("writing to path: ", out_path) src_dir = IMAGENET_TRAIN_PATH for synset in files.list_subdirs(src_dir): subdir_path = os.path.join(src_dir, synset) img_paths = sorted(files.list_files(subdir_path, abs_paths=True)) img_paths = img_paths[:k] new_subdir = os.path.join(out_path, synset) files.ensure_dir_exists(new_subdir) for path in img_paths: fname = os.path.basename(path) new_path = os.path.join(new_subdir, fname) shutil.copy(path, new_path) if __name__ == '__main__': # example_imagenet_train() main() # _folderize_imagenet_one_of_each() # _make_imagenet_k_of_each(10) # _make_imagenet_k_of_each(25) # _make_imagenet_k_of_each(50) # _make_imagenet_k_of_each(100)
#!/usr/bin/env/python import os import numpy as np from joblib import Memory import pandas as pd from . import paths _memory = Memory('.', verbose=1, compress=9) UCR_DATASETS_DIR = paths.UCR UCR_INFO_PATH = paths.UCR_INFO # ================================================================ # Public # ================================================================ def all_ucr_datasets(): for dataDir in sorted(all_ucr_dataset_dirs()): yield UCRDataset(dataDir) class UCRDataset(object): def __init__(self, dataset_dir, sep='\t', precondition=True, znorm=True): self.name = name_from_dir(dataset_dir) self.X_train, y_train = read_ucr_train_data(dataset_dir, sep=sep) self.X_test, y_test = read_ucr_test_data(dataset_dir, sep=sep) # self.y_train = y_train # self.y_test = y_test all_lbls = np.r_[y_train, y_test] uniq_lbls = np.unique(all_lbls) new_lbls = np.argsort(uniq_lbls) # same if labels are 0..(nclasses-1) mapping = dict(zip(uniq_lbls, new_lbls)) self.y_train = np.array([mapping[lbl] for lbl in y_train]) self.y_test = np.array([mapping[lbl] for lbl in y_test]) # self.nclasses = len(uniq_lbls) # MelbournePedestrian has nans, even though not in missing data list for X in (self.X_train, self.X_test): for d in range(X.shape[1]): col = X[:, d] nan_idxs = np.isnan(col) if nan_idxs.sum() > 0: # print("self.name: ", self.name) # print("original number of nans: ", np.sum(nan_idxs)) # X[nan_idxs, d] = col.mean() fillval = np.nanmedian(col) if np.isnan(fillval): # handle all-nan cols, which happens in Crop fillval = np.nanmedian(X) col[nan_idxs] = fillval # np.nan_to_num(col, copy=False, nan=np.median(col)) # print("new number of nans: ", np.isnan(X[:, d]).sum()) # print("new number of nans: ", np.isnan(col).sum()) if znorm: self.X_train -= self.X_train.mean(axis=1, keepdims=True) self.X_test -= self.X_test.mean(axis=1, keepdims=True) eps = 1e-20 self.X_train *= 1 / (self.X_train.std(axis=1, keepdims=True) + eps) self.X_test *= 1 / (self.X_test.std(axis=1, keepdims=True) + eps) elif precondition: # weaker than znormalization since one offset and scale applied # to all dims and all samples in both train and test sets; this # is basically just here because the values in MelbournePedestrian # are huge and screw up numerical algorithms self.orig_mean = np.mean(self.X_train) self.X_train -= self.orig_mean self.X_test -= self.orig_mean self.orig_std = np.std(self.X_train) self.X_train /= self.orig_std self.X_test /= self.orig_std assert len(self.X_train) == len(self.y_train) assert len(self.X_test) == len(self.y_test) # if self.name == 'MelbournePedestrian': # print("I am MelbournePedestrian!") # print('new labels: ', new_lbls) # print("X_train num nans", np.sum(np.isnan(self.X_train))) # print("X_test num nans", np.sum(np.isnan(self.X_test))) # # import sys; sys.exit() # if self.name == 'Wafer': # print("original uniq labels train", np.unique(self.y_train)) # print("original uniq labels test", np.unique(self.y_test)) def all_ucr_dataset_dirs(): return _ucr_datasets_in_dir(UCR_DATASETS_DIR) # ================================================================ # Private # ================================================================ def _ucr_datasets_in_dir(dirpath): datasetsPath = os.path.expanduser(dirpath) files = os.listdir(datasetsPath) rm_dir = 'Missing_value_and_variable_length_datasets_adjusted' if rm_dir in files: files.remove(rm_dir) for i in range(len(files)): files[i] = os.path.join(datasetsPath, files[i]) dirs = list(filter(os.path.isdir, files)) return dirs @_memory.cache def _readtxt(path, sep=None): return np.genfromtxt(path, delimiter=sep).astype(np.float32) def read_data_file(path, sep=None, mean_norm=False): D = _readtxt(path, sep=sep) labels = D[:, 0].astype(np.int) X = D[:, 1:] if mean_norm: X -= np.mean(X, axis=1, keepdims=True) return (X, labels) def name_from_dir(datasetDir): return os.path.basename(datasetDir) def dir_from_name(datasetName): return os.path.join(paths.UCR, datasetName) def read_ucr_data_in_dir(datasetDir, train, sep=None): datasetName = name_from_dir(datasetDir) if train: fileName = datasetName + "_TRAIN.tsv" else: fileName = datasetName + "_TEST.tsv" filePath = os.path.join(datasetDir, fileName) return read_data_file(filePath, sep=sep) def read_ucr_train_data(datasetDir, sep=None): return read_ucr_data_in_dir(datasetDir, train=True, sep=sep) def read_ucr_test_data(datasetDir, sep=None): return read_ucr_data_in_dir(datasetDir, train=False, sep=sep) # combines train and test data def read_all_ucr_data(ucrDatasetDir): X_train, Y_train = read_ucr_train_data(ucrDatasetDir) X_test, Y_test = read_ucr_test_data(ucrDatasetDir) X = np.r_[X_train, X_test] Y = np.r_[Y_train, Y_test] return X, Y @_memory.cache def load_ucr_dset_stats(): df = pd.read_csv(UCR_INFO_PATH) df['l2-1nn-acc'] = 1. - df['ED (w=0)'] return df # ================================================================ Main @_memory.cache def _load_ucr_stats_df(): stats = [] for i, datasetDir in enumerate(all_ucr_dataset_dirs()): # Xtrain, _ = read_ucr_train_data(datasetDir) # Xtest, Ytest = read_ucr_test_data(datasetDir) dset = UCRDataset(datasetDir) N, D = dset.X_train.shape M, D = dset.X_test.shape nclasses = len(np.unique(dset.y_test)) stats.append({'Dataset': dset.name, 'N': N, 'D': D, 'M': M, 'nclasses': nclasses}) # print('%30s:\t%d\t%d\t%d\t%d' % (name_from_dir(datasetDir), # N, M, D, nclasses) return pd.DataFrame.from_records(stats) def main(): # dsets = all_ucr_datasets() # for dset in dsets: # print("loaded ucr dset:", dset.name) # # return df = _load_ucr_stats_df() # df = df.sort_values(axis=1) # df = df.loc[df['N'] > 100] # df = df.loc[df['M'] > 100] print("ucr dset stats:") # print(df['M'].sort_values(ascending=False)) print("number of dsets:", df.shape[0]) print("mean, median Ntrain: ", df['N'].mean(), df['N'].median()) print("mean, median Ntest: ", df['M'].mean(), df['M'].median()) print("mean, median length: ", df['D'].mean(), df['D'].median()) mvals = df['M'].to_numpy() mvals = np.sort(mvals) length = len(mvals) total_sizes = np.array([m * (length - i) for i, m in enumerate(mvals)]) max_idx = np.argmax(total_sizes) best_m_cutoff = mvals[max_idx] print("best num dsets, m, sz = ", length - max_idx, best_m_cutoff, total_sizes[max_idx]) print("type of mvals: ", type(mvals)) for cutoff in [100, 200, 256, 300, 400, 500, 512, 1000]: ndsets = (mvals >= cutoff).sum() total_sz = total_sizes[ndsets-1] print(f"n >= {cutoff}: {ndsets} dsets, total sz = {total_sz}") # import matplotlib.pyplot as plt # xvals = length - np.arange(length) # # xvals = np.arange(length) # # plt.plot(xvals, total_sizes[::-1]) # plt.plot(xvals, total_sizes) # plt.plot(xvals, mvals) # plt.show() # df = df.loc[df['M'] >= best_m_cutoff] # print("---- after cutting off M to maximize mat sizes:") df = df.loc[df['N'] >= 128] print("---- after cutting off N to number of centroids:") print("number of dsets: ", len(df)) print("mean, median Ntrain: ", df['N'].mean(), df['N'].median()) print("mean, median Ntest: ", df['M'].mean(), df['M'].median()) print("mean, median length: ", df['D'].mean(), df['D'].median()) print("mean, median nclasses: ", df['nclasses'].mean(), df['nclasses'].median()) print("min, max nclasses: ", df['nclasses'].min(), df['nclasses'].max()) if __name__ == '__main__': np.set_printoptions(formatter={'float': lambda f: "{:.3}".format(f)}) main()
#!/bin/env python from __future__ import absolute_import, division, print_function from scipy import io import numpy as np import os from joblib import Memory _memory = Memory('.', verbose=1) DATADIR = '../datasets/svhn' TRAIN_PATH = os.path.join(DATADIR, 'train_32x32.mat') TEST_PATH = os.path.join(DATADIR, 'test_32x32.mat') EXTRA_PATH = os.path.join(DATADIR, 'extra_32x32.mat') def extract_data_from_mat_file(path): matlab_dict = io.loadmat(path) X, y = matlab_dict['X'], matlab_dict['y'].ravel() X = np.transpose(X, (3, 0, 1, 2)) # make classes be 0-9 instead of 1-10; this way the classes line up # with the actual digits y[y == 10] = 0 assert len(y.shape) == 1 assert X.shape[0] == len(y) assert X.shape[1] == 32 assert X.shape[2] == 32 assert X.shape[-1] == 3 return X, y @_memory.cache def load_data(): X_train, y_train = extract_data_from_mat_file(TRAIN_PATH) X_test, y_test = extract_data_from_mat_file(TEST_PATH) return (X_train, y_train), (X_test, y_test) def load_extra_data(): return extract_data_from_mat_file(EXTRA_PATH) def main(): import matplotlib.pyplot as plt (X_train, y_train), (X_test, y_test) = load_data() # hacky way to visualize extra data using same code # X_extra, y_extra = load_extra_data() # X_train, X_test = X_extra, X_extra # y_train, y_test = y_extra, y_extra _, axes = plt.subplots(4, 4, figsize=(9, 9)) for i, ax in enumerate(axes.ravel()): X = X_test if i % 2 else X_train y = y_test if i % 2 else y_train idx = np.random.choice(X.shape[0]) ax.imshow(X[idx]) ax.set_title("class = {}".format(y[idx])) plt.tight_layout() plt.show() if __name__ == '__main__': main()
#!/bin/env/python """utility functions for data munging""" from __future__ import absolute_import, division, print_function import numpy as np import sklearn def split_train_test(X, Y, train_frac=.8, random_state=123): """Returns X_train, X_test, y_train, y_test""" np.random.seed(123) return sklearn.model_selection.train_test_split( X, Y, train_size=train_frac, random_state=random_state) def stratified_split_train_test(X, Y, train_frac=.8, random_state=123): """Returns X_train, X_test, y_train, y_test""" return sklearn.model_selection.train_test_split( X, Y, train_size=train_frac, stratify=Y, random_state=random_state)
#!/bin/env python # Load 3-lead ECG recordings from SHAREE Database: # https://physionet.org/content/shareedb/1.0.0/ from __future__ import division, print_function import matplotlib.pyplot as plt import numpy as np import os from . import paths from . import files from joblib import Memory _memory = Memory('.', verbose=0) DATA_DIR = paths.SHAREE_ECG NUM_RECORDINGS = 139 NUM_CHANNELS = 3 RAW_DTYPE = np.uint16 # RAW_DTYPE = np.int16 SAMPLES_PER_SEC = 128 SAMPLES_PER_MIN = SAMPLES_PER_SEC * 60 SAMPLES_PER_HOUR = SAMPLES_PER_MIN * 60 @_memory.cache def load_recording_ids(): fpaths = files.list_files(DATA_DIR, abs_paths=False, endswith='.dat') assert len(fpaths) == NUM_RECORDINGS return fpaths @_memory.cache def load_recording(rec_id, limit_nhours=None, dtype=np.float32): # dtype = np.float32 if dtype is None else dtype path = os.path.join(DATA_DIR, rec_id) a = np.fromfile(path, dtype=RAW_DTYPE) assert len(a) % NUM_CHANNELS == 0 a = a.reshape(-1, NUM_CHANNELS) # looks like it's rowmajor # a = a.reshape(NUM_CHANNELS, -1).T # is colmajor clearly wrong? EDIT: yes if limit_nhours and limit_nhours > 0: a = a[:int(limit_nhours * SAMPLES_PER_HOUR)] a = a[SAMPLES_PER_MIN:] # often a bunch of garbage at the beginning a = a.astype(dtype) # small amount of smoothing since heavily oversampled + noisy # filt = np.hamming(5).astype(np.float32) filt = np.hamming(5).astype(np.float32) filt /= np.sum(filt) for j in range(a.shape[1]): a[:, j] = np.convolve(a[:, j], filt, mode='same') return a # def load_recordings(generator=False, plot=False, **kwargs): def load_recordings(plot=False, **kwargs): rec_ids = load_recording_ids() recs = [] for i, rec_id in enumerate(rec_ids): print("loading rec id: ", rec_id) rec = load_recording(rec_id, **kwargs) recs.append(rec) if plot: if i < 5: offset = SAMPLES_PER_MIN a = rec[offset:(offset + 1000)] print('about to plot recording', rec_id) plt.figure(figsize=(9, 7)) plt.plot(a) plt.show() else: return return recs if __name__ == '__main__': # print("done") print("about to call load_recordings") load_recordings(plot=True) # print("rec ids: ", load_recording_ids()) print("called load_recordings")
#!/bin/env python # Load 3-lead ECG recordings from SHAREE Database: # https://physionet.org/content/shareedb/1.0.0/ from __future__ import division, print_function import matplotlib.pyplot as plt import numpy as np import os from . import paths from . import files from joblib import Memory _memory = Memory('.', verbose=0) DATA_DIR = paths.INCART_ECG NUM_RECORDINGS = 75 NUM_CHANNELS = 12 RAW_DTYPE = np.int16 SAMPLES_PER_SEC = 257 SAMPLES_PER_MIN = SAMPLES_PER_SEC * 60 SAMPLES_PER_HOUR = SAMPLES_PER_MIN * 60 @_memory.cache def load_recording_ids(): fpaths = files.list_files(DATA_DIR, abs_paths=False, endswith='.dat') assert len(fpaths) == NUM_RECORDINGS return fpaths @_memory.cache def load_recording(rec_id, limit_nhours=None, dtype=np.float32): path = os.path.join(DATA_DIR, rec_id) a = np.fromfile(path, dtype=RAW_DTYPE) assert len(a) % NUM_CHANNELS == 0 a = a.reshape(-1, NUM_CHANNELS) # yep, clearly rowmajor when plotted if limit_nhours and limit_nhours > 0: a = a[:int(limit_nhours * SAMPLES_PER_HOUR)] a = a[SAMPLES_PER_MIN:] # often a bunch of garbage at the beginning a = a.astype(dtype) a -= a.mean(axis=0) # just so r_sq values are more meaningful # small amount of smoothing since heavily oversampled + noisy # filt = np.hamming(5).astype(np.float32) # filt = np.hamming(5).astype(np.float32) # filt /= np.sum(filt) # for j in range(a.shape[1]): # a[:, j] = np.convolve(a[:, j], filt, mode='same') return a def load_recordings(plot=False, **kwargs): rec_ids = load_recording_ids() recs = [] for i, rec_id in enumerate(rec_ids): print("loading rec id: ", rec_id) rec = load_recording(rec_id, **kwargs) recs.append(rec) if plot: if i < 5: offset = 0 a = rec[offset:(offset + 1000)] print("plotting recording {} with shape: {}".format( rec_id, rec.shape)) plt.figure(figsize=(9, 7)) plt.plot(a) plt.show() else: return return recs if __name__ == '__main__': print("about to call load_recordings") load_recordings(plot=True) print("called load_recordings")
#!/bin/env python # from __future__ import absolute_import, division, print_function from __future__ import division, print_function import numpy as np from . import paths from . import image_utils as imgs from joblib import Memory _memory = Memory('.', verbose=1) DATADIR_101 = paths.CALTECH_101 DATADIR_256 = paths.CALTECH_256 # _DEFAULT_CALTECH_KWARGS = dict(resample=(224, 224), crop='center', verbose=2) _DEFAULT_CALTECH_KWARGS = dict(resample=(224, 224), crop='center') _CALTECH_101_KWARGS = dict( dirpath=DATADIR_101, remove_classes='BACKGROUND_Google') _CALTECH_256_KWARGS = dict( dirpath=DATADIR_256, remove_classes='257.clutter') @_memory.cache def load_caltech101(**kwargs): [kwargs.setdefault(*item) for item in _DEFAULT_CALTECH_KWARGS.items()] return imgs.load_jpegs_from_dir(**_CALTECH_101_KWARGS, **kwargs) @_memory.cache def load_caltech256(**kwargs): [kwargs.setdefault(*item) for item in _DEFAULT_CALTECH_KWARGS.items()] return imgs.load_jpegs_from_dir(**_CALTECH_256_KWARGS, **kwargs) @_memory.cache def load_caltech101_ids(**kwargs): return imgs.load_jpegs_from_dir( **_CALTECH_101_KWARGS, only_return_path=True, **kwargs) @_memory.cache def load_caltech256_ids(**kwargs): return imgs.load_jpegs_from_dir( **_CALTECH_256_KWARGS, only_return_path=True, **kwargs) # @_memory.cache def load_caltech_img(img_id, **kwargs): [kwargs.setdefault(*item) for item in _DEFAULT_CALTECH_KWARGS.items()] path = img_id # load_jpegs_from_dir returns abs path as id return imgs.load_jpg(path, **kwargs).astype(np.float32) # img = imgs.load_jpg(path, **kwargs).astype(np.float32) # print("img.shape", img.shape) # assert img.shape[:2] == (224, 224) # return img def main(): import matplotlib.pyplot as plt # caltech 101 (X, y), label2cls = imgs.load_jpegs_from_dir( # DATADIR_101, remove_classes='BACKGROUND_Google') # DATADIR_101, remove_classes='BACKGROUND_Google', crop='center') DATADIR_101, remove_classes='BACKGROUND_Google', pad='square') # # DATADIR_101, remove_classes='BACKGROUND_Google', resample=(224, 224)) # caltech 256 # (X, y), label2cls = imgs.load_jpegs_from_dir( # DATADIR_256, remove_classes='257.clutter', verbose=2) if isinstance(X, np.ndarray): print("X shape: ", X.shape) else: print("X is a list of length", len(X)) print("X[0] has shape: ", X[0].shape) print("y shape: ", y.shape) _, axes = plt.subplots(4, 4, figsize=(9, 9)) for i, ax in enumerate(axes.ravel()): idx = np.random.choice(len(X)) ax.imshow(X[idx]) label = label2cls[y[idx]] ax.set_title(label) plt.tight_layout() plt.show() if __name__ == '__main__': main()
#!/bin/env python from __future__ import absolute_import, division, print_function import numpy as np from python import image_utils as imgs from joblib import Memory _memory = Memory('.', verbose=1) DATADIR_101 = '../datasets/caltech/101_ObjectCategories' def main(): import matplotlib.pyplot as plt # caltech 101 (X, y), label2cls = imgs.load_jpegs_from_dir( # TODO ) if isinstance(X, np.ndarray): print("X shape: ", X.shape) else: print("X is a list of length", len(X)) print("X[0] has shape: ", X[0].shape) print("y shape: ", y.shape) _, axes = plt.subplots(4, 4, figsize=(9, 9)) for i, ax in enumerate(axes.ravel()): idx = np.random.choice(len(X)) ax.imshow(X[idx]) label = label2cls[y[idx]] ax.set_title(label) plt.tight_layout() plt.show() if __name__ == '__main__': main()
#!/usr/bin/env python from __future__ import print_function import numpy as np from sklearn.datasets import load_digits import timeit import bolt # ================================================================ utils def _dists_sq(X, q): diffs = X - q return np.sum(diffs * diffs, axis=-1) def _dists_l1(X, q): diffs = np.abs(X - q) return np.sum(diffs, axis=-1) def _element_size_bytes(x): return np.dtype(x.dtype).itemsize def _corr(x, y): x, y = x.astype(np.float64), y.astype(np.float64) x = x.ravel() - np.mean(x) y = y.ravel() - np.mean(y) r = np.mean(x * y) / (np.std(x) * np.std(y)) assert -1.00001 <= r <= 1.00001 return r def _sq_dists_to_vectors(X, queries, rowNorms=None, queryNorms=None): Q = queries.shape[0] mat_size = X.shape[0] * Q mat_size_bytes = _element_size_bytes(X[0] + queries[0]) if mat_size_bytes > int(1e9): print("WARNING: _sq_dists_to_vectors: attempting to create a matrix" "of size {} ({}B)".format(mat_size, mat_size_bytes)) if rowNorms is None: rowNorms = np.sum(X * X, axis=1, keepdims=True) if queryNorms is None: queryNorms = np.sum(queries * queries, axis=1) dotProds = np.dot(X, queries.T) return (-2 * dotProds) + rowNorms + queryNorms # len(X) x len(queries) def top_k_idxs(elements, k, smaller_better=True, axis=-1): if smaller_better: # return indices of lowest elements which_nn = np.arange(k) return np.argpartition(elements, kth=which_nn, axis=axis)[:k] else: # return indices of highest elements which_nn = (elements.shape[axis] - 1 - np.arange(k))[::-1] # print "elements.shape", elements.shape # print "using which_nn: ", which_nn return np.argpartition(elements, kth=which_nn, axis=axis)[-k:][::-1] def _knn(X, Q, k=1000, print_every=5, block_sz=128): nqueries = Q.shape[0] nblocks = int(np.ceil(nqueries / float(block_sz))) truth = np.full((nqueries, k), -999, dtype=np.int32) if nqueries <= block_sz: dists = _sq_dists_to_vectors(Q, X) assert dists.shape == (Q.shape[0], X.shape[0]) for i in range(nqueries): truth[i, :] = top_k_idxs(dists[i, :], k) return truth for b in range(nblocks): # recurse to fill in knn for each block start = b * block_sz end = min(start + block_sz, nqueries) rows = Q[start:end, :] truth[start:end, :] = _knn(X, rows, k=k, block_sz=block_sz) if b % print_every == 0: print("computing top k for query block " \ "{} (queries {}-{})...".format(b, start, end)) assert np.all(truth != -999) return truth def _create_randn_encoder(Ntrain=100, Ntest=20, D=64): enc = bolt.Encoder() X_train = np.random.randn(Ntrain, D) X_test = np.random.randn(Ntest, D) enc.fit(X_train, just_train=True) enc.set_data(X_test) return enc # ================================================================ tests def test_smoketest(): """Test that `bolt.Encoder`'s methods don't crash""" D = 64 enc = _create_randn_encoder(D=D) Nqueries = 5 Q = np.random.randn(Nqueries, D) [enc.transform(q) for q in Q] for k in [1, 3]: [enc.knn(q, k) for q in Q] def _fmt_float(x): return '{}.'.format(int(x)) if x == int(x) else '{:.3f}'.format(x) def _load_digits_X_Q(nqueries): X, _ = load_digits(return_X_y=True) return X[:-nqueries], X[-nqueries:] # X, Q def test_time_space_savings(): # mostly to verify readme code np.set_printoptions(formatter={'float_kind': _fmt_float}) nqueries = 20 X, Q = _load_digits_X_Q(nqueries) enc = bolt.Encoder(accuracy='lowest', reduction=bolt.Reductions.DOT_PRODUCT) enc.fit(X) # massive space savings print("original space usage: {}B".format(X.nbytes)) # 1777 * 64 * 8B = 909KB print("bolt space usage: {}B".format(enc.nbytes)) # 1777 * 2B = 3.55KB # massive time savings (~10x here, but often >100x on larger datasets # with less Python overhead; see the Bolt paper) t_np = timeit.Timer(lambda: [np.dot(X, q) for q in Q]).timeit(5) # ~8ms t_bolt = timeit.Timer(lambda: [enc.transform(q) for q in Q]).timeit(5) # ~800us print("Numpy / BLAS time, Bolt time: {:.3f}ms, {:.3f}ms".format( t_np * 1000, t_bolt * 1000)) def test_unquantize(): X, Q = _load_digits_X_Q(nqueries=20) enc = bolt.Encoder('dot', accuracy='high').fit(X) dots_true = [np.dot(X, q) for q in Q] dots_bolt = [enc.transform(q, unquantize=True) for q in Q] diffs = [true_vals - bolt_vals for true_vals, bolt_vals in zip(dots_true, dots_bolt)] mse = np.mean([np.mean(diff*diff) for diff in diffs]) var = np.mean([np.var(true_vals) for true_vals in dots_true]) print("dot product unquantize mse / variance: ", mse / var) assert (mse / var) < .01 # print "true, bolt dot prods" # print dots_true[0][:20].astype(np.int32) # print dots_bolt[0][:20].astype(np.int32) enc = bolt.Encoder('l2', accuracy='high').fit(X) dists_true = [_dists_sq(X, q) for q in Q] dists_bolt = [enc.transform(q, unquantize=True) for q in Q] diffs = [true_vals - bolt_vals for true_vals, bolt_vals in zip(dists_true, dists_bolt)] mse = np.mean([np.mean(diff*diff) for diff in diffs]) var = np.mean([np.var(true_vals) for true_vals in dots_true]) print("squared l2 unquantize mse / variance: ", mse / var) assert (mse / var) < .01 def test_basic(): # np.set_printoptions(precision=3) np.set_printoptions(formatter={'float_kind': _fmt_float}) nqueries = 20 # nqueries = 10 # nqueries = 3 X, Q = _load_digits_X_Q(nqueries) # TODO rm this block # shift = 100. # shift = 100 # scaleby = 1. # scaleby = 3.5 # acc goes to **** at accelerating rate as this gets larger... # scaleby = 4 # scaleby = 1.0 # X, Q = X + shift, Q + shift # X, Q = X * scaleby, Q * scaleby # X = X[:200] # X = X[:50] # X = X[:20] # X, _ = load_digits(return_X_y=True) # Q = X[-nqueries:] # X = X[:-nqueries] # print "X.shape", X.shape # print "X nbytes", X.nbytes # ------------------------------------------------ squared l2 enc = bolt.Encoder(accuracy='low', reduction=bolt.Reductions.SQUARED_EUCLIDEAN) enc.fit(X) l2_corrs = np.empty(len(Q)) for i, q in enumerate(Q): l2_true = _dists_sq(X, q).astype(np.int) l2_bolt = enc.transform(q) l2_corrs[i] = _corr(l2_true, l2_bolt) if i == nqueries - 1: print("l2 true: ", l2_true) print("l2 bolt: ", l2_bolt) print("corr: ", l2_corrs[i]) mean_l2 = np.mean(l2_corrs) std_l2 = np.std(l2_corrs) assert mean_l2 > .95 print("--> squared l2 dist correlation: {} +/- {}".format(mean_l2, std_l2)) # return # ------------------------------------------------ dot product enc = bolt.Encoder(accuracy='low', reduction=bolt.Reductions.DOT_PRODUCT) enc.fit(X) dot_corrs = np.empty(nqueries) for i, q in enumerate(Q): dots_true = np.dot(X, q) dots_bolt = enc.transform(q) dot_corrs[i] = _corr(dots_true, dots_bolt) mean_dot = np.mean(dot_corrs) std_dot = np.std(dot_corrs) assert mean_dot > .95 print("--> dot product correlation: {} +/- {}".format(mean_dot, std_dot)) # ------------------------------------------------ l2 knn enc = bolt.Encoder(accuracy='low', reduction='l2') enc.fit(X) k_bolt = 10 # tell bolt to search for true knn k_true = 10 # compute this many true neighbors true_knn = _knn(X, Q, k_true) bolt_knn = [enc.knn(q, k_bolt) for q in Q] contained = np.empty((nqueries, k_bolt), dtype=np.bool) for i in range(nqueries): true_neighbors = true_knn[i] bolt_neighbors = bolt_knn[i] for j in range(k_bolt): contained[i, j] = bolt_neighbors[j] in true_neighbors precision = np.mean(contained) print("--> l2 knn precision@{}: {}".format(k_bolt, precision)) assert precision > .6 # # print "true_knn, bolt_knn:" # # print true_knn[:20, :20] # # print bolt_knn[:20] # ------------------------------------------------ dot knn enc = bolt.Encoder(accuracy='low', reduction='dot') # enc = bolt.Encoder(accuracy='high', reduction='dot') enc.fit(X) k_bolt = 10 # tell bolt to search for true knn k_true = 10 # compute this many true neighbors true_dists = np.dot(X, Q.T) # true_dists = [np.dot(X, q) for q in Q] true_knn = np.empty((nqueries, k_true), dtype=np.int64) for i in range(nqueries): true_knn[i, :] = top_k_idxs( true_dists[:, i], k_true, smaller_better=False) bolt_knn = [enc.knn(q, k_bolt) for q in Q] contained = np.empty((len(Q), k_bolt), dtype=np.bool) for i in range(len(Q)): true_neighbors = true_knn[i] # bolt_dists = enc.transform(Q[i]) # bolt_neighbors = top_k_idxs(bolt_dists, k_bolt, smaller_better=True) bolt_neighbors = bolt_knn[i] # TODO uncomment for j in range(k_bolt): contained[i, j] = bolt_neighbors[j] in true_neighbors precision = np.mean(contained) print("--> max inner product knn precision@{}: {}".format( k_bolt, precision)) assert precision > .6 # print("true_knn, bolt_knn:") # print(true_knn[:5]) # print(bolt_knn[:5]) if __name__ == '__main__': test_basic()
# This file is part of Eigen, a lightweight C++ template library # for linear algebra. # # Copyright (C) 2012 Keir Mierle <mierle@gmail.com> # # This Source Code Form is subject to the terms of the Mozilla # Public License v. 2.0. If a copy of the MPL was not distributed # with this file, You can obtain one at http://mozilla.org/MPL/2.0/. # # Author: mierle@gmail.com (Keir Mierle) # # Make the long-awaited conversion to MPL. lgpl3_header = ''' // Eigen is free software; you can redistribute it and/or // modify it under the terms of the GNU Lesser General Public // License as published by the Free Software Foundation; either // version 3 of the License, or (at your option) any later version. // // Alternatively, you can redistribute it and/or // modify it under the terms of the GNU General Public License as // published by the Free Software Foundation; either version 2 of // the License, or (at your option) any later version. // // Eigen is distributed in the hope that it will be useful, but WITHOUT ANY // WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS // FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License or the // GNU General Public License for more details. // // You should have received a copy of the GNU Lesser General Public // License and a copy of the GNU General Public License along with // Eigen. If not, see <http://www.gnu.org/licenses/>. ''' mpl2_header = """ // This Source Code Form is subject to the terms of the Mozilla // Public License v. 2.0. If a copy of the MPL was not distributed // with this file, You can obtain one at http://mozilla.org/MPL/2.0/. """ import os import sys exclusions = set(['relicense.py']) def update(text): if text.find(lgpl3_header) == -1: return text, False return text.replace(lgpl3_header, mpl2_header), True rootdir = sys.argv[1] for root, sub_folders, files in os.walk(rootdir): for basename in files: if basename in exclusions: print 'SKIPPED', filename continue filename = os.path.join(root, basename) fo = file(filename) text = fo.read() fo.close() text, updated = update(text) if updated: fo = file(filename, "w") fo.write(text) fo.close() print 'UPDATED', filename else: print ' ', filename
# Intentionally empty
# -*- coding: utf-8 -*- # This file is part of Eigen, a lightweight C++ template library # for linear algebra. # # Copyright (C) 2009 Benjamin Schindler <bschindler@inf.ethz.ch> # # This Source Code Form is subject to the terms of the Mozilla Public # License, v. 2.0. If a copy of the MPL was not distributed with this # file, You can obtain one at http://mozilla.org/MPL/2.0/. # Pretty printers for Eigen::Matrix # This is still pretty basic as the python extension to gdb is still pretty basic. # It cannot handle complex eigen types and it doesn't support any of the other eigen types # Such as quaternion or some other type. # This code supports fixed size as well as dynamic size matrices # To use it: # # * Create a directory and put the file as well as an empty __init__.py in # that directory. # * Create a ~/.gdbinit file, that contains the following: # python # import sys # sys.path.insert(0, '/path/to/eigen/printer/directory') # from printers import register_eigen_printers # register_eigen_printers (None) # end import gdb import re import itertools class EigenMatrixPrinter: "Print Eigen Matrix or Array of some kind" def __init__(self, variety, val): "Extract all the necessary information" # Save the variety (presumably "Matrix" or "Array") for later usage self.variety = variety # The gdb extension does not support value template arguments - need to extract them by hand type = val.type if type.code == gdb.TYPE_CODE_REF: type = type.target() self.type = type.unqualified().strip_typedefs() tag = self.type.tag regex = re.compile('\<.*\>') m = regex.findall(tag)[0][1:-1] template_params = m.split(',') template_params = [x.replace(" ", "") for x in template_params] if template_params[1] == '-0x00000000000000001' or template_params[1] == '-0x000000001' or template_params[1] == '-1': self.rows = val['m_storage']['m_rows'] else: self.rows = int(template_params[1]) if template_params[2] == '-0x00000000000000001' or template_params[2] == '-0x000000001' or template_params[2] == '-1': self.cols = val['m_storage']['m_cols'] else: self.cols = int(template_params[2]) self.options = 0 # default value if len(template_params) > 3: self.options = template_params[3]; self.rowMajor = (int(self.options) & 0x1) self.innerType = self.type.template_argument(0) self.val = val # Fixed size matrices have a struct as their storage, so we need to walk through this self.data = self.val['m_storage']['m_data'] if self.data.type.code == gdb.TYPE_CODE_STRUCT: self.data = self.data['array'] self.data = self.data.cast(self.innerType.pointer()) class _iterator: def __init__ (self, rows, cols, dataPtr, rowMajor): self.rows = rows self.cols = cols self.dataPtr = dataPtr self.currentRow = 0 self.currentCol = 0 self.rowMajor = rowMajor def __iter__ (self): return self def next(self): return self.__next__() # Python 2.x compatibility def __next__(self): row = self.currentRow col = self.currentCol if self.rowMajor == 0: if self.currentCol >= self.cols: raise StopIteration self.currentRow = self.currentRow + 1 if self.currentRow >= self.rows: self.currentRow = 0 self.currentCol = self.currentCol + 1 else: if self.currentRow >= self.rows: raise StopIteration self.currentCol = self.currentCol + 1 if self.currentCol >= self.cols: self.currentCol = 0 self.currentRow = self.currentRow + 1 item = self.dataPtr.dereference() self.dataPtr = self.dataPtr + 1 if (self.cols == 1): #if it's a column vector return ('[%d]' % (row,), item) elif (self.rows == 1): #if it's a row vector return ('[%d]' % (col,), item) return ('[%d,%d]' % (row, col), item) def children(self): return self._iterator(self.rows, self.cols, self.data, self.rowMajor) def to_string(self): return "Eigen::%s<%s,%d,%d,%s> (data ptr: %s)" % (self.variety, self.innerType, self.rows, self.cols, "RowMajor" if self.rowMajor else "ColMajor", self.data) class EigenQuaternionPrinter: "Print an Eigen Quaternion" def __init__(self, val): "Extract all the necessary information" # The gdb extension does not support value template arguments - need to extract them by hand type = val.type if type.code == gdb.TYPE_CODE_REF: type = type.target() self.type = type.unqualified().strip_typedefs() self.innerType = self.type.template_argument(0) self.val = val # Quaternions have a struct as their storage, so we need to walk through this self.data = self.val['m_coeffs']['m_storage']['m_data']['array'] self.data = self.data.cast(self.innerType.pointer()) class _iterator: def __init__ (self, dataPtr): self.dataPtr = dataPtr self.currentElement = 0 self.elementNames = ['x', 'y', 'z', 'w'] def __iter__ (self): return self def next(self): return self.__next__() # Python 2.x compatibility def __next__(self): element = self.currentElement if self.currentElement >= 4: #there are 4 elements in a quanternion raise StopIteration self.currentElement = self.currentElement + 1 item = self.dataPtr.dereference() self.dataPtr = self.dataPtr + 1 return ('[%s]' % (self.elementNames[element],), item) def children(self): return self._iterator(self.data) def to_string(self): return "Eigen::Quaternion<%s> (data ptr: %s)" % (self.innerType, self.data) def build_eigen_dictionary (): pretty_printers_dict[re.compile('^Eigen::Quaternion<.*>$')] = lambda val: EigenQuaternionPrinter(val) pretty_printers_dict[re.compile('^Eigen::Matrix<.*>$')] = lambda val: EigenMatrixPrinter("Matrix", val) pretty_printers_dict[re.compile('^Eigen::Array<.*>$')] = lambda val: EigenMatrixPrinter("Array", val) def register_eigen_printers(obj): "Register eigen pretty-printers with objfile Obj" if obj == None: obj = gdb obj.pretty_printers.append(lookup_function) def lookup_function(val): "Look-up and return a pretty-printer that can print va." type = val.type if type.code == gdb.TYPE_CODE_REF: type = type.target() type = type.unqualified().strip_typedefs() typename = type.tag if typename == None: return None for function in pretty_printers_dict: if function.search(typename): return pretty_printers_dict[function](val) return None pretty_printers_dict = {} build_eigen_dictionary ()
from attention_tensorflow_mesh.attention_tensorflow_mesh import transformer_lm, transformer, attention
import math import mesh_tensorflow as mtf import tensorflow.compat.v1 as tf # helpers def default(val, d): return val if val is not None else d # simple linear layer def linear(x, dim_out, scope = 'linear', bias = True): with tf.variable_scope(scope): *_, dim_in = x.shape w_init_stdev = 1 / math.sqrt(dim_in.size) return mtf.layers.dense(x, new_dims=[dim_out], reduced_dims=[dim_in], name=scope, use_bias=bias, kernel_initializer=tf.random_normal_initializer(stddev=w_init_stdev, dtype=tf.float32)) # norm def norm(x, axis = None, epsilon=1e-5): axis = default(axis, x.shape[-1]) u = mtf.reduce_mean(x, reduced_dim=axis) s = mtf.reduce_mean(mtf.square(x - u), reduced_dim=axis) u = mtf.broadcast(u, x.shape) s = mtf.broadcast(s, x.shape) return (x - u) * mtf.rsqrt(s + epsilon) def scale_norm(x, scope, *, axis=None, epsilon=1e-5, params=None): if axis is None: axis = x.shape[-1] with tf.variable_scope(scope): n_state = x.shape[-1] dt = tf.float32 g = mtf.get_variable(x.mesh, 'g', [], initializer=tf.constant_initializer(1, dtype=dt), dtype=dt) x = norm(x, axis, epsilon) x = x * g return x def prenorm(fn, scope): def inner(x, *args, **kwargs): return fn(scale_norm(x, scope), *args, **kwargs) return inner def residual(fn): def inner(x, *args, **kwargs): return fn(x, *args, **kwargs) + x return inner # full multi-head attention def attention(x, dim_head, dim_features_head, scope = 'attn', causal = False): with tf.variable_scope(scope): mesh, batch, seq, dim = x.mesh, *x.shape dim_heads = mtf.Dimension('dim_heads', dim_head.size * dim_features_head.size) dim_intermediate = mtf.Dimension('qkv_dimension', dim_heads.size * 3) qkv = linear(x, dim_intermediate, bias = False, scope='to_qkv') q, k, v = mtf.split(qkv, dim_intermediate, 3) q, k, v = map(lambda t: mtf.reshape(t, [batch, seq, dim_head, dim_features_head]), (q, k, v)) q, k, v = map(lambda t: mtf.transpose(t, [batch, dim_head, seq, dim_features_head]), (q, k, v)) k, v = map(lambda t: mtf.rename_dimension(t, seq.name, 'memory_length'), (k, v)) mem_len_dim = v.shape[-2] dots = mtf.layers.us_einsum([q, k], [batch, dim_head, seq, mem_len_dim]) if causal: i = mtf.range(mesh, seq, tf.int32) j = mtf.range(mesh, mem_len_dim, tf.int32) i, j = map(lambda t: mtf.broadcast(t, [seq, mem_len_dim]), (i, j)) mask = mtf.less(i + mem_len_dim.size - seq.size, j) mask = mtf.cast(mask, tf.float32) * -1e10 dots += mask attn = mtf.softmax(dots, mem_len_dim) out = mtf.einsum([attn, v], [batch, dim_head, seq, dim_features_head]) out = mtf.transpose(out, [batch, seq, dim_head, dim_features_head]) out = mtf.reshape(out, [batch, seq, dim_heads]) combined_out = linear(out, dim, scope='combine_output') return combined_out # feed forward def ff(x, mult = 4, scope = 'ff'): *_, dim = x.shape with tf.variable_scope(scope): dim_intermediate = mtf.Dimension('ff_intermediate', dim.size * mult) h = linear(x, dim_intermediate, scope='w1') h = mtf.gelu(h) h = linear(h, dim, scope='w2') return h # block def transformer(x, *, depth, dim_head, dim_features_head, causal = False): attn_fn = residual(prenorm(attention, 'norm1')) ff_fn = residual(prenorm(ff, 'norm2')) for i in range(depth): with tf.variable_scope(f'layer_{i}'): x = attn_fn(x, dim_head, dim_features_head, causal = causal) x = ff_fn(x) return x # language model def transformer_lm(x, *, dim, num_tokens, depth, max_seq_len, dim_head, dim_features_head, causal = False): mesh, batch, seq_dim = x.mesh, *x.shape dim = mtf.Dimension('dim', dim) dim_head = mtf.Dimension('dim_head', dim_head) dim_features_head = mtf.Dimension('dim_features_head', dim_features_head) dim_num_tokens = mtf.Dimension('vocab_size', num_tokens) dim_max_seq_len = mtf.Dimension('max_seq_len', max_seq_len) wte = mtf.get_variable(mesh, name='wte', shape=mtf.Shape([dim_num_tokens, dim]), dtype=tf.float32) wpe = mtf.get_variable(mesh, name='wpe', shape=mtf.Shape([seq_dim, dim]), dtype=tf.float32) x = mtf.gather(wte, x, dim_num_tokens) p = mtf.gather(wpe, mtf.range(mesh, seq_dim, dtype=tf.int32), dim_max_seq_len) x = x + p x = transformer(x, depth = depth, dim_head = dim_head, dim_features_head = dim_features_head, causal = causal) logits = linear(x, dim_num_tokens, scope='to_logits') return logits
# helpers def exists(val): return val is not None def lcm(*numbers): return int(functools.reduce(lambda x, y: int((x * y) / gcd(x, y)), numbers, 1)) def masked_mean(tensor, mask, dim = -1): diff_len = len(tensor.shape) - len(mask.shape) mask = mask[(..., *((None,) * diff_len))] tensor.masked_fill_(~mask, 0.) total_el = mask.sum(dim = dim) mean = tensor.sum(dim = dim) / total_el.clamp(min = 1.) mean.masked_fill_(total_el == 0, 0.) return mean def next_divisible_length(seqlen, multiple): return math.ceil(seqlen / multiple) * multiple def pad_to_multiple(tensor, multiple, *, seq_dim, dim = -1, value = 0.): seqlen = tensor.shape[seq_dim] length = next_divisible_length(seqlen, multiple) if length == seqlen: return tensor remainder = length - seqlen pad_offset = (0,) * (-1 - dim) * 2 return F.pad(tensor, (*pad_offset, 0, remainder), value = value) # helper classes class Pad(nn.Module): def __init__(self, padding, value = 0.): super().__init__() self.padding = padding self.value = value def forward(self, x): return F.pad(x, self.padding, value = self.value) class DepthwiseConv1d(nn.Module): def __init__(self, dim_in, dim_out, kernel_size): super().__init__() self.conv = nn.Conv1d(dim_in, dim_out, kernel_size, groups = dim_in) self.proj_out = nn.Conv1d(dim_out, dim_out, 1) def forward(self, x): x = self.conv(x) return self.proj_out(x) # main class class GBST(nn.Module): def __init__( self, *, num_tokens, dim, max_block_size = None, blocks = None, downsample_factor = 4, score_consensus_attn = True ): super().__init__() assert exists(max_block_size) ^ exists(blocks), 'either max_block_size or blocks are given on initialization' self.token_emb = nn.Embedding(num_tokens, dim) if exists(blocks): assert isinstance(blocks, tuple), 'blocks must be a tuple of block sizes' self.blocks = tuple(map(lambda el: el if isinstance(el, tuple) else (el, 0), blocks)) assert all([(offset < block_size) for block_size, offset in self.blocks]), 'offset must be always smaller than the block size' max_block_size = max(list(map(lambda t: t[0], self.blocks))) else: self.blocks = tuple(map(lambda el: (el, 0), range(1, max_block_size + 1))) self.pos_conv = nn.Sequential( Pad((0, 0, 0, max_block_size - 1)), Rearrange('b n d -> b d n'), DepthwiseConv1d(dim, dim, kernel_size = max_block_size), Rearrange('b d n -> b n d') ) self.score_fn = nn.Sequential( nn.Linear(dim, 1), Rearrange('... () -> ...') ) self.score_consensus_attn = score_consensus_attn assert downsample_factor <= max_block_size, 'final downsample factor should be less than the maximum block size' self.block_pad_multiple = lcm(*[block_size for block_size, _ in self.blocks]) self.downsample_factor = downsample_factor def forward(self, x, mask = None): b, n, block_mult, ds_factor, device = *x.shape, self.block_pad_multiple, self.downsample_factor, x.device m = next_divisible_length(n, ds_factor) # get character token embeddings x = self.token_emb(x) # do a conv to generate the positions for the tokens x = self.pos_conv(x) # pad both sequence and mask to length visibile by all block sizes from 0 to max block size x = pad_to_multiple(x, block_mult, seq_dim = 1, dim = -2) if exists(mask): mask = pad_to_multiple(mask, block_mult, seq_dim = 1, dim = -1, value = False) # compute representations for all blocks by mean pooling block_masks = [] block_reprs = [] for block_size, offset in self.blocks: # clone the input sequence as well as the mask, in order to pad for offsets block_x = x.clone() if exists(mask): block_mask = mask.clone() # pad for offsets, if needed need_padding = offset > 0 if need_padding: left_offset, right_offset = (block_size - offset), offset block_x = F.pad(block_x, (0, 0, left_offset, right_offset), value = 0.) if exists(mask): block_mask = F.pad(block_mask, (left_offset, right_offset), value = False) # group input sequence into blocks blocks = rearrange(block_x, 'b (n m) d -> b n m d', m = block_size) # either mean pool the blocks, or do a masked mean if exists(mask): mask_blocks = rearrange(block_mask, 'b (n m) -> b n m', m = block_size) block_repr = masked_mean(blocks, mask_blocks, dim = -2) else: block_repr = blocks.mean(dim = -2) # append the block representations, as well as the pooled block masks block_repr = repeat(block_repr, 'b n d -> b (n m) d', m = block_size) if need_padding: block_repr = block_repr[:, left_offset:-right_offset] block_reprs.append(block_repr) if exists(mask): mask_blocks = torch.any(mask_blocks, dim = -1) mask_blocks = repeat(mask_blocks, 'b n -> b (n m)', m = block_size) if need_padding: mask_blocks = mask_blocks[:, left_offset:-right_offset] block_masks.append(mask_blocks) # stack all the block representations block_reprs = torch.stack(block_reprs, dim = 2) # calculate scores and softmax across the block size dimension scores = self.score_fn(block_reprs) if exists(mask): block_masks = torch.stack(block_masks, dim = 2) max_neg_value = -torch.finfo(scores.dtype).max scores = scores.masked_fill(~block_masks, max_neg_value) scores = scores.softmax(dim = 2) # do the cheap consensus attention, eq (5) in paper if self.score_consensus_attn: score_sim = einsum('b i d, b j d -> b i j', scores, scores) if exists(mask): cross_mask = rearrange(mask, 'b i -> b i ()') * rearrange(mask, 'b j -> b () j') max_neg_value = -torch.finfo(score_sim.dtype).max score_sim = score_sim.masked_fill(~cross_mask, max_neg_value) score_attn = score_sim.softmax(dim = -1) scores = einsum('b i j, b j m -> b i m', score_attn, scores) # multiply the block representations by the position-wise scores scores = rearrange(scores, 'b n m -> b n m ()') x = (block_reprs * scores).sum(dim = 2) # truncate to length divisible by downsample factor x = x[:, :m] if exists(mask): mask = mask[:, :m] # final mean pooling downsample x = rearrange(x, 'b (n m) d -> b n m d', m = ds_factor) if exists(mask): mask = rearrange(mask, 'b (n m) -> b n m', m = ds_factor) x = masked_mean(x, mask, dim = 2) mask = torch.any(mask, dim = -1) else: x = x.mean(dim = -2) return x, mask
""" Bonito Aligner """ def align_map(aligner, sequences, n_thread=4): """ Align `sequences` with minimap using `n_thread` threads. """ return ThreadMap(partial(MappyWorker, aligner), sequences, n_thread) class MappyWorker(Thread): """ Process that reads items from an input_queue, applies a func to them and puts them on an output_queue """ def __init__(self, aligner, input_queue=None, output_queue=None): super().__init__() self.aligner = aligner self.input_queue = input_queue self.output_queue = output_queue def run(self): thrbuf = ThreadBuffer() while True: item = self.input_queue.get() if item is StopIteration: self.output_queue.put(item) break k, v = item mapping = next(self.aligner.map(v['sequence'], buf=thrbuf, MD=True), None) self.output_queue.put((k, {**v, 'mapping': mapping}))
""" Bonito Fast5 Utils """ class Read: def __init__(self, read, filename): self.read_id = read.read_id self.filename = filename.name self.run_id = read.get_run_id() if type(self.run_id) in (bytes, np.bytes_): self.run_id = self.run_id.decode() read_attrs = read.handle[read.raw_dataset_group_name].attrs channel_info = read.handle[read.global_key + 'channel_id'].attrs self.offset = int(channel_info['offset']) self.sampling_rate = channel_info['sampling_rate'] self.scaling = channel_info['range'] / channel_info['digitisation'] self.mux = read_attrs['start_mux'] self.channel = channel_info['channel_number'] if type(self.channel) in (bytes, np.bytes_): self.channel = self.channel.decode() self.start = read_attrs['start_time'] / self.sampling_rate self.duration = read_attrs['duration'] / self.sampling_rate raw = read.handle[read.raw_dataset_name][:] scaled = np.array(self.scaling * (raw + self.offset), dtype=np.float32) trim_start, _ = trim(scaled[:8000]) scaled = scaled[trim_start:] self.template_start = self.start + (1 / self.sampling_rate) * trim_start self.template_duration = self.duration - (1 / self.sampling_rate) * trim_start if len(scaled) > 8000: med, mad = med_mad(scaled) self.signal = (scaled - med) / mad else: self.signal = norm_by_noisiest_section(scaled) def __repr__(self): return "Read('%s')" % self.read_id class ReadChunk: def __init__(self, read, chunk, i, n): self.read_id = "%s:%i:%i" % (read.read_id, i, n) self.run_id = read.run_id self.filename = read.filename self.mux = read.mux self.channel = read.channel self.start = read.start self.duration = read.duration self.template_start = self.start self.template_duration = self.duration self.signal = chunk def __repr__(self): return "ReadChunk('%s')" % self.read_id def trim(signal, window_size=40, threshold_factor=2.4, min_elements=3): min_trim = 10 signal = signal[min_trim:] med, mad = med_mad(signal[-(window_size*100):]) threshold = med + mad * threshold_factor num_windows = len(signal) // window_size seen_peak = False for pos in range(num_windows): start = pos * window_size end = start + window_size window = signal[start:end] if len(window[window > threshold]) > min_elements or seen_peak: seen_peak = True if window[-1] > threshold: continue return min(end + min_trim, len(signal)), len(signal) return min_trim, len(signal) def med_mad(x, factor=1.4826): """ Calculate signal median and median absolute deviation """ med = np.median(x) mad = np.median(np.absolute(x - med)) * factor return med, mad def norm_by_noisiest_section(signal, samples=100, threshold=6.0): """ Normalise using the medmad from the longest continuous region where the noise is above some threshold relative to the std of the full signal. """ threshold = signal.std() / threshold noise = np.ones(signal.shape) for idx in np.arange(signal.shape[0] // samples): window = slice(idx * samples, (idx + 1) * samples) noise[window] = np.where(signal[window].std() > threshold, 1, 0) # start and end low for peak finding noise[0] = 0; noise[-1] = 0 peaks, info = find_peaks(noise, width=(None, None)) if len(peaks): widest = np.argmax(info['widths']) med, mad = med_mad(signal[info['left_bases'][widest]: info['right_bases'][widest]]) else: med, mad = med_mad(signal) return (signal - med) / mad def read_chunks(read, chunksize=4000, overlap=400): """ Split a Read in fixed sized ReadChunks """ if len(read.signal) < chunksize: return _, offset = divmod(len(read.signal) - chunksize, chunksize - overlap) signal = torch.from_numpy(read.signal[offset:]) blocks = signal.unfold(0, chunksize, chunksize - overlap) for i, block in enumerate(blocks): yield ReadChunk(read, block.numpy(), i+1, blocks.shape[0]) def get_raw_data(filename, read_ids=None, skip=False): """ Get the raw signal and read id from the fast5 files """ with get_fast5_file(filename, 'r') as f5_fh: for read_id in f5_fh.get_read_ids(): if read_ids is None or (read_id in read_ids) ^ skip: yield Read(f5_fh.get_read(read_id), filename) def get_read_ids(filename, read_ids=None, skip=False): """ Get all the read_ids from the file `filename`. """ with get_fast5_file(filename, 'r') as f5_fh: ids = [(filename, rid) for rid in f5_fh.get_read_ids()] if read_ids is None: return ids return [rid for rid in ids if (rid[1] in read_ids) ^ skip] def get_raw_data_for_read(info): """ Get the raw signal from the fast5 file for a given filename, read_id pair """ filename, read_id = info with get_fast5_file(filename, 'r') as f5_fh: return Read(f5_fh.get_read(read_id), filename) def get_reads(directory, read_ids=None, skip=False, max_read_size=0, n_proc=1, recursive=False, cancel=None): """ Get all reads in a given `directory`. """ pattern = "**/*.fast5" if recursive else "*.fast5" get_filtered_reads = partial(get_read_ids, read_ids=read_ids, skip=skip) with Pool(n_proc) as pool: for job in chain(pool.imap(get_filtered_reads, (Path(x) for x in glob(directory + "/" + pattern, recursive=True)))): for read in pool.imap(get_raw_data_for_read, job): if max_read_size > 0 and len(read.signal) > max_read_size: sys.stderr.write( "> skipping long read %s (%s samples)\n" % (read.read_id, len(read.signal)) ) continue yield read if cancel is not None and cancel.is_set(): return
""" Bonito utils """ try: from claragenomics.bindings import cuda from claragenomics.bindings.cudapoa import CudaPoaBatch except ImportError: pass __dir__ = os.path.dirname(os.path.realpath(__file__)) __data__ = os.path.join(__dir__, "data") __models__ = os.path.join(__dir__, "models") __configs__ = os.path.join(__dir__, "models/configs") split_cigar = re.compile(r"(?P<len>\d+)(?P<op>\D+)") default_data = os.path.join(__data__, "dna_r9.4.1") default_config = os.path.join(__configs__, "dna_r9.4.1@v3.1.toml") def init(seed, device): """ Initialise random libs and setup cudnn https://pytorch.org/docs/stable/notes/randomness.html """ random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if device == "cpu": return torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False assert(torch.cuda.is_available()) def permute(x, input_layout, output_layout): """ Permute `x` from `input_layout` to `output_layout` >>> permute(x, 'TNC', 'NTC') """ if input_layout == output_layout: return x return x.permute(*[input_layout.index(x) for x in output_layout]) def concat(xs, dim=0): """ Type agnostic concat. """ if isinstance(xs[0], torch.Tensor): return torch.cat(xs, dim=dim) elif isinstance(xs[0], np.ndarray): return np.concatenate(xs, axis=dim) elif isinstance(xs[0], list): return [x for l in xs for x in l] elif isinstance(xs[0], str): return ''.join(xs) elif isinstance(xs[0], dict): return {k: concat([x[k] for x in xs], dim) for k in xs[0].keys()} else: raise TypeError def select_range(x, start, end, dim=0): """ Type agnostic range select. """ if isinstance(x, dict): return {k: select_range(v, start, end, dim) for (k, v) in x.items()} if dim == 0 or isinstance(x, list): return x[start:end] return x[(*(slice(None),)*dim, slice(start, end))] def size(x, dim=0): """ Type agnostic size. """ if hasattr(x, 'shape'): return x.shape[dim] elif dim == 0: return len(x) raise TypeError def half_supported(): """ Returns whether FP16 is support on the GPU """ try: return get_device_capability()[0] >= 7 except: return False def phred(prob, scale=1.0, bias=0.0): """ Converts `prob` into a ascii encoded phred quality score between 0 and 40. """ p = max(1 - prob, 1e-4) q = -10 * np.log10(p) * scale + bias return chr(int(np.round(q) + 33)) def mean_qscore_from_qstring(qstring): """ Convert qstring into a mean qscore """ if len(qstring) == 0: return 0.0 err_probs = [10**((ord(c) - 33) / -10) for c in qstring] mean_err = np.mean(err_probs) return -10 * np.log10(max(mean_err, 1e-4)) def decode_ref(encoded, labels): """ Convert a integer encoded reference into a string and remove blanks """ return ''.join(labels[e] for e in encoded if e) def column_to_set(filename, idx=0, skip_header=False): """ Pull a column from a file and return a set of the values. """ if filename and os.path.isfile(filename): with open(filename, 'r') as tsv: if skip_header: next(tsv) return {line.strip().split()[idx] for line in tsv.readlines()} def chunk(signal, chunksize, overlap): """ Convert a read into overlapping chunks before calling """ T = signal.shape[0] if chunksize == 0: chunks = signal[None, :] elif T < chunksize: chunks = torch.nn.functional.pad(signal, (chunksize - T, 0))[None, :] else: stub = (T - overlap) % (chunksize - overlap) chunks = signal[stub:].unfold(0, chunksize, chunksize - overlap) if stub > 0: chunks = torch.cat([signal[None, :chunksize], chunks], dim=0) return chunks.unsqueeze(1) def stitch(chunks, chunksize, overlap, length, stride, reverse=False): """ Stitch chunks together with a given overlap """ if chunks.shape[0] == 1: return chunks.squeeze(0) semi_overlap = overlap // 2 start, end = semi_overlap // stride, (chunksize - semi_overlap) // stride stub = (length - overlap) % (chunksize - overlap) first_chunk_end = (stub + semi_overlap) // stride if (stub > 0) else end if reverse: chunks = list(chunks) return concat([ chunks[-1][:-start], *(x[-end:-start] for x in reversed(chunks[1:-1])), chunks[0][-first_chunk_end:] ]) else: return concat([ chunks[0, :first_chunk_end], *chunks[1:-1, start:end], chunks[-1, start:] ]) def batchify(items, batchsize, dim=0): """ Batch up items up to `batch_size`. """ stack, pos = [], 0 for k, v in items: breaks = range(batchsize - pos, size(v, dim), batchsize) for start, end in zip([0, *breaks], [*breaks, size(v, dim)]): sub_batch = select_range(v, start, end, dim) stack.append(((k, (pos, pos + end - start)), sub_batch)) if pos + end - start == batchsize: ks, vs = zip(*stack) yield ks, concat(vs, dim) stack, pos = [], 0 else: pos += end - start if len(stack): ks, vs = zip(*stack) yield ks, concat(vs, dim) def unbatchify(batches, dim=0): """ Reconstruct batches. """ batches = ( (k, select_range(v, start, end, dim)) for sub_batches, v in batches for k, (start, end) in sub_batches ) return ( (k, concat([v for (k, v) in group], dim)) for k, group in groupby(batches, itemgetter(0)) ) def load_data(limit=None, directory=None): """ Load the training data """ if directory is None: directory = default_data chunks = np.load(os.path.join(directory, "chunks.npy"), mmap_mode='r') targets = np.load(os.path.join(directory, "references.npy"), mmap_mode='r') lengths = np.load(os.path.join(directory, "reference_lengths.npy"), mmap_mode='r') indices = os.path.join(directory, "indices.npy") if os.path.exists(indices): idx = np.load(indices, mmap_mode='r') idx = idx[idx < lengths.shape[0]] if limit: idx = idx[:limit] return chunks[idx, :], targets[idx, :], lengths[idx] if limit: chunks = chunks[:limit] targets = targets[:limit] lengths = lengths[:limit] return np.array(chunks), np.array(targets), np.array(lengths) def load_symbol(config, symbol): """ Dynamic load a symbol from module specified in model config. """ if not isinstance(config, dict): if not os.path.isdir(config) and os.path.isdir(os.path.join(__models__, config)): dirname = os.path.join(__models__, config) else: dirname = config config = toml.load(os.path.join(dirname, 'config.toml')) imported = import_module(config['model']['package']) return getattr(imported, symbol) def match_names(state_dict, model): keys_and_shapes = lambda state_dict: zip(*[ (k, s) for s, i, k in sorted([(v.shape, i, k) for i, (k, v) in enumerate(state_dict.items())]) ]) k1, s1 = keys_and_shapes(state_dict) k2, s2 = keys_and_shapes(model.state_dict()) assert s1 == s2 remap = dict(zip(k1, k2)) return OrderedDict([(k, remap[k]) for k in state_dict.keys()]) def load_model(dirname, device, weights=None, half=None, chunksize=0): """ Load a model from disk """ if not os.path.isdir(dirname) and os.path.isdir(os.path.join(__models__, dirname)): dirname = os.path.join(__models__, dirname) if not weights: # take the latest checkpoint weight_files = glob(os.path.join(dirname, "weights_*.tar")) if not weight_files: raise FileNotFoundError("no model weights found in '%s'" % dirname) weights = max([int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in weight_files]) device = torch.device(device) config = toml.load(os.path.join(dirname, 'config.toml')) weights = os.path.join(dirname, 'weights_%s.tar' % weights) Model = load_symbol(config, "Model") model = Model(config) state_dict = torch.load(weights, map_location=device) state_dict = {k2: state_dict[k1] for k1, k2 in match_names(state_dict, model).items()} new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k.replace('module.', '') new_state_dict[name] = v model.load_state_dict(new_state_dict) if half is None: half = half_supported() if half: model = model.half() model.eval() model.to(device) return model def parasail_to_sam(result, seq): """ Extract reference start and sam compatible cigar string. :param result: parasail alignment result. :param seq: query sequence. :returns: reference start coordinate, cigar string. """ cigstr = result.cigar.decode.decode() first = re.search(split_cigar, cigstr) first_count, first_op = first.groups() prefix = first.group() rstart = result.cigar.beg_ref cliplen = result.cigar.beg_query clip = '' if cliplen == 0 else '{}S'.format(cliplen) if first_op == 'I': pre = '{}S'.format(int(first_count) + cliplen) elif first_op == 'D': pre = clip rstart = int(first_count) else: pre = '{}{}'.format(clip, prefix) mid = cigstr[len(prefix):] end_clip = len(seq) - result.end_query - 1 suf = '{}S'.format(end_clip) if end_clip > 0 else '' new_cigstr = ''.join((pre, mid, suf)) return rstart, new_cigstr def accuracy(ref, seq, balanced=False, min_coverage=0.0): """ Calculate the accuracy between `ref` and `seq` """ alignment = parasail.sw_trace_striped_32(seq, ref, 8, 4, parasail.dnafull) counts = defaultdict(int) q_coverage = len(alignment.traceback.query) / len(seq) r_coverage = len(alignment.traceback.ref) / len(ref) if r_coverage < min_coverage: return 0.0 _, cigar = parasail_to_sam(alignment, seq) for count, op in re.findall(split_cigar, cigar): counts[op] += int(count) if balanced: accuracy = (counts['='] - counts['I']) / (counts['='] + counts['X'] + counts['D']) else: accuracy = counts['='] / (counts['='] + counts['I'] + counts['X'] + counts['D']) return accuracy * 100 def print_alignment(ref, seq): """ Print the alignment between `ref` and `seq` """ alignment = parasail.sw_trace_striped_32(seq, ref, 8, 4, parasail.dnafull) print(alignment.traceback.ref) print(alignment.traceback.comp) print(alignment.traceback.query) print(" Score=%s" % alignment.score) return alignment.score def poa(groups, max_poa_sequences=100, gpu_mem_per_batch=0.9): """ Generate consensus for POA groups. Args: groups : A list of lists of sequences for which consensus is to be generated. """ free, total = cuda.cuda_get_mem_info(cuda.cuda_get_device()) gpu_mem_per_batch *= free batch = CudaPoaBatch(max_poa_sequences, gpu_mem_per_batch, stream=None, output_type="consensus") results = [] for i, group in enumerate(groups, start=1): group_status, seq_status = batch.add_poa_group(group) # Once batch is full, run POA processing if group_status == 1 or i == len(groups): batch.generate_poa() consensus, coverage, status = batch.get_consensus() results.extend(consensus) batch.reset() group_status, seq_status = batch.add_poa_group(group) return results
""" Bonito nn modules. """ layers = {} def register(layer): layer.name = layer.__name__.lower() layers[layer.name] = layer return layer register(torch.nn.ReLU) register(torch.nn.Tanh) @register class Swish(torch.nn.SiLU): pass @register class Serial(torch.nn.Sequential): def __init__(self, sublayers): super().__init__(*sublayers) def to_dict(self, include_weights=False): return { 'sublayers': [to_dict(layer, include_weights) for layer in self._modules.values()] } @register class Reverse(Module): def __init__(self, sublayers): super().__init__() self.layer = Serial(sublayers) if isinstance(sublayers, list) else sublayers def forward(self, x): return self.layer(x.flip(0)).flip(0) def to_dict(self, include_weights=False): if isinstance(self.layer, Serial): return self.layer.to_dict(include_weights) else: return {'sublayers': to_dict(self.layer, include_weights)} @register class Convolution(Module): def __init__(self, insize, size, winlen, stride=1, padding=0, bias=True, activation=None): super().__init__() self.conv = torch.nn.Conv1d(insize, size, winlen, stride=stride, padding=padding, bias=bias) self.activation = layers.get(activation, lambda: activation)() def forward(self, x): if self.activation is not None: return self.activation(self.conv(x)) return self.conv(x) def to_dict(self, include_weights=False): res = { "insize": self.conv.in_channels, "size": self.conv.out_channels, "bias": self.conv.bias is not None, "winlen": self.conv.kernel_size[0], "stride": self.conv.stride[0], "padding": self.conv.padding[0], "activation": self.activation.name if self.activation else None, } if include_weights: res['params'] = { 'W': self.conv.weight, 'b': self.conv.bias if self.conv.bias is not None else [] } return res @register class LinearCRFEncoder(Module): def __init__(self, insize, n_base, state_len, bias=True, scale=None, activation=None, blank_score=None): super().__init__() self.n_base = n_base self.state_len = state_len self.blank_score = blank_score size = (n_base + 1) * n_base**state_len if blank_score is None else n_base**(state_len + 1) self.linear = torch.nn.Linear(insize, size, bias=bias) self.activation = layers.get(activation, lambda: activation)() self.scale = scale def forward(self, x): scores = self.linear(x) if self.activation is not None: scores = self.activation(scores) if self.scale is not None: scores = scores * self.scale if self.blank_score is not None: T, N, C = scores.shape s = torch.tensor(self.blank_score, device=scores.device, dtype=scores.dtype) scores = torch.cat([s.expand(T, N, C//self.n_base, 1), scores.reshape(T, N, C//self.n_base, self.n_base)], axis=-1).reshape(T, N, -1) return scores def to_dict(self, include_weights=False): res = { 'insize': self.linear.in_features, 'n_base': self.n_base, 'state_len': self.state_len, 'bias': self.linear.bias is not None, 'scale': self.scale, 'activation': self.activation.name if self.activation else None, 'blank_score': self.blank_score, } if include_weights: res['params'] = { 'W': self.linear.weight, 'b': self.linear.bias if self.linear.bias is not None else [] } return res @register class SHA(Module): def __init__(self, dim): super().__init__() self.scale = dim ** -0.5 self.to_q = nn.Sequential(nn.Linear(dim, dim), nn.LayerNorm(dim)) def forward(self, x, kv): x = x.transpose(0, 1) kv = kv.transpose(0, 1) q = self.to_q(x) sim = torch.matmul(q, kv.transpose(-1, -2)) * self.scale attn = sim.softmax(dim=-1) out = torch.matmul(attn, kv) return out.transpose(0, 1) @register class SHABlock(Module): """ https://arxiv.org/abs/1911.11423 """ def __init__(self, dim, ff_mult=4): super().__init__() self.attn_query_norm = nn.LayerNorm(dim) self.attn_kv_norm = nn.LayerNorm(dim) self.attn = SHA(dim=dim) self.ff_residual_norm = nn.LayerNorm(dim) self.ff = Serial([ nn.LayerNorm(dim), nn.Linear(dim, dim * ff_mult), nn.GELU(), nn.Linear(dim * ff_mult, dim), ]) def forward(self, x): kv = self.attn_kv_norm(x) x = self.attn_query_norm(x) x = self.attn(x, kv) + x x = self.ff(x) + self.ff_residual_norm(x) return x @register class Permute(Module): def __init__(self, dims): super().__init__() self.dims = dims def forward(self, x): return x.permute(*self.dims) def to_dict(self, include_weights=False): return {'dims': self.dims} def truncated_normal(size, dtype=torch.float32, device=None, num_resample=5): x = torch.empty(size + (num_resample,), dtype=torch.float32, device=device).normal_() i = ((x < 2) & (x > -2)).max(-1, keepdim=True)[1] return torch.clamp_(x.gather(-1, i).squeeze(-1), -2, 2) class RNNWrapper(Module): def __init__( self, rnn_type, *args, reverse=False, orthogonal_weight_init=True, disable_state_bias=True, bidirectional=False, **kwargs ): super().__init__() if reverse and bidirectional: raise Exception("'reverse' and 'bidirectional' should not both be set to True") self.reverse = reverse self.rnn = rnn_type(*args, bidirectional=bidirectional, **kwargs) self.init_orthogonal(orthogonal_weight_init) self.init_biases() if disable_state_bias: self.disable_state_bias() def forward(self, x): if self.reverse: x = x.flip(0) y, h = self.rnn(x) if self.reverse: y = y.flip(0) return y def init_biases(self, types=('bias_ih',)): for name, param in self.rnn.named_parameters(): if any(k in name for k in types): with torch.no_grad(): param.set_(0.5*truncated_normal(param.shape, dtype=param.dtype, device=param.device)) def init_orthogonal(self, types=True): if not types: return if types == True: types = ('weight_ih', 'weight_hh') for name, x in self.rnn.named_parameters(): if any(k in name for k in types): for i in range(0, x.size(0), self.rnn.hidden_size): orthogonal_(x[i:i+self.rnn.hidden_size]) def disable_state_bias(self): for name, x in self.rnn.named_parameters(): if 'bias_hh' in name: x.requires_grad = False x.zero_() @register class LSTM(RNNWrapper): def __init__(self, size, insize, bias=True, reverse=False): super().__init__(torch.nn.LSTM, size, insize, bias=bias, reverse=reverse) def to_dict(self, include_weights=False): res = { 'size': self.rnn.hidden_size, 'insize': self.rnn.input_size, 'bias': self.rnn.bias, 'reverse': self.reverse, } if include_weights: res['params'] = { 'iW': self.rnn.weight_ih_l0.reshape(4, self.rnn.hidden_size, self.rnn.input_size), 'sW': self.rnn.weight_hh_l0.reshape(4, self.rnn.hidden_size, self.rnn.hidden_size), 'b': self.rnn.bias_ih_l0.reshape(4, self.rnn.hidden_size) } return res def to_dict(layer, include_weights=False): if hasattr(layer, 'to_dict'): return {'type': layer.name, **layer.to_dict(include_weights)} return {'type': layer.name} def from_dict(model_dict, layer_types=None): model_dict = model_dict.copy() if layer_types is None: layer_types = layers type_name = model_dict.pop('type') typ = layer_types[type_name] if 'sublayers' in model_dict: sublayers = model_dict['sublayers'] model_dict['sublayers'] = [ from_dict(x, layer_types) for x in sublayers ] if isinstance(sublayers, list) else from_dict(sublayers, layer_types) try: layer = typ(**model_dict) except Exception as e: raise Exception(f'Failed to build layer of type {typ} with args {model_dict}') from e return layer
""" Bonito Input/Output """ logger = getLogger('bonito') class CSVLogger: def __init__(self, filename, sep=','): self.filename = str(filename) if os.path.exists(self.filename): with open(self.filename) as f: self.columns = csv.DictReader(f).fieldnames else: self.columns = None self.fh = open(self.filename, 'a', newline='') self.csvwriter = csv.writer(self.fh, delimiter=sep) self.count = 0 def set_columns(self, columns): if self.columns: raise Exception('Columns already set') self.columns = list(columns) self.csvwriter.writerow(self.columns) def append(self, row): if self.columns is None: self.set_columns(row.keys()) self.csvwriter.writerow([row.get(k, '-') for k in self.columns]) self.count += 1 if self.count > 100: self.count = 0 self.fh.flush() def close(self): self.fh.close() def __enter__(self): return self def __exit__(self, *args): self.close() @contextmanager def devnull(*args, **kwds): """ A context manager that sends all out stdout & stderr to devnull. """ save_fds = [os.dup(1), os.dup(2)] null_fds = [os.open(os.devnull, os.O_RDWR) for _ in range(2)] os.dup2(null_fds[0], 1) os.dup2(null_fds[1], 2) try: yield finally: os.dup2(save_fds[0], 1) os.dup2(save_fds[1], 2) for fd in null_fds + save_fds: os.close(fd) def write_fasta(header, sequence, fd=sys.stdout): """ Write a fasta record to a file descriptor. """ fd.write(">%s\n" % header) fd.write("%s\n" % sequence) fd.flush() def write_fastq(header, sequence, qstring, fd=sys.stdout): """ Write a fastq record to a file descriptor. """ fd.write("@%s\n" % header) fd.write("%s\n" % sequence) fd.write("+\n") fd.write("%s\n" % qstring) fd.flush() def write_sam_header(aligner, fd=sys.stdout, sep='\t'): """ Write the SQ & PG sam headers to a file descriptor. """ fd.write('%s\n' % os.linesep.join([ sep.join([ '@SQ', 'SN:%s' % name, 'LN:%s' % len(aligner.seq(name)) ]) for name in aligner.seq_names ])) fd.write('%s\n' % sep.join([ '@PG', 'ID:bonito', 'PN:bonito', 'VN:%s' % bonito.__version__, 'CL:%s' % ' '.join(sys.argv), ])) fd.flush() def write_sam(read_id, sequence, qstring, mapping, fd=sys.stdout, unaligned=False, sep='\t'): """ Write a sam record to a file descriptor. """ if unaligned: fd.write("%s\n" % sep.join(map(str, [ read_id, 4, '*', 0, 0, '*', '*', 0, 0, sequence, qstring, 'NM:i:0' ]))) else: softclip = [ '%sS' % mapping.q_st if mapping.q_st else '', mapping.cigar_str, '%sS' % (len(sequence) - mapping.q_en) if len(sequence) - mapping.q_en else '' ] fd.write("%s\n" % sep.join(map(str, [ read_id, 0 if mapping.strand == +1 else 16, mapping.ctg, mapping.r_st + 1, mapping.mapq, ''.join(softclip if mapping.strand == +1 else softclip[::-1]), '*', 0, 0, sequence if mapping.strand == +1 else revcomp(sequence), qstring, 'NM:i:%s' % mapping.NM, 'MD:Z:%s' % mapping.MD, ]))) fd.flush() def summary_file(): """ Return the filename to use for the summary tsv. """ stdout = realpath('/dev/fd/1') if sys.stdout.isatty() or stdout.startswith('/proc'): return 'summary.tsv' return '%s_summary.tsv' % splitext(stdout)[0] summary_field_names = [ 'filename', 'read_id', 'run_id', 'channel', 'mux', 'start_time', 'duration', 'template_start', 'template_duration', 'sequence_length_template', 'mean_qscore_template', #if alignment 'alignment_genome', 'alignment_genome_start', 'alignment_genome_end', 'alignment_strand_start', 'alignment_strand_end', 'alignment_direction', 'alignment_length', 'alignment_num_aligned', 'alignment_num_correct', 'alignment_num_insertions', 'alignment_num_deletions', 'alignment_num_substitutions', 'alignment_mapq', 'alignment_strand_coverage', 'alignment_identity', 'alignment_accuracy', ] def summary_row(read, seqlen, qscore, alignment=False): """ Summary tsv row. """ fields = [ read.filename, read.read_id, read.run_id, read.channel, read.mux, read.start, read.duration, read.template_start, read.template_duration, seqlen, qscore, ] if alignment: ins = sum(count for count, op in alignment.cigar if op == 1) dels = sum(count for count, op in alignment.cigar if op == 2) subs = alignment.NM - ins - dels length = alignment.blen matches = length - ins - dels correct = alignment.mlen fields.extend([ alignment.ctg, alignment.r_st, alignment.r_en, alignment.q_st if alignment.strand == +1 else seqlen - alignment.q_en, alignment.q_en if alignment.strand == +1 else seqlen - alignment.q_st, '+' if alignment.strand == +1 else '-', length, matches, correct, ins, dels, subs, alignment.mapq, (alignment.q_en - alignment.q_st) / seqlen, correct / matches, correct / length, ]) elif alignment is None: fields.extend( ['*', -1, -1, -1, -1, '*', 0, 0, 0, 0, 0, 0, 0, 0.0, 0.0, 0.0] ) return dict(zip(summary_field_names, fields)) duplex_summary_field_names = [ 'filename_template', 'read_id_template', 'filename_complement', 'read_id_complement', 'run_id', 'channel_template', 'mux_template', 'channel_complement', 'mux_complement', 'sequence_length_duplex', 'mean_qscore_duplex', #if alignment 'alignment_genome', 'alignment_genome_start', 'alignment_genome_end', 'alignment_strand_start', 'alignment_strand_end', 'alignment_direction', 'alignment_length', 'alignment_num_aligned', 'alignment_num_correct', 'alignment_num_insertions', 'alignment_num_deletions', 'alignment_num_substitutions', 'alignment_mapq', 'alignment_strand_coverage', 'alignment_identity', 'alignment_accuracy', ] def duplex_summary_row(read_temp, comp_read, seqlen, qscore, alignment=False): """ Duplex summary tsv row. """ fields = [ read_temp.filename, read_temp.read_id, comp_read.filename, comp_read.read_id, read_temp.run_id, read_temp.channel, read_temp.mux, comp_read.channel, comp_read.mux, seqlen, qscore, ] if alignment: ins = sum(count for count, op in alignment.cigar if op == 1) dels = sum(count for count, op in alignment.cigar if op == 2) subs = alignment.NM - ins - dels length = alignment.blen matches = length - ins - dels correct = alignment.mlen fields.extend([ alignment.ctg, alignment.r_st, alignment.r_en, alignment.q_st if alignment.strand == +1 else seqlen - alignment.q_en, alignment.q_en if alignment.strand == +1 else seqlen - alignment.q_st, '+' if alignment.strand == +1 else '-', length, matches, correct, ins, dels, subs, alignment.mapq, (alignment.q_en - alignment.q_st) / seqlen, correct / matches, correct / length, ]) elif alignment is None: fields.extend( ['*', -1, -1, -1, -1, '*', 0, 0, 0, 0, 0, 0, 0, 0.0, 0.0, 0.0] ) return dict(zip(duplex_summary_field_names, fields)) class Writer(Thread): def __init__(self, iterator, aligner, fd=sys.stdout, fastq=False, duplex=False): super().__init__() self.fd = fd self.log = [] self.fastq = fastq self.duplex = duplex self.aligner = aligner self.iterator = iterator self.write_headers() def write_headers(self): if self.aligner: write_sam_header(self.aligner, fd=self.fd) def run(self): with CSVLogger(summary_file(), sep='\t') as summary: for read, res in self.iterator: seq = res['sequence'] qstring = res.get('qstring', '*') mean_qscore = res.get('mean_qscore', 0.0) mapping = res.get('mapping', False) if self.duplex: samples = len(read[0].signal) + len(read[1].signal) read_id = '%s;%s' % (read[0].read_id, read[1].read_id) else: samples = len(read.signal) read_id = read.read_id if len(seq): if self.aligner: write_sam(read_id, seq, qstring, mapping, fd=self.fd, unaligned=mapping is None) else: if self.fastq: write_fastq(read_id, seq, qstring, fd=self.fd) else: write_fasta(read_id, seq, fd=self.fd) if self.duplex: summary.append(duplex_summary_row(read[0], read[1], len(seq), mean_qscore, alignment=mapping)) else: summary.append(summary_row(read, len(seq), mean_qscore, alignment=mapping)) self.log.append((read_id, samples)) else: logger.warn("> skipping empty sequence %s", read_id) class CTCWriter(Thread): """ CTC writer process that writes output numpy training data. """ def __init__(self, iterator, aligner, min_coverage, min_accuracy, fd=sys.stdout): super().__init__() self.fd = fd self.log = [] self.aligner = aligner self.iterator = iterator self.min_coverage = min_coverage self.min_accuracy = min_accuracy self.write_headers() def write_headers(self): if self.aligner: write_sam_header(self.aligner, fd=self.fd) def run(self): chunks = [] targets = [] lengths = [] with CSVLogger(summary_file(), sep='\t') as summary: for read, ctc_data in self.iterator: seq = ctc_data['sequence'] qstring = ctc_data['qstring'] mean_qscore = ctc_data['mean_qscore'] mapping = ctc_data.get('mapping', False) self.log.append((read.read_id, len(read.signal))) if len(seq) == 0 or mapping is None: continue cov = (mapping.q_en - mapping.q_st) / len(seq) acc = mapping.mlen / mapping.blen refseq = self.aligner.seq(mapping.ctg, mapping.r_st, mapping.r_en) if acc < self.min_accuracy or cov < self.min_coverage or 'N' in refseq: continue write_sam(read.read_id, seq, qstring, mapping, fd=self.fd, unaligned=mapping is None) summary.append(summary_row(read, len(seq), mean_qscore, alignment=mapping)) if mapping.strand == -1: refseq = revcomp(refseq) target = [int(x) for x in refseq.translate({65: '1', 67: '2', 71: '3', 84: '4'})] targets.append(target) chunks.append(read.signal) lengths.append(len(target)) if len(chunks) == 0: sys.stderr.write("> no suitable ctc data to write\n") return chunks = np.array(chunks, dtype=np.float16) targets_ = np.zeros((chunks.shape[0], max(lengths)), dtype=np.uint8) for idx, target in enumerate(targets): targets_[idx, :len(target)] = target lengths = np.array(lengths, dtype=np.uint16) indices = np.random.permutation(typical_indices(lengths)) chunks = chunks[indices] targets_ = targets_[indices] lengths = lengths[indices] summary = pd.read_csv(summary_file(), sep='\t') summary.iloc[indices].to_csv(summary_file(), sep='\t', index=False) output_directory = '.' if sys.stdout.isatty() else dirname(realpath('/dev/fd/1')) np.save(os.path.join(output_directory, "chunks.npy"), chunks) np.save(os.path.join(output_directory, "references.npy"), targets_) np.save(os.path.join(output_directory, "reference_lengths.npy"), lengths) sys.stderr.write("> written ctc training data\n") sys.stderr.write(" - chunks.npy with shape (%s)\n" % ','.join(map(str, chunks.shape))) sys.stderr.write(" - references.npy with shape (%s)\n" % ','.join(map(str, targets_.shape))) sys.stderr.write(" - reference_lengths.npy shape (%s)\n" % ','.join(map(str, lengths.shape))) def stop(self): self.join()
modules = [ 'basecaller', 'train', 'evaluate', 'view', 'convert', 'download', 'export', 'duplex', ] __version__ = '0.4.0' def main(): parser = ArgumentParser( 'bonito', formatter_class=ArgumentDefaultsHelpFormatter ) parser.add_argument( '-v', '--version', action='version', version='%(prog)s {}'.format(__version__) ) subparsers = parser.add_subparsers( title='subcommands', description='valid commands', help='additional help', dest='command' ) subparsers.required = True for module in modules: mod = globals()[module] p = subparsers.add_parser(module, parents=[mod.argparser()]) p.set_defaults(func=mod.main) args = parser.parse_args() args.func(args)
""" Bonito Multiprocesing """ def process_iter(iterator, maxsize=1): """ Take an iterator and run it on another process. """ return iter(ProcessIterator(iterator, maxsize=maxsize)) def thread_iter(iterator, maxsize=1): """ Take an iterator and run it on another thread. """ return iter(ThreadIterator(iterator, maxsize=maxsize)) def process_cancel(): """ Register an cancel event on sigint """ event = Event() signal(SIGINT, lambda *a: event.set()) return event def process_map(func, iterator, n_proc=4, maxsize=0): """ Take an `iterator` of key, value pairs and apply `func` to all values using `n_proc` processes. """ if n_proc == 0: return ((k, func(v)) for k, v in iterator) return iter(ProcessMap(func, iterator, n_proc, output_queue=Queue(maxsize))) def thread_map(func, iterator, n_thread=4, maxsize=2): """ Take an `iterator` of key, value pairs and apply `func` to all values using `n_thread` threads. """ if n_thread == 0: return ((k, func(v)) for k, v in iterator) return iter(ThreadMap(partial(MapWorkerThread, func), iterator, n_thread, maxsize=maxsize)) class BackgroundIterator: """ Runs an iterator in the background. """ def __init__(self, iterator, maxsize=10): super().__init__() self.iterator = iterator self.queue = self.QueueClass(maxsize) def __iter__(self): self.start() while True: item = self.queue.get() if item is StopIteration: break yield item def run(self): for item in self.iterator: self.queue.put(item) self.queue.put(StopIteration) def stop(self): self.join() class ThreadIterator(BackgroundIterator, Thread): """ Runs an iterator in a separate process. """ QueueClass = queue.Queue class ProcessIterator(BackgroundIterator, Process): """ Runs an iterator in a separate process. """ QueueClass = Queue class MapWorker(Process): """ Process that reads items from an input_queue, applies a func to them and puts them on an output_queue """ def __init__(self, func, input_queue, output_queue): super().__init__() self.func = func self.input_queue = input_queue self.output_queue = output_queue def run(self): while True: item = self.input_queue.get() if item is StopIteration: break k, v = item self.output_queue.put((k, self.func(v))) class ProcessMap(Thread): def __init__(self, func, iterator, n_proc, output_queue=None): super().__init__() self.key_map = {} self.iterator = iterator self.work_queue = Queue(n_proc * 2) self.output_queue = output_queue or Queue() self.processes = [MapWorker(func, self.work_queue, self.output_queue) for _ in range(n_proc)] def start(self): for process in self.processes: process.start() super().start() def run(self): for (k, v) in self.iterator: self.work_queue.put((id(k), v)) self.key_map[id(k)] = k for _ in self.processes: self.work_queue.put(StopIteration) for process in self.processes: process.join() self.output_queue.put(StopIteration) def __iter__(self): self.start() while True: item = self.output_queue.get() if item is StopIteration: break k, v = item yield self.key_map.pop(k), v class MapWorkerThread(Thread): """ Process that reads items from an input_queue, applies a func to them and puts them on an output_queue """ def __init__(self, func, input_queue=None, output_queue=None): super().__init__() self.func = func self.input_queue = input_queue self.output_queue = output_queue def run(self): while True: item = self.input_queue.get() if item is StopIteration: self.output_queue.put(item) break k, v = item self.output_queue.put((k, self.func(v))) class ThreadMap(Thread): def __init__(self, worker_type, iterator, n_thread, maxsize=2): super().__init__() self.iterator = iterator self.n_thread = n_thread self.work_queues = [queue.Queue(maxsize) for _ in range(n_thread)] self.output_queues = [queue.Queue(maxsize) for _ in range(n_thread)] self.workers = [worker_type(input_queue=in_q, output_queue=out_q) for (in_q, out_q) in zip(self.work_queues, self.output_queues)] def start(self): for worker in self.workers: worker.start() super().start() def __iter__(self): self.start() for i in count(): item = self.output_queues[i % self.n_thread].get() if item is StopIteration: #do we need to empty output_queues in order to join worker threads? for j in range(i + 1, i + self.n_thread): self.output_queues[j % self.n_thread].get() break yield item def run(self): for i, (k, v) in enumerate(self.iterator): self.work_queues[i % self.n_thread].put((k, v)) for q in self.work_queues: q.put(StopIteration) for worker in self.workers: worker.join()
""" Bonito train """ class ChunkDataSet: def __init__(self, chunks, targets, lengths): self.chunks = np.expand_dims(chunks, axis=1) self.targets = targets self.lengths = lengths def __getitem__(self, i): return ( self.chunks[i].astype(np.float32), self.targets[i].astype(np.int64), self.lengths[i].astype(np.int64), ) def __len__(self): return len(self.lengths) def const_schedule(y): """ Constant Scheduler """ return lambda t: y def linear_schedule(y0, y1): """ Linear Scheduler """ return lambda t: y0 + (y1 - y0) * t def cosine_decay_schedule(y0, y1): """ Cosine Decay Scheduler """ return lambda t: y1 + 0.5 * (y0 - y1) * (np.cos(t * np.pi) + 1.0) def piecewise_schedule(knots, funcs): """ Piecewise Scheduler """ def f(t): i = np.searchsorted(knots, t) t0 = 0.0 if i == 0 else knots[i - 1] t1 = 1.0 if i == len(knots) else knots[i] return funcs[i]((t - t0) / (t1 - t0)) return f def func_scheduler(optimizer, func, total_steps, warmup_steps=None, warmup_ratio=0.1, start_step=0): """ Learning Rate Scheduler """ if warmup_steps: y0 = func(0.0) func = piecewise_schedule( [warmup_steps / total_steps], [linear_schedule(warmup_ratio * y0, y0), func] ) return LambdaLR(optimizer, (lambda step: func((step + start_step) / total_steps))) def load_state(dirname, device, model): """ Load a model state dict from disk """ model.to(device) weight_no = None weight_files = glob(os.path.join(dirname, "weights_*.tar")) if weight_files: weight_no = max([int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in weight_files]) if weight_no: print("[picking up from epoch %s]" % weight_no) state_dict = torch.load( os.path.join(dirname, 'weights_%s.tar' % weight_no), map_location=device ) state_dict = {k2: state_dict[k1] for k1, k2 in match_names(state_dict, model).items()} new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k.replace('module.', '') new_state_dict[name] = v model.load_state_dict(new_state_dict) epoch = weight_no else: epoch = 0 return epoch class Trainer: def __init__(self, model, device, train_loader, valid_loader, criterion=None, use_amp=True): self.model = model.to(device) self.device = device self.train_loader = train_loader self.valid_loader = valid_loader self.criterion = criterion or (model.seqdist.ctc_loss if hasattr(model, 'seqdist') else model.ctc_label_smoothing_loss) self.use_amp = use_amp self.scaler = torch.cuda.amp.GradScaler(enabled=use_amp) self.optimizer = None def train_one_step(self, batch): data, targets, lengths = batch self.optimizer.zero_grad() with amp.autocast(enabled=self.use_amp): scores = self.model(data.to(self.device)) losses = self.criterion(scores, targets.to(self.device), lengths.to(self.device)) if not isinstance(losses, dict): losses = {'loss': losses} self.scaler.scale(losses['loss']).backward() self.scaler.unscale_(self.optimizer) grad_norm = torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=2.0).item() self.scaler.step(self.optimizer) self.scaler.update() return losses, grad_norm def train_one_epoch(self, loss_log, lr_scheduler): t0 = perf_counter() chunks = 0 self.model.train() progress_bar = tqdm( total=len(self.train_loader), desc='[0/{}]'.format(len(self.train_loader.dataset)), ascii=True, leave=True, ncols=100, bar_format='{l_bar}{bar}| [{elapsed}{postfix}]' ) smoothed_loss = None with progress_bar: for batch in self.train_loader: chunks += batch[0].shape[0] losses, grad_norm = self.train_one_step(batch) losses = {k: v.item() for k,v in losses.items()} if lr_scheduler is not None: lr_scheduler.step() smoothed_loss = losses['loss'] if smoothed_loss is None else (0.01 * losses['loss'] + 0.99 * smoothed_loss) progress_bar.set_postfix(loss='%.4f' % smoothed_loss) progress_bar.set_description("[{}/{}]".format(chunks, len(self.train_loader.dataset))) progress_bar.update() if loss_log is not None: loss_log.append({'chunks': chunks, 'time': perf_counter() - t0, 'grad_norm': grad_norm, **losses}) return smoothed_loss, perf_counter() - t0 def validate_one_step(self, batch): data, targets, lengths = batch scores = self.model(data.to(self.device)) losses = self.criterion(scores, targets.to(self.device), lengths.to(self.device)) losses = {k: v.item() for k, v in losses.items()} if isinstance(losses, dict) else losses.item() if hasattr(self.model, 'decode_batch'): seqs = self.model.decode_batch(scores) else: seqs = [self.model.decode(x) for x in permute(scores, 'TNC', 'NTC')] refs = [decode_ref(target, self.model.alphabet) for target in targets] accs = [ accuracy(ref, seq, min_coverage=0.5) if len(seq) else 0. for ref, seq in zip(refs, seqs) ] return seqs, refs, accs, losses def validate_one_epoch(self): self.model.eval() with torch.no_grad(): seqs, refs, accs, losses = zip(*(self.validate_one_step(batch) for batch in self.valid_loader)) seqs, refs, accs = (sum(x, []) for x in (seqs, refs, accs)) loss = np.mean([(x['ctc_loss'] if isinstance(x, dict) else x) for x in losses]) return loss, np.mean(accs), np.median(accs) def init_optimizer(self, lr, **kwargs): self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=lr, **kwargs) def get_lr_scheduler(self, epochs, last_epoch=0): return func_scheduler( self.optimizer, cosine_decay_schedule(1.0, 0.1), epochs * len(self.train_loader), warmup_steps=500, start_step=last_epoch*len(self.train_loader) ) def fit(self, workdir, epochs=1, lr=2e-3, last_epoch=0): if self.optimizer is None: self.init_optimizer(lr) lr_scheduler = self.get_lr_scheduler(epochs, last_epoch=last_epoch) for epoch in range(1 + last_epoch, epochs + 1 + last_epoch): try: with bonito.io.CSVLogger(os.path.join(workdir, 'losses_{}.csv'.format(epoch))) as loss_log: train_loss, duration = self.train_one_epoch(loss_log, lr_scheduler) model_state = self.model.module.state_dict() if hasattr(self.model, 'module') else self.model.state_dict() torch.save(model_state, os.path.join(workdir, "weights_%s.tar" % epoch)) val_loss, val_mean, val_median = self.validate_one_epoch() except KeyboardInterrupt: break print("[epoch {}] directory={} loss={:.4f} mean_acc={:.3f}% median_acc={:.3f}%".format( epoch, workdir, val_loss, val_mean, val_median )) with bonito.io.CSVLogger(os.path.join(workdir, 'training.csv')) as training_log: training_log.append({ 'time': datetime.today(), 'duration': int(duration), 'epoch': epoch, 'train_loss': train_loss, 'validation_loss': val_loss, 'validation_mean': val_mean, 'validation_median': val_median })
""" Bonito Download """ class File: """ Small class for downloading models and training assets. """ __url__ = "https://nanoporetech.box.com/shared/static/" def __init__(self, path, url_frag, force): self.path = path self.force = force self.url = os.path.join(self.__url__, url_frag) def location(self, filename): return os.path.join(self.path, filename) def exists(self, filename): return os.path.exists(self.location(filename)) def download(self): """ Download the remote file """ # create the requests for the file req = requests.get(self.url, stream=True) total = int(req.headers.get('content-length', 0)) fname = re.findall('filename="([^"]+)', req.headers['content-disposition'])[0] # skip download if local file is found if self.exists(fname.strip('.zip')) and not self.force: print("[skipping %s]" % fname) return if self.exists(fname.strip('.zip')) and self.force: rmtree(self.location(fname.strip('.zip'))) # download the file with tqdm(total=total, unit='iB', ascii=True, ncols=100, unit_scale=True, leave=False) as t: with open(self.location(fname), 'wb') as f: for data in req.iter_content(1024): f.write(data) t.update(len(data)) print("[downloaded %s]" % fname) # unzip .zip files if fname.endswith('.zip'): with ZipFile(self.location(fname), 'r') as zfile: zfile.extractall(self.path) os.remove(self.location(fname)) # convert chunkify training files to bonito if fname.endswith('.hdf5'): print("[converting %s]" % fname) args = cargparser().parse_args([ self.location(fname), self.location(fname).strip('.hdf5') ]) convert(args) r9_models = [ "n8c07gc9ro09zt0ivgcoeuz6krnwsnf6.zip", # dna_r9.4.1@v1 "nas0uhf46fd1lh2jndhx2a54a9vvhxp4.zip", # dna_r9.4.1@v2 "1wodp3ur4jhvqvu5leowfg6lrw54jxp2.zip", # dna_r9.4.1@v3 "uetgwsnb8yfqvuyoka8p09mxilgskqc7.zip", # dna_r9.4.1@v3.1 "47t2y48zw4waly25lmzx6sagf4bbbqqz.zip", # dna_r9.4.1@v3.2 "hrv649cvx8lvomu1u0tsd47e5u2bbabt.zip", # dna_r9.4.1@v3.3 "arqi4qwcj9btsd6bbjsnlbai0s6dg8yd.zip", ] r10_models = [ "e70s615lh3i24rkhz006i0e4u4m8y2xa.zip", # dna_r10.3_q20ea "hnr5mwlm8vmdsfpvn5fsxn3mvhbucy5f.zip", # dna_r10.3@v3 "yesf11tisfrncmod5hj2xtx9kbdveuqt.zip", # dna_r10.3@v3.2 "ci6xdu7d4wczmhorhw1sweyg4gczx97t.zip", # dna_r10.3@v3.3 "4cunv5z7nwjag7v2bun0g7vk2lf8rqnc.zip", ] training = [ "cmh91cxupa0are1kc3z9aok425m75vrb.hdf5", ] def main(args): """ Download models and training sets """ if args.models or args.all: print("[downloading models]") for model in r9_models[-1 if args.latest else 0:]: File(__models__, model, args.force).download() for model in r10_models[-1 if args.latest else 0:]: File(__models__, model, args.force).download() if args.training or args.all: print("[downloading training data]") for train in training: File(__data__, train, args.force).download() def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) group = parser.add_mutually_exclusive_group() group.add_argument('--all', action='store_true') group.add_argument('--models', action='store_true') group.add_argument('--training', action='store_true') parser.add_argument('-f', '--force', action='store_true') parser.add_argument('--latest', action='store_true') return parser
#!/usr/bin/env python """ Convert a Taiyaki chunkify training file to set of Bonito CTC .npy files """ def align(samples, pointers, reference): """ align to the start of the mapping """ squiggle_duration = len(samples) mapped_off_the_start = len(pointers[pointers < 0]) mapped_off_the_end = len(pointers[pointers >= squiggle_duration]) pointers = pointers[mapped_off_the_start:len(pointers) - mapped_off_the_end] reference = reference[mapped_off_the_start:len(reference) - mapped_off_the_end] return samples[pointers[0]:pointers[-1]], pointers - pointers[0], reference def scale(read, normalise=True): """ scale and normalise a read """ samples = read['Dacs'][:] scaling = read.attrs['range'] / read.attrs['digitisation'] scaled = (scaling * (samples + read.attrs['offset'])).astype(np.float32) if normalise: return (scaled - read.attrs['shift_frompA']) / read.attrs['scale_frompA'] return scaled def pad_lengths(ragged_array, max_len=None): lengths = np.array([len(x) for x in ragged_array], dtype=np.uint16) padded = np.zeros((len(ragged_array), max_len or np.max(lengths)), dtype=ragged_array[0].dtype) for x, y in zip(ragged_array, padded): y[:len(x)] = x return padded, lengths def regular_break_points(n, chunk_len, overlap=0, align='mid'): num_chunks, remainder = divmod(n - overlap, chunk_len - overlap) start = {'left': 0, 'mid': remainder // 2, 'right': remainder}[align] starts = np.arange(start, start + num_chunks*(chunk_len - overlap), (chunk_len - overlap)) return np.vstack([starts, starts + chunk_len]).T def get_chunks(read, break_points): sample = scale(read) pointers = read['Ref_to_signal'][:] target = read['Reference'][:] + 1 # CTC convention return ( (sample[i:j], target[ti:tj]) for (i, j), (ti, tj) in zip(break_points, np.searchsorted(pointers, break_points)) ) def chunk_dataset(reads, chunk_len, num_chunks=None): all_chunks = ( (chunk, target) for read in reads for chunk, target in get_chunks(reads[read], regular_break_points(len(reads[read]['Dacs']), chunk_len)) ) chunks, targets = zip(*tqdm(take(all_chunks, num_chunks), total=num_chunks)) targets, target_lens = pad_lengths(targets) # convert refs from ragged arrray return ChunkDataSet(chunks, targets, target_lens) def validation_split(reads, num_valid=1000): reads = np.random.permutation(sorted(reads.items())) return OrderedDict(reads[:-num_valid]), OrderedDict(reads[-num_valid:]) def typical_indices(x, n=2.5): mu, sd = np.mean(x), np.std(x) idx, = np.where((mu - n*sd < x) & (x < mu + n*sd)) return idx def filter_chunks(ds, idx): filtered = ChunkDataSet(ds.chunks.squeeze(1)[idx], ds.targets[idx], ds.lengths[idx]) filtered.targets = filtered.targets[:, :filtered.lengths.max()] return filtered def save_chunks(chunks, output_directory): os.makedirs(output_directory, exist_ok=True) np.save(os.path.join(output_directory, "chunks.npy"), chunks.chunks.squeeze(1)) np.save(os.path.join(output_directory, "references.npy"), chunks.targets) np.save(os.path.join(output_directory, "reference_lengths.npy"), chunks.lengths) print() print("> data written to %s:" % output_directory) print(" - chunks.npy with shape", chunks.chunks.squeeze(1).shape) print(" - references.npy with shape", chunks.targets.shape) print(" - reference_lengths.npy shape", chunks.lengths.shape) def main(args): random.seed(args.seed) np.random.seed(args.seed) reads = h5py.File(args.chunkify_file, 'r')['Reads'] training, validation = validation_split(reads, args.validation_reads) print("> preparing training chunks\n") training_chunks = chunk_dataset(training, args.chunksize) training_indices = typical_indices(training_chunks.lengths) training_chunks = filter_chunks(training_chunks, np.random.permutation(training_indices)) save_chunks(training_chunks, args.output_directory) print("\n> preparing validation chunks\n") validation_chunks = chunk_dataset(validation, args.chunksize) validation_indices = typical_indices(validation_chunks.lengths) validation_chunks = filter_chunks(validation_chunks, validation_indices) save_chunks(validation_chunks, os.path.join(args.output_directory, "validation")) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("chunkify_file") parser.add_argument("output_directory") parser.add_argument("--seed", default=25, type=int) parser.add_argument("--chunksize", default=3600, type=int) parser.add_argument("--validation-reads", default=1000, type=int) return parser
""" Bonito Export """ class JsonEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, np.integer): return int(obj) elif isinstance(obj, np.floating): return float(obj) elif isinstance(obj, np.ndarray): return obj.tolist() elif isinstance(obj, torch.nn.Parameter): return obj.data elif isinstance(obj, torch.Tensor): return obj.detach().numpy() else: return super(JsonEncoder, self).default(obj) def file_md5(filename, nblock=1024): """ Get md5 string from file. """ hasher = hashlib.md5() block_size = nblock * hasher.block_size with open(filename, "rb") as fh: for blk in iter((lambda: fh.read(block_size)), b""): hasher.update(blk) return hasher.hexdigest() def reformat_output_layer(layer_dict): n_base, state_len, blank_score = [layer_dict.pop(k) for k in ['n_base', 'state_len', 'blank_score']] layer_dict['size'] = (n_base + 1) * n_base**state_len layer_dict['type'] = 'GlobalNormTransducer' if blank_score is not None: assert layer_dict['activation'] == 'tanh' params = layer_dict['params'] params['W'] = torch.nn.functional.pad( params['W'].reshape([n_base**state_len, n_base, -1]), (0, 0, 1, 0), value=0. ).reshape((n_base + 1) * n_base**state_len, -1) params['b'] = torch.nn.functional.pad( params['b'].reshape(n_base**state_len, n_base), (1, 0), value=np.arctanh(blank_score / layer_dict['scale']) ).reshape(-1) return layer_dict def to_guppy_dict(model, include_weights=True): guppy_dict = bonito.nn.to_dict(model.encoder, include_weights=include_weights) guppy_dict['sublayers'] = [x for x in guppy_dict['sublayers'] if x['type'] != 'permute'] guppy_dict['sublayers'] = [dict(x, type='LSTM', activation='tanh', gate='sigmoid') if x['type'] == 'lstm' else x for x in guppy_dict['sublayers']] guppy_dict['sublayers'] = [dict(x, padding=(x['padding'], x['padding'])) if x['type'] == 'convolution' else x for x in guppy_dict['sublayers']] guppy_dict['sublayers'] = [{'type': 'reverse', 'sublayers': x} if x.pop('reverse', False) else x for x in guppy_dict['sublayers']] guppy_dict['sublayers'][-1] = reformat_output_layer(guppy_dict['sublayers'][-1]) return guppy_dict def main(args): if not os.path.isdir(args.model): print("[error] file given - please provide a model directory to export.", file=sys.stderr) return 1 model = bonito.util.load_model(args.model, device='cpu') jsn = to_guppy_dict(model) weight_files = glob(os.path.join(args.model, "weights_*.tar")) weights = max([int(re.sub(".*_([0-9]+).tar", "\\1", w)) for w in weight_files]) jsn["md5sum"] = file_md5(os.path.join(args.model, 'weights_%s.tar' % weights)) json.dump(jsn, sys.stdout, cls=JsonEncoder) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument('model') return parser
""" Bonito model viewer - display a model architecture for a given config. """ def main(args): config = toml.load(args.config) Model = load_symbol(config, "Model") model = Model(config) print(model) print("Total parameters in model", sum(p.numel() for p in model.parameters())) def argparser(): parser = argparse.ArgumentParser( formatter_class=argparse.ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("config") return parser
""" Bonito Basecaller """ def main(args): if args.save_ctc and not args.reference: sys.stderr.write("> a reference is needed to output ctc training data\n") exit(1) sys.stderr.write("> loading model\n") model = load_model(args.model_directory, args.device, weights=int(args.weights)) if args.reference: sys.stderr.write("> loading reference\n") aligner = Aligner(args.reference, preset='ont-map', best_n=1) if not aligner: sys.stderr.write("> failed to load/build index\n") exit(1) else: aligner = None reads = get_reads( args.reads_directory, n_proc=8, recursive=args.recursive, read_ids=column_to_set(args.read_ids), skip=args.skip, cancel=process_cancel() ) if args.max_reads: reads = take(reads, args.max_reads) basecall = load_symbol(args.model_directory, "basecall") if args.save_ctc: reads = ( chunk for read in reads for chunk in read_chunks(read, chunksize=args.chunksize) ) basecalls = basecall( model, reads, batchsize=64, chunksize=args.chunksize, aligner=aligner, qscores=args.fastq, reverse=args.revcomp, ) writer = CTCWriter( tqdm(basecalls, desc="> calling", unit=" reads", leave=False), aligner, args.ctc_min_coverage, args.ctc_min_accuracy ) else: basecalls = basecall( model, reads, aligner=aligner, reverse=args.revcomp, qscores=args.fastq, batchsize=args.batchsize, chunksize=args.chunksize, ) writer = Writer( tqdm(basecalls, desc="> calling", unit=" reads", leave=False), aligner, fastq=args.fastq ) t0 = perf_counter() writer.start() writer.join() duration = perf_counter() - t0 num_samples = sum(num_samples for read_id, num_samples in writer.log) sys.stderr.write("> completed reads: %s\n" % len(writer.log)) sys.stderr.write("> duration: %s\n" % timedelta(seconds=np.round(duration))) sys.stderr.write("> samples per second %.1E\n" % (num_samples / duration)) sys.stderr.write("> done\n") def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("model_directory") parser.add_argument("reads_directory") parser.add_argument("--reference") parser.add_argument("--read-ids") parser.add_argument("--device", default="cuda") parser.add_argument("--weights", default="0", type=str) parser.add_argument("--skip", action="store_true", default=False) parser.add_argument("--fastq", action="store_true", default=False) parser.add_argument("--save-ctc", action="store_true", default=False) parser.add_argument("--revcomp", action="store_true", default=False) parser.add_argument("--recursive", action="store_true", default=False) parser.add_argument("--ctc-min-coverage", default=0.9, type=float) parser.add_argument("--ctc-min-accuracy", default=0.9, type=float) parser.add_argument("--batchsize", default=32, type=int) parser.add_argument("--chunksize", default=4000, type=int) parser.add_argument("--max-reads", default=0, type=int) return parser
""" Bonito Duplex consensus decoding. https://www.biorxiv.org/content/10.1101/2020.02.25.956771v1 """ def poagen(groups, gpu_percent=0.8): free, total = cuda.cuda_get_mem_info(cuda.cuda_get_device()) gpu_mem_per_batch = gpu_percent * free max_seq_sz = 0 max_sequences_per_poa = 0 for group in groups: longest_seq = len(max(group, key=len)) max_seq_sz = longest_seq if longest_seq > max_seq_sz else max_seq_sz seq_in_poa = len(group) max_sequences_per_poa = seq_in_poa if seq_in_poa > max_sequences_per_poa else max_sequences_per_poa batch = CudaPoaBatch( max_sequences_per_poa, max_seq_sz, gpu_mem_per_batch, output_type="consensus", cuda_banded_alignment=True, alignment_band_width=256, ) poa_index = 0 initial_count = 0 while poa_index < len(groups): group = groups[poa_index] group_status, seq_status = batch.add_poa_group(group) # If group was added and more space is left in batch, continue onto next group. if group_status == 0: for seq_index, status in enumerate(seq_status): if status != 0: print("Could not add sequence {} to POA {} - error {}".format(seq_index, poa_index, status_to_str(status)), file=sys.stderr) poa_index += 1 # Once batch is full or no groups are left, run POA processing. if ((group_status == 1) or ((group_status == 0) and (poa_index == len(groups)))): batch.generate_poa() consensus, coverage, con_status = batch.get_consensus() for p, status in enumerate(con_status): if status != 0: print("Could not get consensus for POA group {} - {}".format(initial_count + p, status_to_str(status)), file=sys.stderr) yield from consensus initial_count = poa_index batch.reset() # In the case where POA group wasn't processed correctly. elif group_status != 0: print("Could not add POA group {} to batch - {}".format(poa_index, status_to_str(group_status)), file=sys.stderr) poa_index += 1 def get_read(readdir, summary, idx): """ Get a single read from row `idx` in the `summary` dataframe. """ return get_raw_data_for_read( (readdir / summary.iloc[idx].filename_fast5, summary.iloc[idx].read_id) ) def read_gen(directory, summary, n_proc=1, cancel=None): """ Generate reads from the given `directory` listed in the `summary` dataframe. """ with Pool(n_proc) as pool: for read in pool.imap(partial(get_read, Path(directory), summary), range(len(summary))): yield read if cancel is not None and cancel.is_set(): return def get_read_ids(filename): """ Return a dictionary of read_id -> filename mappings. """ with get_fast5_file(filename, 'r') as f5: return { read.read_id: basename(filename) for read in f5.get_reads() } def build_index(files, n_proc=1): """ Build an index of read ids to filename mappings """ index = {} with ProcessPoolExecutor(max_workers=n_proc) as pool: for res in tqdm(pool.map(get_read_ids, files), leave=False): index.update(res) return index def build_envelope(len1, seq1, path1, len2, seq2, path2, padding=15): # needleman-wunsch alignment with constant gap penalty. aln = parasail.nw_trace_striped_32(seq2, seq1, 2, 2, parasail.dnafull) # pair up positions alignment = np.column_stack([ np.cumsum([x != '-' for x in aln.traceback.ref]) - 1, np.cumsum([x != '-' for x in aln.traceback.query]) - 1 ]) path_range1 = np.column_stack([path1, path1[1:] + [len1]]) path_range2 = np.column_stack([path2, path2[1:] + [len2]]) envelope = np.full((len1, 2), -1, dtype=int) for idx1, idx2 in alignment.clip(0): st_1, en_1 = path_range1[idx1] st_2, en_2 = path_range2[idx2] for idx in range(st_1, en_1): if st_2 < envelope[idx, 0] or envelope[idx, 0] < 0: envelope[idx, 0] = st_2 if en_2 > envelope[idx, 1] or envelope[idx, 1] < 0: envelope[idx, 1] = en_2 # add a little padding to ensure some overlap envelope[:, 0] = envelope[:, 0] - padding envelope[:, 1] = envelope[:, 1] + padding envelope = np.clip(envelope, 0, len2) prev_end = 0 for i in range(envelope.shape[0]): if envelope[i, 0] > envelope[i, 1]: envelope[i, 0] = 0 if envelope[i, 0] > prev_end: envelope[i, 0] = prev_end prev_end = envelope[i, 1] return envelope.astype(np.uint64) def find_follow_on(df, gap=5, distance=51, cov=0.85, min_len=100): """ Find follow on reads from a sequencing summary file. """ df = df[ df.alignment_coverage.astype('float32').gt(cov) & df.sequence_length_template.astype('int32').gt(min_len) ] df = df.sort_values(['run_id', 'channel', 'mux', 'start_time']) genome_start = np.array(df.alignment_genome_start, dtype=np.int32) genome_end = np.array(df.alignment_genome_end, dtype=np.int32) direction = np.array(df.alignment_direction) start_time = np.array(df.start_time, dtype=np.float32) end_time = np.array(df.start_time + df.duration, dtype=np.float32) channel = np.array(df.channel, dtype=np.int32) mux = np.array(df.mux, dtype=np.int32) filt = ( (channel[1:] == channel[:-1]) & (mux[1:] == mux[:-1]) & (np.abs(genome_start[1:] - genome_start[:-1]) < distance) & (np.abs(genome_end[1:] - genome_end[:-1]) < distance) & (direction[1:] != direction[:-1]) & (start_time[1:] - end_time[:-1] < gap) ) mask = np.full(len(filt) + 1, False) mask[:-1] = mask[:-1] | filt mask[1:] = mask[1:] | filt return df[mask] def compute_scores(model, batch, reverse=False): with torch.no_grad(): device = next(model.parameters()).device dtype = torch.float16 if half_supported() else torch.float32 scores = model.encoder(batch.to(dtype).to(device)) if reverse: scores = model.seqdist.reverse_complement(scores) betas = model.seqdist.backward_scores(scores.to(torch.float32)) trans, init = model.seqdist.compute_transition_probs(scores, betas) return { 'trans': trans.to(dtype).transpose(0, 1), 'init': init.to(dtype).unsqueeze(1), } def basecall(model, reads, chunksize=4000, overlap=500, batchsize=32, reverse=False): reads = ( read_chunk for read in reads for read_chunk in split_read(read, chunksize * batchsize)[::-1 if reverse else 1] ) chunks = ( ((read, start, end), chunk(torch.from_numpy(read.signal[start:end]), chunksize, overlap)) for (read, start, end) in reads ) batches = ( (k, compute_scores(model, batch, reverse=reverse)) for k, batch in batchify(chunks, batchsize=batchsize) ) stitched = ( (read, stitch(x, chunksize, overlap, end - start, model.stride, reverse=reverse)) for ((read, start, end), x) in unbatchify(batches) ) transferred = thread_map(transfer, stitched, n_thread=1) return ( (read, concat([part for k, part in parts])) for read, parts in groupby(transferred, lambda x: x[0]) ) def beam_search_duplex(seq1, path1, t1, b1, seq2, path2, t2, b2, alphabet='NACGT', beamsize=5, pad=40, T=0.01): env = build_envelope(t1.shape[0], seq1, path1, t2.shape[0], seq2, path2, padding=pad) return crf_beam_search_duplex( t1, b1, t2, b2, alphabet=alphabet, beam_size=beamsize, beam_cut_threshold=T, envelope=env, ) def decode(res, beamsize_1=5, pad_1=40, cut_1=0.01, beamsize_2=5, pad_2=40, cut_2=0.01, match=80, alphabet="NACGT"): temp_probs, init1 = res[0]['trans'].astype(np.float32), res[0]['init'][0].astype(np.float32) comp_probs, init2 = res[1]['trans'].astype(np.float32), res[1]['init'][0].astype(np.float32) simplex1, path1 = crf_beam_search(temp_probs, init1, alphabet, beam_size=5, beam_cut_threshold=0.01) simplex2, path2 = crf_beam_search(comp_probs, init2, alphabet, beam_size=5, beam_cut_threshold=0.01) if len(simplex1) < 10 or len(simplex2) < 10: return [simplex1, simplex2] if accuracy(simplex1, simplex2) < match: return [simplex1, simplex2] duplex1 = beam_search_duplex( simplex1, path1, temp_probs, init1, simplex2, path2, comp_probs, init2, pad=pad_1, beamsize=5, T=cut_1 ) duplex2 = beam_search_duplex( simplex2, path2, comp_probs, init2, simplex1, path1, temp_probs, init1, pad=pad_2, beamsize=5, T=cut_2 ) return [duplex1, duplex2, simplex1, simplex2] def poa(seqs, allseq=False): con, msa = spoa.poa(seqs, genmsa=False) if allseq: return (con, *seqs) return (con, ) def call(model, reads_directory, templates, complements, aligner=None, cudapoa=True): temp_reads = read_gen(reads_directory, templates, n_proc=8, cancel=process_cancel()) comp_reads = read_gen(reads_directory, complements, n_proc=8, cancel=process_cancel()) temp_scores = basecall(model, temp_reads, reverse=False) comp_scores = basecall(model, comp_reads, reverse=True) scores = (((r1, r2), (s1, s2)) for (r1, s1), (r2, s2) in zip(temp_scores, comp_scores)) calls = thread_map(decode, scores, n_thread=12) if cudapoa: sequences = ((reads, [seqs, ]) for reads, seqs in calls if len(seqs) > 2) consensus = (zip(reads, poagen(calls)) for reads, calls in batchify(sequences, 100)) res = ((reads[0], {'sequence': seq}) for seqs in consensus for reads, seq in seqs) else: sequences = ((reads, seqs) for reads, seqs in calls if len(seqs) > 2) consensus = process_map(poa, sequences, n_proc=4) res = ((reads, {'sequence': seq}) for reads, seqs in consensus for seq in seqs) if aligner is None: return res return align_map(aligner, res) def main(args): sys.stderr.write("> loading model\n") model = load_model(args.model, args.device) if args.reference: sys.stderr.write("> loading reference\n") aligner = Aligner(args.reference, preset='ont-map') if not aligner: sys.stderr.write("> failed to load/build index\n") exit(1) else: aligner = None if args.summary: sys.stderr.write("> finding follow on strands\n") pairs = pd.read_csv(args.summary, '\t', low_memory=False) pairs = pairs[pairs.sequence_length_template.gt(0)] if 'filename' in pairs.columns: pairs = pairs.rename(columns={'filename': 'filename_fast5'}) if 'alignment_strand_coverage' in pairs.columns: pairs = pairs.rename(columns={'alignment_strand_coverage': 'alignment_coverage'}) valid_fast5s = [ f for f in pairs.filename_fast5.unique() if ((args.reads_directory / Path(f)).exists()) ] pairs = pairs[pairs.filename_fast5.isin(valid_fast5s)] pairs = find_follow_on(pairs) sys.stderr.write("> found %s follow strands in summary\n" % (len(pairs) // 2)) if args.max_reads > 0: pairs = pairs.head(args.max_reads) temp_reads = pairs.iloc[0::2] comp_reads = pairs.iloc[1::2] else: if args.index is not None: sys.stderr.write("> loading read index\n") index = json.load(open(args.index, 'r')) else: sys.stderr.write("> building read index\n") files = list(glob(os.path.join(args.reads_directory, '*.fast5'))) index = build_index(files, n_proc=8) if args.save_index: with open('bonito-read-id.idx', 'w') as f: json.dump(index, f) pairs = pd.read_csv(args.pairs, sep=args.sep, names=['read_1', 'read_2']) if args.max_reads > 0: pairs = pairs.head(args.max_reads) pairs['file_1'] = pairs['read_1'].apply(index.get) pairs['file_2'] = pairs['read_2'].apply(index.get) pairs = pairs.dropna().reset_index() temp_reads = pairs[['read_1', 'file_1']].rename( columns={'read_1': 'read_id', 'file_1': 'filename_fast5'} ) comp_reads = pairs[['read_2', 'file_2']].rename( columns={'read_2': 'read_id', 'file_2': 'filename_fast5'} ) if len(pairs) == 0: print("> no matched pairs found in given directory", file=sys.stderr) exit(1) # https://github.com/clara-parabricks/GenomeWorks/issues/648 with devnull(): CudaPoaBatch(1000, 1000, 3724032) basecalls = call(model, args.reads_directory, temp_reads, comp_reads, aligner=aligner) writer = Writer(tqdm(basecalls, desc="> calling", unit=" reads", leave=False), aligner, duplex=True) t0 = perf_counter() writer.start() writer.join() duration = perf_counter() - t0 num_samples = sum(num_samples for read_id, num_samples in writer.log) print("> duration: %s" % timedelta(seconds=np.round(duration)), file=sys.stderr) print("> samples per second %.1E" % (num_samples / duration), file=sys.stderr) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("model") parser.add_argument("reads_directory") group = parser.add_mutually_exclusive_group() group.add_argument("--summary", default=None) group.add_argument("--pairs", default=None) parser.add_argument("--sep", default=' ') parser.add_argument("--index", default=None) parser.add_argument("--save-index", action="store_true", default=False) parser.add_argument("--reference") parser.add_argument("--device", default="cuda") parser.add_argument("--max-reads", default=0, type=int) return parser
#!/usr/bin/env python3 """ Bonito training. """ def main(args): workdir = os.path.expanduser(args.training_directory) if os.path.exists(workdir) and not args.force: print("[error] %s exists, use -f to force continue training." % workdir) exit(1) init(args.seed, args.device) device = torch.device(args.device) print("[loading data]") train_data = load_data(limit=args.chunks, directory=args.directory) if os.path.exists(os.path.join(args.directory, 'validation')): valid_data = load_data(directory=os.path.join(args.directory, 'validation')) else: print("[validation set not found: splitting training set]") split = np.floor(len(train_data[0]) * 0.97).astype(np.int32) valid_data = [x[split:] for x in train_data] train_data = [x[:split] for x in train_data] train_loader = DataLoader(ChunkDataSet(*train_data), batch_size=args.batch, shuffle=True, num_workers=4, pin_memory=True) valid_loader = DataLoader(ChunkDataSet(*valid_data), batch_size=args.batch, num_workers=4, pin_memory=True) if args.pretrained: dirname = args.pretrained if not os.path.isdir(dirname) and os.path.isdir(os.path.join(__models__, dirname)): dirname = os.path.join(__models__, dirname) config_file = os.path.join(dirname, 'config.toml') else: config_file = args.config config = toml.load(config_file) argsdict = dict(training=vars(args)) os.makedirs(workdir, exist_ok=True) toml.dump({**config, **argsdict}, open(os.path.join(workdir, 'config.toml'), 'w')) print("[loading model]") if args.pretrained: print("[using pretrained model {}]".format(args.pretrained)) model = load_model(args.pretrained, device, half=False) else: model = load_symbol(config, 'Model')(config) last_epoch = load_state(workdir, args.device, model) if args.multi_gpu: from torch.nn import DataParallel model = DataParallel(model) model.decode = model.module.decode model.alphabet = model.module.alphabet trainer = Trainer(model, device, train_loader, valid_loader, use_amp=half_supported() and not args.no_amp) trainer.fit(workdir, args.epochs, args.lr, last_epoch=last_epoch) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("training_directory") group = parser.add_mutually_exclusive_group() group.add_argument('--config', default=default_config) group.add_argument('--pretrained', default="") parser.add_argument("--directory", default=default_data) parser.add_argument("--device", default="cuda") parser.add_argument("--lr", default=2e-3, type=float) parser.add_argument("--seed", default=25, type=int) parser.add_argument("--epochs", default=5, type=int) parser.add_argument("--batch", default=64, type=int) parser.add_argument("--chunks", default=0, type=int) parser.add_argument("--no-amp", action="store_true", default=False) parser.add_argument("--multi-gpu", action="store_true", default=False) parser.add_argument("-f", "--force", action="store_true", default=False) return parser
""" Bonito model evaluator """ def main(args): poas = [] init(args.seed, args.device) print("* loading data") directory = args.directory if os.path.exists(os.path.join(directory, 'validation')): directory = os.path.join(directory, 'validation') testdata = ChunkDataSet( *load_data( limit=args.chunks, directory=directory ) ) dataloader = DataLoader(testdata, batch_size=args.batchsize) accuracy_with_cov = lambda ref, seq: accuracy(ref, seq, min_coverage=args.min_coverage) for w in [int(i) for i in args.weights.split(',')]: seqs = [] print("* loading model", w) model = load_model(args.model_directory, args.device, weights=w) print("* calling") t0 = time.perf_counter() with torch.no_grad(): for data, *_ in dataloader: if half_supported(): data = data.type(torch.float16).to(args.device) else: data = data.to(args.device) log_probs = model(data) if hasattr(model, 'decode_batch'): seqs.extend(model.decode_batch(log_probs)) else: seqs.extend([model.decode(p) for p in permute(log_probs, 'TNC', 'NTC')]) duration = time.perf_counter() - t0 refs = [decode_ref(target, model.alphabet) for target in dataloader.dataset.targets] accuracies = [accuracy_with_cov(ref, seq) if len(seq) else 0. for ref, seq in zip(refs, seqs)] if args.poa: poas.append(sequences) print("* mean %.2f%%" % np.mean(accuracies)) print("* median %.2f%%" % np.median(accuracies)) print("* time %.2f" % duration) print("* samples/s %.2E" % (args.chunks * data.shape[2] / duration)) if args.poa: print("* doing poa") t0 = time.perf_counter() # group each sequence prediction per model together poas = [list(seq) for seq in zip(*poas)] consensuses = poa(poas) duration = time.perf_counter() - t0 accuracies = list(starmap(accuracy_with_coverage_filter, zip(references, consensuses))) print("* mean %.2f%%" % np.mean(accuracies)) print("* median %.2f%%" % np.median(accuracies)) print("* time %.2f" % duration) def argparser(): parser = ArgumentParser( formatter_class=ArgumentDefaultsHelpFormatter, add_help=False ) parser.add_argument("model_directory") parser.add_argument("--directory", default=None) parser.add_argument("--device", default="cuda") parser.add_argument("--seed", default=9, type=int) parser.add_argument("--weights", default="0", type=str) parser.add_argument("--chunks", default=1000, type=int) parser.add_argument("--batchsize", default=96, type=int) parser.add_argument("--beamsize", default=5, type=int) parser.add_argument("--poa", action="store_true", default=False) parser.add_argument("--min-coverage", default=0.5, type=float) return parser
""" Bonito CTC-CRF Model. """ def get_stride(m): if hasattr(m, 'stride'): return m.stride if isinstance(m.stride, int) else m.stride[0] if isinstance(m, Convolution): return get_stride(m.conv) if isinstance(m, Serial): return int(np.prod([get_stride(x) for x in m])) return 1 class CTC_CRF(SequenceDist): def __init__(self, state_len, alphabet): super().__init__() self.alphabet = alphabet self.state_len = state_len self.n_base = len(alphabet[1:]) self.idx = torch.cat([ torch.arange(self.n_base**(self.state_len))[:, None], torch.arange( self.n_base**(self.state_len) ).repeat_interleave(self.n_base).reshape(self.n_base, -1).T ], dim=1).to(torch.int32) def n_score(self): return len(self.alphabet) * self.n_base**(self.state_len) def logZ(self, scores, S:semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, len(self.alphabet)) alpha_0 = Ms.new_full((N, self.n_base**(self.state_len)), S.one) beta_T = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return seqdist.sparse.logZ(Ms, self.idx, alpha_0, beta_T, S) def normalise(self, scores): return (scores - self.logZ(scores)[:, None] / len(scores)) def forward_scores(self, scores, S: semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, self.n_base + 1) alpha_0 = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return seqdist.sparse.fwd_scores_cupy(Ms, self.idx, alpha_0, S, K=1) def backward_scores(self, scores, S: semiring=Log): T, N, _ = scores.shape Ms = scores.reshape(T, N, -1, self.n_base + 1) beta_T = Ms.new_full((N, self.n_base**(self.state_len)), S.one) return seqdist.sparse.bwd_scores_cupy(Ms, self.idx, beta_T, S, K=1) def compute_transition_probs(self, scores, betas): T, N, C = scores.shape # add bwd scores to edge scores log_trans_probs = (scores.reshape(T, N, -1, self.n_base + 1) + betas[1:, :, :, None]) # transpose from (new_state, dropped_base) to (old_state, emitted_base) layout log_trans_probs = torch.cat([ log_trans_probs[:, :, :, [0]], log_trans_probs[:, :, :, 1:].transpose(3, 2).reshape(T, N, -1, self.n_base) ], dim=-1) # convert from log probs to probs by exponentiating and normalising trans_probs = torch.softmax(log_trans_probs, dim=-1) #convert first bwd score to initial state probabilities init_state_probs = torch.softmax(betas[0], dim=-1) return trans_probs, init_state_probs def reverse_complement(self, scores): T, N, C = scores.shape expand_dims = T, N, *(self.n_base for _ in range(self.state_len)), self.n_base + 1 scores = scores.reshape(*expand_dims) blanks = torch.flip(scores[..., 0].permute( 0, 1, *range(self.state_len + 1, 1, -1)).reshape(T, N, -1, 1), [0, 2] ) emissions = torch.flip(scores[..., 1:].permute( 0, 1, *range(self.state_len, 1, -1), self.state_len +2, self.state_len + 1).reshape(T, N, -1, self.n_base), [0, 2, 3] ) return torch.cat([blanks, emissions], dim=-1).reshape(T, N, -1) def viterbi(self, scores): traceback = self.posteriors(scores, Max) paths = traceback.argmax(2) % len(self.alphabet) return paths def path_to_str(self, path): alphabet = np.frombuffer(''.join(self.alphabet).encode(), dtype='u1') seq = alphabet[path[path != 0]] return seq.tobytes().decode() def prepare_ctc_scores(self, scores, targets): # convert from CTC targets (with blank=0) to zero indexed targets = torch.clamp(targets - 1, 0) T, N, C = scores.shape scores = scores.to(torch.float32) n = targets.size(1) - (self.state_len - 1) stay_indices = sum( targets[:, i:n + i] * self.n_base ** (self.state_len - i - 1) for i in range(self.state_len) ) * len(self.alphabet) move_indices = stay_indices[:, 1:] + targets[:, :n - 1] + 1 stay_scores = scores.gather(2, stay_indices.expand(T, -1, -1)) move_scores = scores.gather(2, move_indices.expand(T, -1, -1)) return stay_scores, move_scores def ctc_loss(self, scores, targets, target_lengths, loss_clip=None, reduction='mean', normalise_scores=True): if normalise_scores: scores = self.normalise(scores) stay_scores, move_scores = self.prepare_ctc_scores(scores, targets) logz = logZ_cupy(stay_scores, move_scores, target_lengths + 1 - self.state_len) loss = - (logz / target_lengths) if loss_clip: loss = torch.clamp(loss, 0.0, loss_clip) if reduction == 'mean': return loss.mean() elif reduction in ('none', None): return loss else: raise ValueError('Unknown reduction type {}'.format(reduction)) def ctc_viterbi_alignments(self, scores, targets, target_lengths): stay_scores, move_scores = self.prepare_ctc_scores(scores, targets) return viterbi_alignments(stay_scores, move_scores, target_lengths + 1 - self.state_len) def conv(c_in, c_out, ks, stride=1, bias=False, activation=None): return Convolution(c_in, c_out, ks, stride=stride, padding=ks//2, bias=bias, activation=activation) def rnn_encoder(n_base, state_len, insize=1, stride=5, winlen=19, activation='swish', rnn_type='lstm', features=768, scale=5.0, blank_score=None, single_head_attn=False): rnn = layers[rnn_type] return Serial([ conv(insize, 4, ks=5, bias=True, activation=activation), conv(4, 16, ks=5, bias=True, activation=activation), conv(16, features, ks=winlen, stride=stride, bias=True, activation=activation), Permute([2, 0, 1]), rnn(features, features, reverse=True), rnn(features, features), rnn(features, features, reverse=True), rnn(features, features), *([SHABlock(features)] if single_head_attn else []), rnn(features, features, reverse=True), LinearCRFEncoder(features, n_base, state_len, bias=True, activation='tanh', scale=scale, blank_score=blank_score) ]) class SeqdistModel(Module): def __init__(self, encoder, seqdist): super().__init__() self.seqdist = seqdist self.encoder = encoder self.stride = get_stride(encoder) self.alphabet = seqdist.alphabet def forward(self, x): return self.encoder(x).to(torch.float32) def decode_batch(self, x): scores = self.seqdist.posteriors(x.to(torch.float32)) + 1e-8 tracebacks = self.seqdist.viterbi(scores.log()).to(torch.int16).T return [self.seqdist.path_to_str(x) for x in tracebacks.cpu().numpy()] def decode(self, x): return self.decode_batch(x.unsqueeze(1))[0] class Model(SeqdistModel): def __init__(self, config): seqdist = CTC_CRF( state_len=config['global_norm']['state_len'], alphabet=config['labels']['labels'] ) if 'type' in config['encoder']: #new-style config encoder = from_dict(config['encoder']) else: #old-style encoder = rnn_encoder(seqdist.n_base, seqdist.state_len, insize=config['input']['features'], **config['encoder']) super().__init__(encoder, seqdist) self.config = config
""" Bonito CRF basecall """ def stitch(chunks, chunksize, overlap, length, stride, reverse=False): """ Stitch chunks together with a given overlap """ if isinstance(chunks, dict): return { k: stitch(v, chunksize, overlap, length, stride, reverse=reverse) for k, v in chunks.items() } return bonito.util.stitch(chunks, chunksize, overlap, length, stride, reverse=reverse) def compute_scores(model, batch, reverse=False): """ Compute scores for model. """ with torch.no_grad(): device = next(model.parameters()).device dtype = torch.float16 if half_supported() else torch.float32 scores = model(batch.to(dtype).to(device)) if reverse: scores = model.seqdist.reverse_complement(scores) betas = model.seqdist.backward_scores(scores.to(torch.float32)) betas -= (betas.max(2, keepdim=True)[0] - 5.0) return { 'scores': scores.transpose(0, 1), 'betas': betas.transpose(0, 1), } def quantise_int8(x, scale=127/5): """ Quantise scores to int8. """ scores = x['scores'] scores *= scale scores = torch.round(scores).to(torch.int8).detach() betas = x['betas'] betas *= scale betas = torch.round(torch.clamp(betas, -127., 128.)).to(torch.int8).detach() return {'scores': scores, 'betas': betas} def transfer(x): """ Device to host transfer using pinned memory. """ torch.cuda.synchronize() with torch.cuda.stream(torch.cuda.Stream()): return { k: torch.empty(v.shape, pin_memory=True, dtype=v.dtype).copy_(v).numpy() for k, v in x.items() } def decode_int8(scores, seqdist, scale=127/5, beamsize=40, beamcut=100.0): """ Beamsearch decode. """ path, _ = beamsearch( scores['scores'], scale, seqdist.n_base, beamsize, guide=scores['betas'], beam_cut=beamcut ) try: return seqdist.path_to_str(path % 4 + 1) except IndexError: return "" def split_read(read, split_read_length=400000): """ Split large reads into manageable pieces. """ if len(read.signal) <= split_read_length: return [(read, 0, len(read.signal))] breaks = np.arange(0, len(read.signal) + split_read_length, split_read_length) return [(read, start, min(end, len(read.signal))) for (start, end) in zip(breaks[:-1], breaks[1:])] def basecall(model, reads, aligner=None, beamsize=40, chunksize=4000, overlap=500, batchsize=32, qscores=False, reverse=False): """ Basecalls a set of reads. """ _decode = partial(decode_int8, seqdist=model.seqdist, beamsize=beamsize) reads = (read_chunk for read in reads for read_chunk in split_read(read)[::-1 if reverse else 1]) chunks = ( ((read, start, end), chunk(torch.from_numpy(read.signal[start:end]), chunksize, overlap)) for (read, start, end) in reads ) batches = ( (k, quantise_int8(compute_scores(model, batch, reverse=reverse))) for k, batch in thread_iter(batchify(chunks, batchsize=batchsize)) ) stitched = ( (read, stitch(x, chunksize, overlap, end - start, model.stride, reverse=reverse)) for ((read, start, end), x) in unbatchify(batches) ) transferred = thread_map(transfer, stitched, n_thread=1) basecalls = thread_map(_decode, transferred, n_thread=8) basecalls = ( (read, ''.join(seq for k, seq in parts)) for read, parts in groupby(basecalls, lambda x: (x[0].parent if hasattr(x[0], 'parent') else x[0])) ) basecalls = ( (read, {'sequence': seq, 'qstring': '?' * len(seq) if qscores else '*', 'mean_qscore': 0.0}) for read, seq in basecalls ) if aligner: return align_map(aligner, basecalls) return basecalls
""" Bonito Model template """ class Model(Module): """ Model template for QuartzNet style architectures https://arxiv.org/pdf/1910.10261.pdf """ def __init__(self, config): super(Model, self).__init__() if 'qscore' not in config: self.qbias = 0.0 self.qscale = 1.0 else: self.qbias = config['qscore']['bias'] self.qscale = config['qscore']['scale'] self.config = config self.stride = config['block'][0]['stride'][0] self.alphabet = config['labels']['labels'] self.features = config['block'][-1]['filters'] self.encoder = Encoder(config) self.decoder = Decoder(self.features, len(self.alphabet)) def forward(self, x): encoded = self.encoder(x) return self.decoder(encoded) def decode(self, x, beamsize=5, threshold=1e-3, qscores=False, return_path=False): x = x.exp().cpu().numpy().astype(np.float32) if beamsize == 1 or qscores: seq, path = viterbi_search(x, self.alphabet, qscores, self.qscale, self.qbias) else: seq, path = beam_search(x, self.alphabet, beamsize, threshold) if return_path: return seq, path return seq def ctc_label_smoothing_loss(self, log_probs, targets, lengths, weights=None): T, N, C = log_probs.shape weights = weights or torch.cat([torch.tensor([0.4]), (0.1 / (C - 1)) * torch.ones(C - 1)]) log_probs_lengths = torch.full(size=(N, ), fill_value=T, dtype=torch.int64) loss = ctc_loss(log_probs.to(torch.float32), targets, log_probs_lengths, lengths, reduction='mean') label_smoothing_loss = -((log_probs * weights.to(log_probs.device)).mean()) return {'loss': loss + label_smoothing_loss, 'ctc_loss': loss, 'label_smooth_loss': label_smoothing_loss} class Encoder(Module): """ Builds the model encoder """ def __init__(self, config): super(Encoder, self).__init__() self.config = config features = self.config['input']['features'] activation = layers[self.config['encoder']['activation']]() encoder_layers = [] for layer in self.config['block']: encoder_layers.append( Block( features, layer['filters'], activation, repeat=layer['repeat'], kernel_size=layer['kernel'], stride=layer['stride'], dilation=layer['dilation'], dropout=layer['dropout'], residual=layer['residual'], separable=layer['separable'], ) ) features = layer['filters'] self.encoder = Sequential(*encoder_layers) def forward(self, x): return self.encoder(x) class TCSConv1d(Module): """ Time-Channel Separable 1D Convolution """ def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=False, separable=False): super(TCSConv1d, self).__init__() self.separable = separable if separable: self.depthwise = Conv1d( in_channels, in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, groups=in_channels ) self.pointwise = Conv1d( in_channels, out_channels, kernel_size=1, stride=1, dilation=dilation, bias=bias, padding=0 ) else: self.conv = Conv1d( in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias ) def forward(self, x): if self.separable: x = self.depthwise(x) x = self.pointwise(x) else: x = self.conv(x) return x class Block(Module): """ TCSConv, Batch Normalisation, Activation, Dropout """ def __init__(self, in_channels, out_channels, activation, repeat=5, kernel_size=1, stride=1, dilation=1, dropout=0.0, residual=False, separable=False): super(Block, self).__init__() self.use_res = residual self.conv = ModuleList() _in_channels = in_channels padding = self.get_padding(kernel_size[0], stride[0], dilation[0]) # add the first n - 1 convolutions + activation for _ in range(repeat - 1): self.conv.extend( self.get_tcs( _in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding, separable=separable ) ) self.conv.extend(self.get_activation(activation, dropout)) _in_channels = out_channels # add the last conv and batch norm self.conv.extend( self.get_tcs( _in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding, separable=separable ) ) # add the residual connection if self.use_res: self.residual = Sequential(*self.get_tcs(in_channels, out_channels)) # add the activation and dropout self.activation = Sequential(*self.get_activation(activation, dropout)) def get_activation(self, activation, dropout): return activation, Dropout(p=dropout) def get_padding(self, kernel_size, stride, dilation): if stride > 1 and dilation > 1: raise ValueError("Dilation and stride can not both be greater than 1") return (kernel_size // 2) * dilation def get_tcs(self, in_channels, out_channels, kernel_size=1, stride=1, dilation=1, padding=0, bias=False, separable=False): return [ TCSConv1d( in_channels, out_channels, kernel_size, stride=stride, dilation=dilation, padding=padding, bias=bias, separable=separable ), BatchNorm1d(out_channels, eps=1e-3, momentum=0.1) ] def forward(self, x): _x = x for layer in self.conv: _x = layer(_x) if self.use_res: _x = _x + self.residual(x) return self.activation(_x) class Decoder(Module): """ Decoder """ def __init__(self, features, classes): super(Decoder, self).__init__() self.layers = Sequential( Conv1d(features, classes, kernel_size=1, bias=True), Permute([2, 0, 1]) ) def forward(self, x): return log_softmax(self.layers(x), dim=-1)
""" Bonito basecall """ def basecall(model, reads, aligner=None, beamsize=5, chunksize=0, overlap=0, batchsize=1, qscores=False, reverse=None): """ Basecalls a set of reads. """ chunks = ( (read, chunk(torch.tensor(read.signal), chunksize, overlap)) for read in reads ) scores = unbatchify( (k, compute_scores(model, v)) for k, v in batchify(chunks, batchsize) ) scores = ( (read, {'scores': stitch(v, chunksize, overlap, len(read.signal), model.stride)}) for read, v in scores ) decoder = partial(decode, decode=model.decode, beamsize=beamsize, qscores=qscores) basecalls = process_map(decoder, scores, n_proc=4) if aligner: return align_map(aligner, basecalls) return basecalls def compute_scores(model, batch): """ Compute scores for model. """ with torch.no_grad(): device = next(model.parameters()).device chunks = batch.to(torch.half).to(device) probs = permute(model(chunks), 'TNC', 'NTC') return probs.cpu().to(torch.float32) def decode(scores, decode, beamsize=5, qscores=False): """ Convert the network scores into a sequence. """ # do a greedy decode to get a sensible qstring to compute the mean qscore from seq, path = decode(scores['scores'], beamsize=1, qscores=True, return_path=True) seq, qstring = seq[:len(path)], seq[len(path):] mean_qscore = mean_qscore_from_qstring(qstring) # beam search will produce a better sequence but doesn't produce a sensible qstring/path if not (qscores or beamsize == 1): try: seq = decode(scores['scores'], beamsize=beamsize) path = None qstring = '*' except: pass return {'sequence': seq, 'qstring': qstring, 'mean_qscore': mean_qscore, 'path': path}
# constants FlashAttentionConfig = namedtuple('FlashAttentionConfig', ['enable_flash', 'enable_math', 'enable_mem_efficient']) # helpers def exists(val): return val is not None def once(fn): called = False @wraps(fn) def inner(x): nonlocal called if called: return called = True return fn(x) return inner print_once = once(print) # main class class Attend(nn.Module): def __init__( self, dropout = 0., flash = False ): super().__init__() self.dropout = dropout self.attn_dropout = nn.Dropout(dropout) self.flash = flash assert not (flash and version.parse(torch.__version__) < version.parse('2.0.0')), 'in order to use flash attention, you must be using pytorch 2.0 or above' # determine efficient attention configs for cuda and cpu self.cpu_config = FlashAttentionConfig(True, True, True) self.cuda_config = None if not torch.cuda.is_available() or not flash: return device_properties = torch.cuda.get_device_properties(torch.device('cuda')) if device_properties.major == 8 and device_properties.minor == 0: print_once('A100 GPU detected, using flash attention if input tensor is on cuda') self.cuda_config = FlashAttentionConfig(True, False, False) else: print_once('Non-A100 GPU detected, using math or mem efficient attention if input tensor is on cuda') self.cuda_config = FlashAttentionConfig(False, True, True) def flash_attn(self, q, k, v): _, heads, q_len, _, k_len, is_cuda, device = *q.shape, k.shape[-2], q.is_cuda, q.device # Check if there is a compatible device for flash attention config = self.cuda_config if is_cuda else self.cpu_config # pytorch 2.0 flash attn: q, k, v, mask, dropout, softmax_scale with torch.backends.cuda.sdp_kernel(**config._asdict()): out = F.scaled_dot_product_attention( q, k, v, dropout_p = self.dropout if self.training else 0. ) return out def forward(self, q, k, v): """ einstein notation b - batch h - heads n, i, j - sequence length (base sequence length, source, target) d - feature dimension """ q_len, k_len, device = q.shape[-2], k.shape[-2], q.device scale = q.shape[-1] ** -0.5 if self.flash: return self.flash_attn(q, k, v) # similarity sim = einsum(f"b h i d, b h j d -> b h i j", q, k) * scale # attention attn = sim.softmax(dim=-1) attn = self.attn_dropout(attn) # aggregate values out = einsum(f"b h i j, b h j d -> b h i d", attn, v) return out
# helper functions def exists(val): return val is not None # norm class RMSNorm(Module): def __init__(self, dim): super().__init__() self.scale = dim ** 0.5 self.gamma = nn.Parameter(torch.ones(dim)) def forward(self, x): return F.normalize(x, dim = -1) * self.scale * self.gamma # attention class FeedForward(Module): def __init__( self, dim, mult = 4, dropout = 0. ): super().__init__() dim_inner = int(dim * mult) self.net = nn.Sequential( RMSNorm(dim), nn.Linear(dim, dim_inner), nn.GELU(), nn.Dropout(dropout), nn.Linear(dim_inner, dim), nn.Dropout(dropout) ) def forward(self, x): return self.net(x) class Attention(Module): def __init__( self, dim, heads = 8, dim_head = 64, dropout = 0., rotary_embed = None, flash = True ): super().__init__() self.heads = heads self.scale = dim_head **-0.5 dim_inner = heads * dim_head self.rotary_embed = rotary_embed self.attend = Attend(flash = flash, dropout = dropout) self.norm = RMSNorm(dim) self.to_qkv = nn.Linear(dim, dim_inner * 3, bias = False) self.to_out = nn.Sequential( nn.Linear(dim_inner, dim, bias = False), nn.Dropout(dropout) ) def forward(self, x): x = self.norm(x) q, k, v = rearrange(self.to_qkv(x), 'b n (qkv h d) -> qkv b h n d', qkv = 3, h = self.heads) if exists(self.rotary_embed): q = self.rotary_embed.rotate_queries_or_keys(q) k = self.rotary_embed.rotate_queries_or_keys(k) out = self.attend(q, k, v) out = rearrange(out, 'b h n d -> b n (h d)') return self.to_out(out) class Transformer(Module): def __init__( self, *, dim, depth, dim_head = 64, heads = 8, attn_dropout = 0., ff_dropout = 0., ff_mult = 4, norm_output = True, rotary_embed = None, flash_attn = True ): super().__init__() self.layers = ModuleList([]) for _ in range(depth): self.layers.append(ModuleList([ Attention(dim = dim, dim_head = dim_head, heads = heads, dropout = attn_dropout, rotary_embed = rotary_embed, flash = flash_attn), FeedForward(dim = dim, mult = ff_mult, dropout = ff_dropout) ])) self.norm = RMSNorm(dim) if norm_output else nn.Identity() def forward(self, x): for attn, ff in self.layers: x = attn(x) + x x = ff(x) + x return self.norm(x) # bandsplit module class BandSplit(Module): @beartype def __init__( self, dim, dim_inputs: Tuple[int, ...] ): super().__init__() self.dim_inputs = dim_inputs self.to_features = ModuleList([]) for dim_in in dim_inputs: net = nn.Sequential( RMSNorm(dim_in), nn.Linear(dim_in, dim) ) self.to_features.append(net) def forward(self, x): x = x.split(self.dim_inputs, dim = -1) outs = [] for split_input, to_feature in zip(x, self.to_features): split_output = to_feature(split_input) outs.append(split_output) return torch.stack(outs, dim = -2) class LinearGLUWithTanH(Module): def __init__(self, dim_in, dim_out): super().__init__() self.proj = nn.Linear(dim_in, dim_out * 2) def forward(self, x): x, gate = self.proj(x).chunk(2, dim = -1) return x.tanh() * gate.sigmoid() class MaskEstimator(Module): @beartype def __init__( self, dim, dim_inputs: Tuple[int, ...], depth ): super().__init__() self.dim_inputs = dim_inputs self.to_freqs = ModuleList([]) for dim_in in dim_inputs: net = [] for ind in range(depth): is_last = ind == (depth - 1) dim_out = dim if not is_last else dim_in net.append(LinearGLUWithTanH(dim, dim_out)) self.to_freqs.append(nn.Sequential(*net)) def forward(self, x): x = x.unbind(dim = -2) outs = [] for band_features, to_freq in zip(x, self.to_freqs): freq_out = to_freq(band_features) outs.append(freq_out) return torch.cat(outs, dim = -1) # main class class BSRoformer(Module): @beartype def __init__( self, dim, *, depth, time_transformer_depth = 2, freq_transformer_depth = 2, freqs_per_bands: Tuple[int, ...] = (256, 257), # in the paper, they divide into ~60 bands, test with 1 for starters dim_head = 64, heads = 8, attn_dropout = 0., ff_dropout = 0., flash_attn = True, dim_freqs_in = 513, stft_n_fft = 1024, stft_hop_length = 256, stft_win_length = 1024, stft_normalized = False, mask_estimator_depth = 1, multi_stft_resolution_loss_weight = 1., multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256), multi_stft_hop_size = 147, multi_stft_normalized = False ): super().__init__() self.layers = ModuleList([]) transformer_kwargs = dict( dim = dim, heads = heads, dim_head = dim_head, attn_dropout = attn_dropout, ff_dropout = ff_dropout, flash_attn = flash_attn ) time_rotary_embed = RotaryEmbedding(dim = dim_head) freq_rotary_embed = RotaryEmbedding(dim = dim_head) for _ in range(depth): self.layers.append(nn.ModuleList([ Transformer(depth = time_transformer_depth, rotary_embed = time_rotary_embed, **transformer_kwargs), Transformer(depth = freq_transformer_depth, rotary_embed = freq_rotary_embed, **transformer_kwargs) ])) self.stft_kwargs = dict( n_fft = stft_n_fft, hop_length = stft_hop_length, win_length = stft_win_length, normalized = stft_normalized ) freqs = torch.stft(torch.randn(1, 1024), **self.stft_kwargs, return_complex = True).shape[1] assert len(freqs_per_bands) > 1 assert sum(freqs_per_bands) == freqs, f'the number of freqs in the bands must equal {freqs} based on the STFT settings' freqs_per_bands_with_complex = tuple(2 * f for f in freqs_per_bands) self.band_split = BandSplit( dim = dim, dim_inputs = freqs_per_bands_with_complex ) self.mask_estimator = MaskEstimator( dim = dim, dim_inputs = freqs_per_bands_with_complex, depth = mask_estimator_depth ) # for the multi-resolution stft loss self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes self.multi_stft_n_fft = stft_n_fft self.multi_stft_kwargs = dict( hop_length = multi_stft_hop_size, normalized = multi_stft_normalized ) def forward( self, raw_audio, target = None, return_loss_breakdown = False ): """ einops b - batch f - freq t - time c - complex (2) d - feature dimension """ # to stft stft_repr = torch.stft(raw_audio, **self.stft_kwargs, return_complex = True) stft_repr = torch.view_as_real(stft_repr) x = rearrange(stft_repr, 'b f t c -> b t (f c)') x = self.band_split(x) # axial / hierarchical attention for time_transformer, freq_transformer in self.layers: x = rearrange(x, 'b t f d -> b f t d') x, ps = pack([x], 'b * d') x = time_transformer(x) x, = unpack(x, ps, 'b * d') x = rearrange(x, 'b f t d -> b t f d') x, ps = pack([x], 'b * d') x = freq_transformer(x) x, = unpack(x, ps, 'b * d') mask = self.mask_estimator(x) mask = rearrange(mask, 'b t (f c) -> b f t c', c = 2) # modulate frequency representation stft_repr = stft_repr * mask # istft stft_repr = torch.view_as_complex(stft_repr) recon_audio = torch.istft(stft_repr, **self.stft_kwargs, return_complex = False) # if a target is passed in, calculate loss for learning if not exists(target): return recon_audio target = target[..., :recon_audio.shape[-1]] # protect against lost length on istft loss = F.l1_loss(recon_audio, target) multi_stft_resolution_loss = 0. for window_size in self.multi_stft_resolutions_window_sizes: res_stft_kwargs = dict( n_fft = max(window_size, self.multi_stft_n_fft), # not sure what n_fft is across multi resolution stft win_length = window_size, return_complex = True, **self.multi_stft_kwargs, ) recon_Y = torch.stft(recon_audio, **res_stft_kwargs) target_Y = torch.stft(target, **res_stft_kwargs) multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y) weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight total_loss = loss + weighted_multi_resolution_loss if not return_loss_breakdown: return total_loss return total_loss, (loss, multi_stft_resolution_loss)
class BlackHole(object): def __setattr__(self, name, value): pass def __call__(self, *args, **kwargs): return self def __getattr__(self, name): return self def seed_all(seed): torch.backends.cudnn.deterministic = True torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed) random.seed(seed) def recursive_to(obj, device): if isinstance(obj, torch.Tensor): try: return obj.cuda(device=device, non_blocking=True) except RuntimeError: return obj.to(device) elif isinstance(obj, list): return [recursive_to(o, device=device) for o in obj] elif isinstance(obj, tuple): return (recursive_to(o, device=device) for o in obj) elif isinstance(obj, dict): return {k: recursive_to(v, device=device) for k, v in obj.items()} else: return obj
NON_STANDARD_SUBSTITUTIONS = { '2AS':'ASP', '3AH':'HIS', '5HP':'GLU', 'ACL':'ARG', 'AGM':'ARG', 'AIB':'ALA', 'ALM':'ALA', 'ALO':'THR', 'ALY':'LYS', 'ARM':'ARG', 'ASA':'ASP', 'ASB':'ASP', 'ASK':'ASP', 'ASL':'ASP', 'ASQ':'ASP', 'AYA':'ALA', 'BCS':'CYS', 'BHD':'ASP', 'BMT':'THR', 'BNN':'ALA', 'BUC':'CYS', 'BUG':'LEU', 'C5C':'CYS', 'C6C':'CYS', 'CAS':'CYS', 'CCS':'CYS', 'CEA':'CYS', 'CGU':'GLU', 'CHG':'ALA', 'CLE':'LEU', 'CME':'CYS', 'CSD':'ALA', 'CSO':'CYS', 'CSP':'CYS', 'CSS':'CYS', 'CSW':'CYS', 'CSX':'CYS', 'CXM':'MET', 'CY1':'CYS', 'CY3':'CYS', 'CYG':'CYS', 'CYM':'CYS', 'CYQ':'CYS', 'DAH':'PHE', 'DAL':'ALA', 'DAR':'ARG', 'DAS':'ASP', 'DCY':'CYS', 'DGL':'GLU', 'DGN':'GLN', 'DHA':'ALA', 'DHI':'HIS', 'DIL':'ILE', 'DIV':'VAL', 'DLE':'LEU', 'DLY':'LYS', 'DNP':'ALA', 'DPN':'PHE', 'DPR':'PRO', 'DSN':'SER', 'DSP':'ASP', 'DTH':'THR', 'DTR':'TRP', 'DTY':'TYR', 'DVA':'VAL', 'EFC':'CYS', 'FLA':'ALA', 'FME':'MET', 'GGL':'GLU', 'GL3':'GLY', 'GLZ':'GLY', 'GMA':'GLU', 'GSC':'GLY', 'HAC':'ALA', 'HAR':'ARG', 'HIC':'HIS', 'HIP':'HIS', 'HMR':'ARG', 'HPQ':'PHE', 'HTR':'TRP', 'HYP':'PRO', 'IAS':'ASP', 'IIL':'ILE', 'IYR':'TYR', 'KCX':'LYS', 'LLP':'LYS', 'LLY':'LYS', 'LTR':'TRP', 'LYM':'LYS', 'LYZ':'LYS', 'MAA':'ALA', 'MEN':'ASN', 'MHS':'HIS', 'MIS':'SER', 'MLE':'LEU', 'MPQ':'GLY', 'MSA':'GLY', 'MSE':'MET', 'MVA':'VAL', 'NEM':'HIS', 'NEP':'HIS', 'NLE':'LEU', 'NLN':'LEU', 'NLP':'LEU', 'NMC':'GLY', 'OAS':'SER', 'OCS':'CYS', 'OMT':'MET', 'PAQ':'TYR', 'PCA':'GLU', 'PEC':'CYS', 'PHI':'PHE', 'PHL':'PHE', 'PR3':'CYS', 'PRR':'ALA', 'PTR':'TYR', 'PYX':'CYS', 'SAC':'SER', 'SAR':'GLY', 'SCH':'CYS', 'SCS':'CYS', 'SCY':'CYS', 'SEL':'SER', 'SEP':'SER', 'SET':'SER', 'SHC':'CYS', 'SHR':'LYS', 'SMC':'CYS', 'SOC':'CYS', 'STY':'TYR', 'SVA':'SER', 'TIH':'ALA', 'TPL':'TRP', 'TPO':'THR', 'TPQ':'ALA', 'TRG':'LYS', 'TRO':'TRP', 'TYB':'TYR', 'TYI':'TYR', 'TYQ':'TYR', 'TYS':'TYR', 'TYY':'TYR' } RESIDUE_SIDECHAIN_POSTFIXES = { 'A': ['B'], 'R': ['B', 'G', 'D', 'E', 'Z', 'H1', 'H2'], 'N': ['B', 'G', 'D1', 'D2'], 'D': ['B', 'G', 'D1', 'D2'], 'C': ['B', 'G'], 'E': ['B', 'G', 'D', 'E1', 'E2'], 'Q': ['B', 'G', 'D', 'E1', 'E2'], 'G': [], 'H': ['B', 'G', 'D1', 'D2', 'E1', 'E2'], 'I': ['B', 'G1', 'G2', 'D1'], 'L': ['B', 'G', 'D1', 'D2'], 'K': ['B', 'G', 'D', 'E', 'Z'], 'M': ['B', 'G', 'D', 'E'], 'F': ['B', 'G', 'D1', 'D2', 'E1', 'E2', 'Z'], 'P': ['B', 'G', 'D'], 'S': ['B', 'G'], 'T': ['B', 'G1', 'G2'], 'W': ['B', 'G', 'D1', 'D2', 'E1', 'E2', 'E3', 'Z2', 'Z3', 'H2'], 'Y': ['B', 'G', 'D1', 'D2', 'E1', 'E2', 'Z', 'H'], 'V': ['B', 'G1', 'G2'], } GLY_INDEX = 5 ATOM_N, ATOM_CA, ATOM_C, ATOM_O, ATOM_CB = 0, 1, 2, 3, 4 def augmented_three_to_one(three): if three in NON_STANDARD_SUBSTITUTIONS: three = NON_STANDARD_SUBSTITUTIONS[three] return three_to_one(three) def augmented_three_to_index(three): if three in NON_STANDARD_SUBSTITUTIONS: three = NON_STANDARD_SUBSTITUTIONS[three] return three_to_index(three) def augmented_is_aa(three): if three in NON_STANDARD_SUBSTITUTIONS: three = NON_STANDARD_SUBSTITUTIONS[three] return is_aa(three, standard=True) def is_hetero_residue(res): return len(res.id[0].strip()) > 0 def get_atom_name_postfix(atom): name = atom.get_name() if name in ('N', 'CA', 'C', 'O'): return name if name[-1].isnumeric(): return name[-2:] else: return name[-1:] def get_residue_pos14(res): pos14 = torch.full([14, 3], float('inf')) suffix_to_atom = {get_atom_name_postfix(a):a for a in res.get_atoms()} atom_order = ['N', 'CA', 'C', 'O'] + RESIDUE_SIDECHAIN_POSTFIXES[augmented_three_to_one(res.get_resname())] for i, atom_suffix in enumerate(atom_order): if atom_suffix not in suffix_to_atom: continue pos14[i,0], pos14[i,1], pos14[i,2] = suffix_to_atom[atom_suffix].get_coord().tolist() return pos14 def parse_pdb(path, model_id=0): warnings.simplefilter('ignore', BiopythonWarning) parser = PDBParser() structure = parser.get_structure(None, path) return parse_complex(structure, model_id) def parse_complex(structure, model_id=None): if model_id is not None: structure = structure[model_id] chains = Selection.unfold_entities(structure, 'C') aa, resseq, icode, seq = [], [], [], [] pos14, pos14_mask = [], [] chain_id, chain_seq = [], [] for i, chain in enumerate(chains): seq_this = 0 for res in chain: resname = res.get_resname() if not augmented_is_aa(resname): continue if not (res.has_id('CA') and res.has_id('C') and res.has_id('N')): continue # Chain chain_id.append(chain.get_id()) chain_seq.append(i+1) # Residue types restype = augmented_three_to_index(resname) aa.append(restype) # Atom coordinates pos14_this = get_residue_pos14(res) pos14_mask_this = pos14_this.isfinite() pos14.append(pos14_this.nan_to_num(posinf=99999)) pos14_mask.append(pos14_mask_this) # Sequential number resseq_this = int(res.get_id()[1]) icode_this = res.get_id()[2] if seq_this == 0: seq_this = 1 else: d_resseq = resseq_this - resseq[-1] if d_resseq == 0: seq_this += 1 else: seq_this += d_resseq resseq.append(resseq_this) icode.append(icode_this) seq.append(seq_this) if len(aa) == 0: return None return { 'name': structure.get_id(), # Chain 'chain_id': ''.join(chain_id), 'chain_seq': torch.LongTensor(chain_seq), # Sequence 'aa': torch.LongTensor(aa), 'resseq': torch.LongTensor(resseq), 'icode': ''.join(icode), 'seq': torch.LongTensor(seq), # Atom positions 'pos14': torch.stack(pos14), 'pos14_mask': torch.stack(pos14_mask), }
class PaddingCollate(object): def __init__(self, length_ref_key='mutation_mask', pad_values={'aa': 20, 'pos14': float('999'), 'icode': ' ', 'chain_id': '-'}, donot_pad={'foldx'}, eight=False): super().__init__() self.length_ref_key = length_ref_key self.pad_values = pad_values self.donot_pad = donot_pad self.eight = eight def _pad_last(self, x, n, value=0): if isinstance(x, torch.Tensor): assert x.size(0) <= n if x.size(0) == n: return x pad_size = [n - x.size(0)] + list(x.shape[1:]) pad = torch.full(pad_size, fill_value=value).to(x) return torch.cat([x, pad], dim=0) elif isinstance(x, list): pad = [value] * (n - len(x)) return x + pad elif isinstance(x, str): if value == 0: # Won't pad strings if not specified return x pad = value * (n - len(x)) return x + pad elif isinstance(x, dict): padded = {} for k, v in x.items(): if k in self.donot_pad: padded[k] = v else: padded[k] = self._pad_last(v, n, value=self._get_pad_value(k)) return padded else: return x @staticmethod def _get_pad_mask(l, n): return torch.cat([ torch.ones([l], dtype=torch.bool), torch.zeros([n-l], dtype=torch.bool) ], dim=0) def _get_pad_value(self, key): if key not in self.pad_values: return 0 return self.pad_values[key] def __call__(self, data_list): max_length = max([data[self.length_ref_key].size(0) for data in data_list]) if self.eight: max_length = math.ceil(max_length / 8) * 8 data_list_padded = [] for data in data_list: data_padded = { k: self._pad_last(v, max_length, value=self._get_pad_value(k)) for k, v in data.items() if k in ('wt', 'mut', 'ddG', 'mutation_mask', 'index', 'mutation') } data_padded['mask'] = self._get_pad_mask(data[self.length_ref_key].size(0), max_length) data_list_padded.append(data_padded) return default_collate(data_list_padded) def _mask_list(l, mask): return [l[i] for i in range(len(l)) if mask[i]] def _mask_string(s, mask): return ''.join([s[i] for i in range(len(s)) if mask[i]]) def _mask_dict_recursively(d, mask): out = {} for k, v in d.items(): if isinstance(v, torch.Tensor) and v.size(0) == mask.size(0): out[k] = v[mask] elif isinstance(v, list) and len(v) == mask.size(0): out[k] = _mask_list(v, mask) elif isinstance(v, str) and len(v) == mask.size(0): out[k] = _mask_string(v, mask) elif isinstance(v, dict): out[k] = _mask_dict_recursively(v, mask) else: out[k] = v return out class KnnResidue(object): def __init__(self, num_neighbors=128): super().__init__() self.num_neighbors = num_neighbors def __call__(self, data): pos_CA = data['wt']['pos14'][:, ATOM_CA] pos_CA_mut = pos_CA[data['mutation_mask']] diff = pos_CA_mut.view(1, -1, 3) - pos_CA.view(-1, 1, 3) dist = torch.linalg.norm(diff, dim=-1) try: mask = torch.zeros([dist.size(0)], dtype=torch.bool) mask[ dist.min(dim=1)[0].argsort()[:self.num_neighbors] ] = True except IndexError as e: print(data) raise e return _mask_dict_recursively(data, mask) def load_wt_mut_pdb_pair(wt_path, mut_path): data_wt = parse_pdb(wt_path) data_mut = parse_pdb(mut_path) transform = KnnResidue() collate_fn = PaddingCollate() mutation_mask = (data_wt['aa'] != data_mut['aa']) batch = collate_fn([transform({'wt': data_wt, 'mut': data_mut, 'mutation_mask': mutation_mask})]) return batch
class ComplexEncoder(nn.Module): def __init__(self, cfg): super().__init__() self.cfg = cfg self.relpos_embedding = nn.Embedding(cfg.max_relpos*2+2, cfg.pair_feat_dim) self.residue_encoder = PerResidueEncoder(cfg.node_feat_dim) if cfg.geomattn is not None: self.ga_encoder = GAEncoder( node_feat_dim = cfg.node_feat_dim, pair_feat_dim = cfg.pair_feat_dim, num_layers = cfg.geomattn.num_layers, spatial_attn_mode = cfg.geomattn.spatial_attn_mode, ) else: self.out_mlp = nn.Sequential( nn.Linear(cfg.node_feat_dim, cfg.node_feat_dim), nn.ReLU(), nn.Linear(cfg.node_feat_dim, cfg.node_feat_dim), nn.ReLU(), nn.Linear(cfg.node_feat_dim, cfg.node_feat_dim), ) def forward(self, pos14, aa, seq, chain, mask_atom): """ Args: pos14: (N, L, 14, 3). aa: (N, L). seq: (N, L). chain: (N, L). mask_atom: (N, L, 14) Returns: (N, L, node_ch) """ same_chain = (chain[:, None, :] == chain[:, :, None]) # (N, L, L) relpos = (seq[:, None, :] - seq[:, :, None]).clamp(min=-self.cfg.max_relpos, max=self.cfg.max_relpos) + self.cfg.max_relpos # (N, L, L) relpos = torch.where(same_chain, relpos, torch.full_like(relpos, fill_value=self.cfg.max_relpos*2+1)) pair_feat = self.relpos_embedding(relpos) # (N, L, L, pair_ch) R = construct_3d_basis(pos14[:, :, ATOM_CA], pos14[:, :, ATOM_C], pos14[:, :, ATOM_N]) # Residue encoder res_feat = self.residue_encoder(aa, pos14, mask_atom) # Geom encoder t = pos14[:, :, ATOM_CA] mask_residue = mask_atom[:, :, ATOM_CA] res_feat = self.ga_encoder(R, t, get_pos_CB(pos14, mask_atom), res_feat, pair_feat, mask_residue) return res_feat class DDGReadout(nn.Module): def __init__(self, feat_dim): super().__init__() self.mlp = nn.Sequential( nn.Linear(feat_dim*2, feat_dim), nn.ReLU(), nn.Linear(feat_dim, feat_dim), nn.ReLU(), nn.Linear(feat_dim, feat_dim), nn.ReLU(), nn.Linear(feat_dim, feat_dim) ) self.project = nn.Linear(feat_dim, 1, bias=False) def forward(self, node_feat_wt, node_feat_mut, mask=None): """ Args: node_feat_wt: (N, L, F). node_feat_mut: (N, L, F). mask: (N, L). """ feat_wm = torch.cat([node_feat_wt, node_feat_mut], dim=-1) feat_mw = torch.cat([node_feat_mut, node_feat_wt], dim=-1) feat_diff = self.mlp(feat_wm) - self.mlp(feat_mw) # (N, L, F) # feat_diff = self.mlp(node_feat_wt) - self.mlp(node_feat_mut) per_residue_ddg = self.project(feat_diff).squeeze(-1) # (N, L) if mask is not None: per_residue_ddg = per_residue_ddg * mask ddg = per_residue_ddg.sum(dim=1) # (N,) return ddg class DDGPredictor(nn.Module): def __init__(self, cfg): super().__init__() self.encoder = ComplexEncoder(cfg) self.ddG_readout = DDGReadout(cfg.node_feat_dim) def forward(self, complex_wt, complex_mut, ddG_true=None): mask_atom_wt = complex_wt['pos14_mask'].all(dim=-1) # (N, L, 14) mask_atom_mut = complex_mut['pos14_mask'].all(dim=-1) feat_wt = self.encoder(complex_wt['pos14'], complex_wt['aa'], complex_wt['seq'], complex_wt['chain_seq'], mask_atom_wt) feat_mut = self.encoder(complex_mut['pos14'], complex_mut['aa'], complex_mut['seq'], complex_mut['chain_seq'], mask_atom_mut) mask_res = mask_atom_wt[:, :, ATOM_CA] ddG_pred = self.ddG_readout(feat_wt, feat_mut, mask_res) # One mask is enough if ddG_true is None: return ddG_pred else: losses = { 'ddG': F.mse_loss(ddG_pred, ddG_true), } return losses, ddG_pred
def _alpha_from_logits(logits, mask, inf=1e5): """ Args: logits: Logit matrices, (N, L_i, L_j, num_heads). mask: Masks, (N, L). Returns: alpha: Attention weights. """ N, L, _, _ = logits.size() mask_row = mask.view(N, L, 1, 1).expand_as(logits) # (N, L, *, *) mask_pair = mask_row * mask_row.permute(0, 2, 1, 3) # (N, L, L, *) logits = torch.where(mask_pair, logits, logits-inf) alpha = torch.softmax(logits, dim=2) # (N, L, L, num_heads) alpha = torch.where(mask_row, alpha, torch.zeros_like(alpha)) return alpha def _heads(x, n_heads, n_ch): """ Args: x: (..., num_heads * num_channels) Returns: (..., num_heads, num_channels) """ s = list(x.size())[:-1] + [n_heads, n_ch] return x.view(*s) class GeometricAttention(nn.Module): def __init__(self, node_feat_dim, pair_feat_dim, spatial_attn_mode='CB', value_dim=16, query_key_dim=16, num_query_points=8, num_value_points=8, num_heads=12): super().__init__() self.node_feat_dim = node_feat_dim self.pair_feat_dim = pair_feat_dim self.value_dim = value_dim self.query_key_dim = query_key_dim self.num_query_points = num_query_points self.num_value_points = num_value_points self.num_heads = num_heads assert spatial_attn_mode in ('CB', 'vpoint') self.spatial_attn_mode = spatial_attn_mode # Node self.proj_query = nn.Linear(node_feat_dim, query_key_dim*num_heads, bias=False) self.proj_key = nn.Linear(node_feat_dim, query_key_dim*num_heads, bias=False) self.proj_value = nn.Linear(node_feat_dim, value_dim*num_heads, bias=False) # Pair self.proj_pair_bias = nn.Linear(pair_feat_dim, num_heads, bias=False) # Spatial self.spatial_coef = nn.Parameter(torch.full([1, 1, 1, self.num_heads], fill_value=np.log(np.exp(1.) - 1.)), requires_grad=True) if spatial_attn_mode == 'vpoint': self.proj_query_point = nn.Linear(node_feat_dim, num_query_points*num_heads*3, bias=False) self.proj_key_point = nn.Linear(node_feat_dim, num_query_points*num_heads*3, bias=False) self.proj_value_point = nn.Linear(node_feat_dim, num_value_points*num_heads*3, bias=False) # Output if spatial_attn_mode == 'CB': self.out_transform = nn.Linear( in_features = (num_heads*pair_feat_dim) + (num_heads*value_dim) + (num_heads*(3+3+1)), out_features = node_feat_dim, ) elif spatial_attn_mode == 'vpoint': self.out_transform = nn.Linear( in_features = (num_heads*pair_feat_dim) + (num_heads*value_dim) + (num_heads*num_value_points*(3+3+1)), out_features = node_feat_dim, ) self.layer_norm = nn.LayerNorm(node_feat_dim) def _node_logits(self, x): query_l = _heads(self.proj_query(x), self.num_heads, self.query_key_dim) # (N, L, n_heads, qk_ch) key_l = _heads(self.proj_key(x), self.num_heads, self.query_key_dim) # (N, L, n_heads, qk_ch) query_l = query_l.permute(0, 2, 1, 3) # (N,L1,H,C) -> (N,H,L1,C) key_l = key_l.permute(0, 2, 3, 1) # (N,L2,H,C) -> (N,H,C,L2) logits = torch.matmul(query_l, key_l) # (N,H,L1,L2) logits = logits.permute(0, 2, 3, 1) # (N,L1,L2,H) # logits = (query_l.unsqueeze(2) * key_l.unsqueeze(1) * (1 / np.sqrt(self.query_key_dim))).sum(-1) # (N, L, L, num_heads) return logits def _pair_logits(self, z): logits_pair = self.proj_pair_bias(z) return logits_pair def _beta_logits(self, R, t, p_CB): N, L, _ = t.size() qk = p_CB[:, :, None, :].expand(N, L, self.num_heads, 3) sum_sq_dist = ((qk.unsqueeze(2) - qk.unsqueeze(1)) ** 2).sum(-1) # (N, L, L, n_heads) gamma = F.softplus(self.spatial_coef) logtis_beta = sum_sq_dist * ((-1 * gamma * np.sqrt(2 / 9)) / 2) return logtis_beta def _spatial_logits(self, R, t, x): N, L, _ = t.size() # Query query_points = _heads(self.proj_query_point(x), self.num_heads*self.num_query_points, 3) # (N, L, n_heads * n_pnts, 3) query_points = local_to_global(R, t, query_points) # Global query coordinates, (N, L, n_heads * n_pnts, 3) query_s = query_points.reshape(N, L, self.num_heads, -1) # (N, L, n_heads, n_pnts*3) # Key key_points = _heads(self.proj_key_point(x), self.num_heads*self.num_query_points, 3) # (N, L, 3, n_heads * n_pnts) key_points = local_to_global(R, t, key_points) # Global key coordinates, (N, L, n_heads * n_pnts, 3) key_s = key_points.reshape(N, L, self.num_heads, -1) # (N, L, n_heads, n_pnts*3) # Q-K Product sum_sq_dist = ((query_s.unsqueeze(2) - key_s.unsqueeze(1)) ** 2).sum(-1) # (N, L, L, n_heads) gamma = F.softplus(self.spatial_coef) logits_spatial = sum_sq_dist * ((-1 * gamma * np.sqrt(2 / (9 * self.num_query_points))) / 2) # (N, L, L, n_heads) return logits_spatial def _pair_aggregation(self, alpha, z): N, L = z.shape[:2] feat_p2n = alpha.unsqueeze(-1) * z.unsqueeze(-2) # (N, L, L, n_heads, C) feat_p2n = feat_p2n.sum(dim=2) # (N, L, n_heads, C) return feat_p2n.reshape(N, L, -1) def _node_aggregation(self, alpha, x): N, L = x.shape[:2] value_l = _heads(self.proj_value(x), self.num_heads, self.query_key_dim) # (N, L, n_heads, v_ch) feat_node = alpha.unsqueeze(-1) * value_l.unsqueeze(1) # (N, L, L, n_heads, *) @ (N, *, L, n_heads, v_ch) feat_node = feat_node.sum(dim=2) # (N, L, n_heads, v_ch) return feat_node.reshape(N, L, -1) def _beta_aggregation(self, alpha, R, t, p_CB, x): N, L, _ = t.size() v = p_CB[:, :, None, :].expand(N, L, self.num_heads, 3) # (N, L, n_heads, 3) aggr = alpha.reshape(N, L, L, self.num_heads, 1) * v.unsqueeze(1) # (N, *, L, n_heads, 3) aggr = aggr.sum(dim=2) feat_points = global_to_local(R, t, aggr) # (N, L, n_heads, 3) feat_distance = feat_points.norm(dim=-1) feat_direction = normalize_vector(feat_points, dim=-1, eps=1e-4) feat_spatial = torch.cat([ feat_points.reshape(N, L, -1), feat_distance.reshape(N, L, -1), feat_direction.reshape(N, L, -1), ], dim=-1) return feat_spatial def _spatial_aggregation(self, alpha, R, t, x): N, L, _ = t.size() value_points = _heads(self.proj_value_point(x), self.num_heads*self.num_value_points, 3) # (N, L, n_heads * n_v_pnts, 3) value_points = local_to_global(R, t, value_points.reshape(N, L, self.num_heads, self.num_value_points, 3)) # (N, L, n_heads, n_v_pnts, 3) aggr_points = alpha.reshape(N, L, L, self.num_heads, 1, 1) * value_points.unsqueeze(1) # (N, *, L, n_heads, n_pnts, 3) aggr_points = aggr_points.sum(dim=2) # (N, L, n_heads, n_pnts, 3) feat_points = global_to_local(R, t, aggr_points) # (N, L, n_heads, n_pnts, 3) feat_distance = feat_points.norm(dim=-1) # (N, L, n_heads, n_pnts) feat_direction = normalize_vector(feat_points, dim=-1, eps=1e-4) # (N, L, n_heads, n_pnts, 3) feat_spatial = torch.cat([ feat_points.reshape(N, L, -1), feat_distance.reshape(N, L, -1), feat_direction.reshape(N, L, -1), ], dim=-1) return feat_spatial def forward_beta(self, R, t, p_CB, x, z, mask): """ Args: R: Frame basis matrices, (N, L, 3, 3_index). t: Frame external (absolute) coordinates, (N, L, 3). x: Node-wise features, (N, L, F). z: Pair-wise features, (N, L, L, C). mask: Masks, (N, L). Returns: x': Updated node-wise features, (N, L, F). """ # Attention logits logits_node = self._node_logits(x) logits_pair = self._pair_logits(z) logits_spatial = self._beta_logits(R, t, p_CB) # Summing logits up and apply `softmax`. logits_sum = logits_node + logits_pair + logits_spatial alpha = _alpha_from_logits(logits_sum * np.sqrt(1 / 3), mask) # (N, L, L, n_heads) # Aggregate features feat_p2n = self._pair_aggregation(alpha, z) feat_node = self._node_aggregation(alpha, x) feat_spatial = self._beta_aggregation(alpha, R, t, p_CB, x) # Finally feat_all = self.out_transform(torch.cat([feat_p2n, feat_node, feat_spatial], dim=-1)) # (N, L, F) feat_all = mask_zero(mask.unsqueeze(-1), feat_all) x_updated = self.layer_norm(x + feat_all) return x_updated def forward_vpoint(self, R, t, p_CB, x, z, mask): """ Args: R: Frame basis matrices, (N, L, 3, 3_index). t: Frame external (absolute) coordinates, (N, L, 3). x: Node-wise features, (N, L, F). z: Pair-wise features, (N, L, L, C). mask: Masks, (N, L). Returns: x': Updated node-wise features, (N, L, F). """ # Attention logits logits_node = self._node_logits(x) logits_pair = self._pair_logits(z) logits_spatial = self._spatial_logits(R, t, x) # Summing logits up and apply `softmax`. logits_sum = logits_node + logits_pair + logits_spatial alpha = _alpha_from_logits(logits_sum * np.sqrt(1 / 3), mask) # (N, L, L, n_heads) # Aggregate features feat_p2n = self._pair_aggregation(alpha, z) feat_node = self._node_aggregation(alpha, x) feat_spatial = self._spatial_aggregation(alpha, R, t, x) # Finally feat_all = self.out_transform(torch.cat([feat_p2n, feat_node, feat_spatial], dim=-1)) # (N, L, F) feat_all = mask_zero(mask.unsqueeze(-1), feat_all) x_updated = self.layer_norm(x + feat_all) return x_updated def forward(self, R, t, p_CB, x, z, mask): if self.spatial_attn_mode == 'CB': return self.forward_beta(R, t, p_CB, x, z, mask) else: return self.forward_vpoint(R, t, p_CB, x, z, mask) class GAEncoder(nn.Module): def __init__(self, node_feat_dim, pair_feat_dim, num_layers, spatial_attn_mode='CB'): super().__init__() self.blocks = nn.ModuleList([ GeometricAttention(node_feat_dim, pair_feat_dim, spatial_attn_mode=spatial_attn_mode) for _ in range(num_layers) ]) def forward(self, R, t, p_CB, x, z, mask): for block in self.blocks: x = block(R, t, p_CB, x, z, mask) # Residual connection within the block return x
class PerResidueEncoder(nn.Module): def __init__(self, feat_dim): super().__init__() self.aatype_embed = nn.Embedding(21, feat_dim) self.torsion_embed = PositionalEncoding() self.mlp = nn.Sequential( nn.Linear(21*14*3 + feat_dim, feat_dim * 2), nn.ReLU(), nn.Linear(feat_dim * 2, feat_dim), nn.ReLU(), nn.Linear(feat_dim, feat_dim), nn.ReLU(), nn.Linear(feat_dim, feat_dim) ) def forward(self, aa, pos14, atom_mask): """ Args: aa: (N, L). pos14: (N, L, 14, 3). atom_mask: (N, L, 14). """ N, L = aa.size() R = construct_3d_basis(pos14[:, :, 1], pos14[:, :, 2], pos14[:, :, 0]) # (N, L, 3, 3) t = pos14[:, :, 1] # (N, L, 3) crd14 = global_to_local(R, t, pos14) # (N, L, 14, 3) crd14_mask = atom_mask[:, :, :, None].expand_as(crd14) crd14 = torch.where(crd14_mask, crd14, torch.zeros_like(crd14)) aa_expand = aa[:, :, None, None, None].expand(N, L, 21, 14, 3) rng_expand = torch.arange(0, 21)[None, None, :, None, None].expand(N, L, 21, 14, 3).to(aa_expand) place_mask = (aa_expand == rng_expand) crd_expand = crd14[:, :, None, :, :].expand(N, L, 21, 14, 3) crd_expand = torch.where(place_mask, crd_expand, torch.zeros_like(crd_expand)) crd_feat = crd_expand.reshape(N, L, 21 * 14 * 3) aa_feat = self.aatype_embed(aa) # (N, L, feat) out_feat = self.mlp(torch.cat([crd_feat, aa_feat], dim=-1)) return out_feat
def get_pos_CB(pos14, atom_mask): """ Args: pos14: (N, L, 14, 3) atom_mask: (N, L, 14) """ N, L = pos14.shape[:2] mask_CB = atom_mask[:, :, ATOM_CB] # (N, L) mask_CB = mask_CB[:, :, None].expand(N, L, 3) pos_CA = pos14[:, :, ATOM_CA] # (N, L, 3) pos_CB = pos14[:, :, ATOM_CB] return torch.where(mask_CB, pos_CB, pos_CA) def mask_zero(mask, value): return torch.where(mask, value, torch.zeros_like(value)) class PositionalEncoding(nn.Module): def __init__(self, num_funcs=6): super().__init__() self.num_funcs = num_funcs self.register_buffer('freq_bands', 2.0 ** torch.linspace(0.0, num_funcs-1, num_funcs)) def get_out_dim(self, in_dim): return in_dim * (2 * self.num_funcs + 1) def forward(self, x): """ Args: x: (..., d). """ shape = list(x.shape[:-1]) + [-1] x = x.unsqueeze(-1) # (..., d, 1) code = torch.cat([x, torch.sin(x * self.freq_bands), torch.cos(x * self.freq_bands)], dim=-1) # (..., d, 2f+1) code = code.reshape(shape) return code def safe_norm(x, dim=-1, keepdim=False, eps=1e-8, sqrt=True): out = torch.clamp(torch.sum(torch.square(x), dim=dim, keepdim=keepdim), min=eps) return torch.sqrt(out) if sqrt else out def normalize_vector(v, dim, eps=1e-6): return v / (torch.linalg.norm(v, ord=2, dim=dim, keepdim=True) + eps) def project_v2v(v, e, dim): """ Description: Project vector `v` onto vector `e`. Args: v: (N, L, 3). e: (N, L, 3). """ return (e * v).sum(dim=dim, keepdim=True) * e def construct_3d_basis(center, p1, p2): """ Args: center: (N, L, 3), usually the position of C_alpha. p1: (N, L, 3), usually the position of C. p2: (N, L, 3), usually the position of N. Returns A batch of orthogonal basis matrix, (N, L, 3, 3cols_index). The matrix is composed of 3 column vectors: [e1, e2, e3]. """ v1 = p1 - center # (N, L, 3) e1 = normalize_vector(v1, dim=-1) v2 = p2 - center # (N, L, 3) u2 = v2 - project_v2v(v2, e1, dim=-1) e2 = normalize_vector(u2, dim=-1) e3 = torch.cross(e1, e2, dim=-1) # (N, L, 3) mat = torch.cat([ e1.unsqueeze(-1), e2.unsqueeze(-1), e3.unsqueeze(-1) ], dim=-1) # (N, L, 3, 3_index) return mat def local_to_global(R, t, p): """ Description: Convert local (internal) coordinates to global (external) coordinates q. q <- Rp + t Args: R: (N, L, 3, 3). t: (N, L, 3). p: Local coordinates, (N, L, ..., 3). Returns: q: Global coordinates, (N, L, ..., 3). """ assert p.size(-1) == 3 p_size = p.size() N, L = p_size[0], p_size[1] p = p.view(N, L, -1, 3).transpose(-1, -2) # (N, L, *, 3) -> (N, L, 3, *) q = torch.matmul(R, p) + t.unsqueeze(-1) # (N, L, 3, *) q = q.transpose(-1, -2).reshape(p_size) # (N, L, 3, *) -> (N, L, *, 3) -> (N, L, ..., 3) return q def global_to_local(R, t, q): """ Description: Convert global (external) coordinates q to local (internal) coordinates p. p <- R^{T}(q - t) Args: R: (N, L, 3, 3). t: (N, L, 3). q: Global coordinates, (N, L, ..., 3). Returns: p: Local coordinates, (N, L, ..., 3). """ assert q.size(-1) == 3 q_size = q.size() N, L = q_size[0], q_size[1] q = q.reshape(N, L, -1, 3).transpose(-1, -2) # (N, L, *, 3) -> (N, L, 3, *) if t is None: p = torch.matmul(R.transpose(-1, -2), q) # (N, L, 3, *) else: p = torch.matmul(R.transpose(-1, -2), (q - t.unsqueeze(-1))) # (N, L, 3, *) p = p.transpose(-1, -2).reshape(q_size) # (N, L, 3, *) -> (N, L, *, 3) -> (N, L, ..., 3) return p
sys.path.append(os.path.dirname(os.path.dirname(__file__))) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('wt_pdb', type=str) parser.add_argument('mut_pdb', type=str) parser.add_argument('--model', type=str, default='./data/model.pt') parser.add_argument('--device', type=str, default='cuda') args = parser.parse_args() batch = load_wt_mut_pdb_pair(args.wt_pdb, args.mut_pdb) batch = recursive_to(batch, args.device) ckpt = torch.load(args.model) config = ckpt['config'] weight = ckpt['model'] model = DDGPredictor(config.model).to(args.device) model.load_state_dict(weight) with torch.no_grad(): model.eval() pred = model(batch['wt'], batch['mut']) print('Predicted ddG: %.2f' % pred.item())
AoA = AttentionOnAttention
def exists(val): return val is not None def default(val, d): return val if exists(val) else d class AttentionOnAttention(nn.Module): def __init__( self, *, dim, dim_head = 64, heads = 8, dropout = 0., aoa_dropout = 0. ): super().__init__() inner_dim = dim_head * heads self.heads = heads self.scale = dim_head ** -0.5 self.to_q = nn.Linear(dim, inner_dim, bias = False) self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False) self.dropout = nn.Dropout(dropout) self.aoa = nn.Sequential( nn.Linear(2 * inner_dim, 2 * dim), nn.GLU(), nn.Dropout(aoa_dropout) ) def forward(self, x, context = None): h = self.heads q_ = self.to_q(x) context = default(context, x) kv = self.to_kv(context).chunk(2, dim = -1) # split heads q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q_, *kv)) dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale # attention attn = dots.softmax(dim = -1) attn = self.dropout(attn) # weighted average of values attn_out = einsum('b h i j, b h j d -> b h i d', attn, v) # concat heads out = rearrange(attn_out, 'b h n d -> b n (h d)', h = h) # attention on attention out = self.aoa(torch.cat((out, q_), dim = -1)) return out
# helpers def exists(val): return val is not None def batched_index_select(values, indices): last_dim = values.shape[-1] return values.gather(1, indices[:, :, None].expand(-1, -1, last_dim)) # helper classes class Residual(nn.Module): def __init__(self, fn): super().__init__() self.fn = fn def forward(self, x, **kwargs): return self.fn(x, **kwargs) + x class PreNorm(nn.Module): def __init__(self, dim, fn): super().__init__() self.fn = fn self.norm = nn.LayerNorm(dim) def forward(self, x, **kwargs): return self.fn(self.norm(x), **kwargs) class FeedForward(nn.Module): def __init__(self, dim, mult = 4, dropout = 0.): super().__init__() self.net = nn.Sequential( nn.Linear(dim, dim * mult), nn.GELU(), nn.Dropout(dropout), nn.Linear(dim * mult, dim) ) def forward(self, x, **kwargs): return self.net(x) # adjacent attention class class AdjacentAttention(nn.Module): def __init__( self, *, dim, dim_head = 64, heads = 4, dropout = 0. ): super().__init__() inner_dim = dim_head * heads self.scale = dim_head ** -0.5 self.heads = heads self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False) self.to_out = nn.Linear(inner_dim, dim) self.null_k = nn.Parameter(torch.randn(heads, dim_head)) self.null_v = nn.Parameter(torch.randn(heads, dim_head)) self.dropout = nn.Dropout(dropout) def forward( self, x, adj_kv_indices, mask ): b, n, d, h = *x.shape, self.heads flat_indices = repeat(adj_kv_indices, 'b n a -> (b h) (n a)', h = h) # derive query, key, value q, k, v = self.to_qkv(x).chunk(3, dim = -1) q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q, k, v)) # gather keys and values according to adjacency matrix k, v = map(lambda t: rearrange(t, 'b h n d -> (b h) n d'), (k, v)) k = batched_index_select(k, flat_indices) v = batched_index_select(v, flat_indices) k, v = map(lambda t: rearrange(t, '(b h) (n a) d -> b h n a d', h = h, n = n), (k, v)) # add null key / value, so a node can attend to nothing # have come across this in GNN literature as some other name nk, nv = map(lambda t: rearrange(t, 'h d -> () h () () d').expand(b, -1, n, 1, -1), (self.null_k, self.null_v)) k = torch.cat((nk, k), dim = -2) v = torch.cat((nv, v), dim = -2) mask = F.pad(mask, (1, 0), value = 1) # similarity of each node to its neighbors sim = einsum('b h n d, b h n a d -> b h n a', q, k) * self.scale # mask out neighbors that are just padding mask_value = -torch.finfo(sim.dtype).max mask = rearrange(mask.bool(), 'b n a -> b () n a') sim.masked_fill_(~mask.bool(), mask_value) # attention attn = sim.softmax(dim = -1) # dropout attn = self.dropout(attn) # get weighted average of the values of all neighbors out = einsum('b h n a, b h n a d -> b h n d', attn, v) out = rearrange(out, 'b h n d -> b n (h d)') # combine output return self.to_out(out) # adjacent network (layers of adjacent attention) class AdjacentAttentionNetwork(nn.Module): def __init__( self, *, dim, depth, dim_head = 64, heads = 4, num_neighbors_cutoff = None, num_global_nodes = 0, attn_dropout = 0., ff_dropout = 0. ): super().__init__() self.num_neighbors_cutoff = num_neighbors_cutoff self.layers = nn.ModuleList([]) for _ in range(depth): global_attn = PreNorm(dim, ISAB( dim = dim, heads = heads, num_induced_points = num_global_nodes )) if num_global_nodes > 0 else None self.layers.append(nn.ModuleList([ Residual(PreNorm(dim, AdjacentAttention( dim = dim, dim_head = dim_head, heads = heads, dropout = attn_dropout ))), global_attn, Residual(PreNorm(dim, FeedForward( dim = dim, dropout = ff_dropout ))) ])) def forward(self, x, adjacency_mat, mask = None): device, n = x.device, x.shape[1] diag = torch.eye(adjacency_mat.shape[-1], device = device).bool() adjacency_mat |= diag # nodes should pay attention itself (self-interacting) # zero out points on adjacency matrix # where the nodes are just padding if exists(mask): adjacency_mat &= (mask[:, :, None] * mask[:, None, :]) adj_mat = adjacency_mat.float() # if we don't set a hard limit to the number of neighbors: # - get the maximum number of neighbors and pad the rest of the nodes with less than that number of neighbors # else: # - randomly sample the cutoff number of neighbors for any node that exceeds the max # - this would be similar to random sparse attention (bigbird) # get the maximum number of neighbors max_neighbors = int(adj_mat.sum(dim = -1).max()) if exists(self.num_neighbors_cutoff) and max_neighbors > self.num_neighbors_cutoff: # to randomly sample the neighbors, add a small uniform noise to the mask and topk noise = torch.empty((n, n), device = device).uniform_(-0.01, 0.01) adj_mat = adj_mat + noise adj_mask, adj_kv_indices = adj_mat.topk(dim = -1, k = self.num_neighbors_cutoff) # cast the mask back to 0s and 1s adj_mask = (adj_mask > 0.5).float() else: # todo - get distribution of number of neighbors, and strategically break up attention (message passing) to multiple steps # - start with a bimodal num neighbors test case, then generalize # use topk to get all the neighbors # also pass the mask into the attention, as some neighbors will be just padding and not actually neighbors adj_mask, adj_kv_indices = adj_mat.topk(dim = -1, k = max_neighbors) for attn, global_attn, ff in self.layers: x = attn( x, adj_kv_indices = adj_kv_indices, mask = adj_mask ) if exists(global_attn): out, _ = global_attn(x, mask = mask) x = x + out x = ff(x) return x
logging.set_verbosity_error() def exists(val): return val is not None def map_values(fn, dictionary): return {k: fn(v) for k, v in dictionary.items()} CONTEXT_EMBED_USE_CPU = os.getenv('CONTEXT_EMBED_USE_CPU', None) is not None if CONTEXT_EMBED_USE_CPU: print('calculating context embed only on cpu') MODELS = dict( pubmed = dict( dim = 768, path = 'microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract', ) ) GLOBAL_VARIABLES = dict(model = None, tokenizer = None) def get_contextual_dim(model_name): assert model_name in MODELS return MODELS[model_name]['dim'] @run_once('init_transformer') def init_transformer(model_name): path = MODELS[model_name]['path'] GLOBAL_VARIABLES['tokenizer'] = AutoTokenizer.from_pretrained(path) model = AutoModelForMaskedLM.from_pretrained(path) if not CONTEXT_EMBED_USE_CPU: model = model.cuda() GLOBAL_VARIABLES['model'] = model @torch.no_grad() def tokenize_text( text, max_length = 256, model_name = 'pubmed', hidden_state_index = -1, return_cls_token = True ): init_transformer(model_name) model = GLOBAL_VARIABLES['model'] tokenizer = GLOBAL_VARIABLES['tokenizer'] encoding = tokenizer.batch_encode_plus( [text], add_special_tokens = True, padding = True, truncation = True, max_length = max_length, return_attention_mask = True, return_tensors = 'pt' ) if not CONTEXT_EMBED_USE_CPU: encoding = map_values(lambda t: t.cuda(), encoding) model.eval() with torch.no_grad(): outputs = model(**encoding, output_hidden_states = True) hidden_state = outputs.hidden_states[hidden_state_index][0] if return_cls_token: return hidden_state[0] return hidden_state.mean(dim = 0) def get_text_repr( texts, *, device, max_length = 256, model_name = 'pubmed', hidden_state_index = -1, return_cls_token = True, ): assert model_name in MODELS, f'{model_name} not found in available text transformers to use' if isinstance(texts, str): texts = [texts] get_context_repr_fn = cache_fn(tokenize_text, path = f'contexts/{model_name}') representations = [get_context_repr_fn(text, max_length = max_length, model_name = model_name, hidden_state_index = hidden_state_index, return_cls_token = return_cls_token) for text in texts] return torch.stack(representations).to(device)
# helpers functions def exists(x): return x is not None def default(val, d): if exists(val): return val return d() if callable(d) else d def cycle(dl): while True: for data in dl: yield data def has_int_squareroot(num): return (math.sqrt(num) ** 2) == num def num_to_groups(num, divisor): groups = num // divisor remainder = num % divisor arr = [divisor] * groups if remainder > 0: arr.append(remainder) return arr def convert_image_to(img_type, image): if image.mode != img_type: return image.convert(img_type) return image # small helper modules class Residual(nn.Module): def __init__(self, fn): super().__init__() self.fn = fn def forward(self, x, *args, **kwargs): return self.fn(x, *args, **kwargs) + x def Upsample(dim, dim_out = None): return nn.Sequential( nn.Upsample(scale_factor = 2, mode = 'nearest'), nn.Conv2d(dim, default(dim_out, dim), 3, padding = 1) ) def Downsample(dim, dim_out = None): return nn.Conv2d(dim, default(dim_out, dim), 4, 2, 1) class LayerNorm(nn.Module): def __init__(self, dim): super().__init__() self.g = nn.Parameter(torch.ones(1, dim, 1, 1)) def forward(self, x): eps = 1e-5 if x.dtype == torch.float32 else 1e-3 var = torch.var(x, dim = 1, unbiased = False, keepdim = True) mean = torch.mean(x, dim = 1, keepdim = True) return (x - mean) * (var + eps).rsqrt() * self.g class PreNorm(nn.Module): def __init__(self, dim, fn): super().__init__() self.fn = fn self.norm = LayerNorm(dim) def forward(self, x): x = self.norm(x) return self.fn(x) # positional embeds class LearnedSinusoidalPosEmb(nn.Module): def __init__(self, dim): super().__init__() assert (dim % 2) == 0 half_dim = dim // 2 self.weights = nn.Parameter(torch.randn(half_dim)) def forward(self, x): x = rearrange(x, 'b -> b 1') freqs = x * rearrange(self.weights, 'd -> 1 d') * 2 * math.pi fouriered = torch.cat((freqs.sin(), freqs.cos()), dim = -1) fouriered = torch.cat((x, fouriered), dim = -1) return fouriered # building block modules class Block(nn.Module): def __init__(self, dim, dim_out, groups = 8): super().__init__() self.proj = nn.Conv2d(dim, dim_out, 3, padding = 1) self.norm = nn.GroupNorm(groups, dim_out) self.act = nn.SiLU() def forward(self, x, scale_shift = None): x = self.proj(x) x = self.norm(x) if exists(scale_shift): scale, shift = scale_shift x = x * (scale + 1) + shift x = self.act(x) return x class ResnetBlock(nn.Module): def __init__(self, dim, dim_out, *, time_emb_dim = None, groups = 8): super().__init__() self.mlp = nn.Sequential( nn.SiLU(), nn.Linear(time_emb_dim, dim_out * 2) ) if exists(time_emb_dim) else None self.block1 = Block(dim, dim_out, groups = groups) self.block2 = Block(dim_out, dim_out, groups = groups) self.res_conv = nn.Conv2d(dim, dim_out, 1) if dim != dim_out else nn.Identity() def forward(self, x, time_emb = None): scale_shift = None if exists(self.mlp) and exists(time_emb): time_emb = self.mlp(time_emb) time_emb = rearrange(time_emb, 'b c -> b c 1 1') scale_shift = time_emb.chunk(2, dim = 1) h = self.block1(x, scale_shift = scale_shift) h = self.block2(h) return h + self.res_conv(x) class LinearAttention(nn.Module): def __init__(self, dim, heads = 4, dim_head = 32): super().__init__() self.scale = dim_head ** -0.5 self.heads = heads hidden_dim = dim_head * heads self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False) self.to_out = nn.Sequential( nn.Conv2d(hidden_dim, dim, 1), LayerNorm(dim) ) def forward(self, x): b, c, h, w = x.shape qkv = self.to_qkv(x).chunk(3, dim = 1) q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv) q = q.softmax(dim = -2) k = k.softmax(dim = -1) q = q * self.scale v = v / (h * w) context = torch.einsum('b h d n, b h e n -> b h d e', k, v) out = torch.einsum('b h d e, b h d n -> b h e n', context, q) out = rearrange(out, 'b h c (x y) -> b (h c) x y', h = self.heads, x = h, y = w) return self.to_out(out) class Attention(nn.Module): def __init__(self, dim, heads = 4, dim_head = 32): super().__init__() self.scale = dim_head ** -0.5 self.heads = heads hidden_dim = dim_head * heads self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False) self.to_out = nn.Conv2d(hidden_dim, dim, 1) def forward(self, x): b, c, h, w = x.shape qkv = self.to_qkv(x).chunk(3, dim = 1) q, k, v = map(lambda t: rearrange(t, 'b (h c) x y -> b h c (x y)', h = self.heads), qkv) q = q * self.scale sim = einsum('b h d i, b h d j -> b h i j', q, k) attn = sim.softmax(dim = -1) out = einsum('b h i j, b h d j -> b h i d', attn, v) out = rearrange(out, 'b h (x y) d -> b (h d) x y', x = h, y = w) return self.to_out(out) # model class Unet(nn.Module): def __init__( self, dim, init_dim = None, dim_mults=(1, 2, 4, 8), channels = 3, resnet_block_groups = 8, learned_sinusoidal_dim = 16 ): super().__init__() # determine dimensions self.channels = channels input_channels = channels * 2 init_dim = default(init_dim, dim) self.init_conv = nn.Conv2d(input_channels, init_dim, 7, padding = 3) dims = [init_dim, *map(lambda m: dim * m, dim_mults)] in_out = list(zip(dims[:-1], dims[1:])) block_klass = partial(ResnetBlock, groups = resnet_block_groups) # time embeddings time_dim = dim * 4 sinu_pos_emb = LearnedSinusoidalPosEmb(learned_sinusoidal_dim) fourier_dim = learned_sinusoidal_dim + 1 self.time_mlp = nn.Sequential( sinu_pos_emb, nn.Linear(fourier_dim, time_dim), nn.GELU(), nn.Linear(time_dim, time_dim) ) # layers self.downs = nn.ModuleList([]) self.ups = nn.ModuleList([]) num_resolutions = len(in_out) for ind, (dim_in, dim_out) in enumerate(in_out): is_last = ind >= (num_resolutions - 1) self.downs.append(nn.ModuleList([ block_klass(dim_in, dim_in, time_emb_dim = time_dim), block_klass(dim_in, dim_in, time_emb_dim = time_dim), Residual(PreNorm(dim_in, LinearAttention(dim_in))), Downsample(dim_in, dim_out) if not is_last else nn.Conv2d(dim_in, dim_out, 3, padding = 1) ])) mid_dim = dims[-1] self.mid_block1 = block_klass(mid_dim, mid_dim, time_emb_dim = time_dim) self.mid_attn = Residual(PreNorm(mid_dim, Attention(mid_dim))) self.mid_block2 = block_klass(mid_dim, mid_dim, time_emb_dim = time_dim) for ind, (dim_in, dim_out) in enumerate(reversed(in_out)): is_last = ind == (len(in_out) - 1) self.ups.append(nn.ModuleList([ block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim), block_klass(dim_out + dim_in, dim_out, time_emb_dim = time_dim), Residual(PreNorm(dim_out, LinearAttention(dim_out))), Upsample(dim_out, dim_in) if not is_last else nn.Conv2d(dim_out, dim_in, 3, padding = 1) ])) self.final_res_block = block_klass(dim * 2, dim, time_emb_dim = time_dim) self.final_conv = nn.Conv2d(dim, channels, 1) def forward(self, x, time, x_self_cond = None): x_self_cond = default(x_self_cond, lambda: torch.zeros_like(x)) x = torch.cat((x_self_cond, x), dim = 1) x = self.init_conv(x) r = x.clone() t = self.time_mlp(time) h = [] for block1, block2, attn, downsample in self.downs: x = block1(x, t) h.append(x) x = block2(x, t) x = attn(x) h.append(x) x = downsample(x) x = self.mid_block1(x, t) x = self.mid_attn(x) x = self.mid_block2(x, t) for block1, block2, attn, upsample in self.ups: x = torch.cat((x, h.pop()), dim = 1) x = block1(x, t) x = torch.cat((x, h.pop()), dim = 1) x = block2(x, t) x = attn(x) x = upsample(x) x = torch.cat((x, r), dim = 1) x = self.final_res_block(x, t) return self.final_conv(x) # chroma class def log(t, eps = 1e-20): return torch.log(t.clamp(min = eps)) def right_pad_dims_to(x, t): padding_dims = x.ndim - t.ndim if padding_dims <= 0: return t return t.view(*t.shape, *((1,) * padding_dims)) def beta_linear_log_snr(t): return -torch.log(expm1(1e-4 + 10 * (t ** 2))) def alpha_cosine_log_snr(t, s: float = 0.008): return -log((torch.cos((t + s) / (1 + s) * math.pi * 0.5) ** -2) - 1, eps = 1e-5) # not sure if this accounts for beta being clipped to 0.999 in discrete version def log_snr_to_alpha_sigma(log_snr): return torch.sqrt(torch.sigmoid(log_snr)), torch.sqrt(torch.sigmoid(-log_snr)) class Chroma(nn.Module): def __init__( self, model, *, image_size, timesteps = 1000, use_ddim = False, noise_schedule = 'cosine', time_difference = 0. ): super().__init__() self.model = model self.channels = self.model.channels self.image_size = image_size if noise_schedule == "linear": self.log_snr = beta_linear_log_snr elif noise_schedule == "cosine": self.log_snr = alpha_cosine_log_snr else: raise ValueError(f'invalid noise schedule {noise_schedule}') self.timesteps = timesteps self.use_ddim = use_ddim # proposed in the paper, summed to time_next # as a way to fix a deficiency in self-conditioning and lower FID when the number of sampling timesteps is < 400 self.time_difference = time_difference @property def device(self): return next(self.model.parameters()).device def get_sampling_timesteps(self, batch, *, device): times = torch.linspace(1., 0., self.timesteps + 1, device = device) times = repeat(times, 't -> b t', b = batch) times = torch.stack((times[:, :-1], times[:, 1:]), dim = 0) times = times.unbind(dim = -1) return times @torch.no_grad() def ddpm_sample(self, shape, time_difference = None): batch, device = shape[0], self.device time_difference = default(time_difference, self.time_difference) time_pairs = self.get_sampling_timesteps(batch, device = device) img = torch.randn(shape, device=device) x_start = None for time, time_next in tqdm(time_pairs, desc = 'sampling loop time step', total = self.timesteps): # add the time delay time_next = (time_next - self.time_difference).clamp(min = 0.) noise_cond = self.log_snr(time) # get predicted x0 x_start = self.model(img, noise_cond, x_start) # clip x0 x_start.clamp_(-1., 1.) # get log(snr) log_snr = self.log_snr(time) log_snr_next = self.log_snr(time_next) log_snr, log_snr_next = map(partial(right_pad_dims_to, img), (log_snr, log_snr_next)) # get alpha sigma of time and next time alpha, sigma = log_snr_to_alpha_sigma(log_snr) alpha_next, sigma_next = log_snr_to_alpha_sigma(log_snr_next) # derive posterior mean and variance c = -expm1(log_snr - log_snr_next) mean = alpha_next * (img * (1 - c) / alpha + c * x_start) variance = (sigma_next ** 2) * c log_variance = log(variance) # get noise noise = torch.where( rearrange(time_next > 0, 'b -> b 1 1 1'), torch.randn_like(img), torch.zeros_like(img) ) img = mean + (0.5 * log_variance).exp() * noise return img @torch.no_grad() def ddim_sample(self, shape, time_difference = None): batch, device = shape[0], self.device time_difference = default(time_difference, self.time_difference) time_pairs = self.get_sampling_timesteps(batch, device = device) img = torch.randn(shape, device = device) x_start = None for times, times_next in tqdm(time_pairs, desc = 'sampling loop time step'): # get times and noise levels log_snr = self.log_snr(times) log_snr_next = self.log_snr(times_next) padded_log_snr, padded_log_snr_next = map(partial(right_pad_dims_to, img), (log_snr, log_snr_next)) alpha, sigma = log_snr_to_alpha_sigma(padded_log_snr) alpha_next, sigma_next = log_snr_to_alpha_sigma(padded_log_snr_next) # add the time delay times_next = (times_next - time_difference).clamp(min = 0.) # predict x0 x_start = self.model(img, log_snr, x_start) # clip x0 x_start.clamp_(-1., 1.) # get predicted noise pred_noise = (img - alpha * x_start) / sigma.clamp(min = 1e-8) # calculate x next img = x_start * alpha_next + pred_noise * sigma_next return img @torch.no_grad() def sample(self, batch_size = 16): image_size, channels = self.image_size, self.channels sample_fn = self.ddpm_sample if not self.use_ddim else self.ddim_sample return sample_fn((batch_size, channels, image_size, image_size)) def forward(self, img, *args, **kwargs): batch, c, h, w, device, img_size, = *img.shape, img.device, self.image_size assert h == img_size and w == img_size, f'height and width of image must be {img_size}' # sample random times times = torch.zeros((batch,), device = device).float().uniform_(0, 1.) # noise sample noise = torch.randn_like(img) noise_level = self.log_snr(times) padded_noise_level = right_pad_dims_to(img, noise_level) alpha, sigma = log_snr_to_alpha_sigma(padded_noise_level) noised_img = alpha * img + sigma * noise # if doing self-conditioning, 50% of the time, predict x_start from current set of times # and condition with unet with that # this technique will slow down training by 25%, but seems to lower FID significantly self_cond = None if random() < 0.5: with torch.no_grad(): self_cond = self.model(noised_img, noise_level).detach_() # predict and take gradient step pred = self.model(noised_img, noise_level, self_cond) return F.mse_loss(pred, img) # trainer class class Trainer(object): def __init__( self, diffusion_model, folder, *, train_batch_size = 16, gradient_accumulate_every = 1, augment_horizontal_flip = True, train_lr = 1e-4, train_num_steps = 100000, ema_update_every = 10, ema_decay = 0.995, adam_betas = (0.9, 0.99), save_and_sample_every = 1000, num_samples = 25, results_folder = './results', amp = False, fp16 = False, split_batches = True, convert_image_to = None ): super().__init__() self.accelerator = Accelerator( split_batches = split_batches, mixed_precision = 'fp16' if fp16 else 'no' ) self.accelerator.native_amp = amp self.model = diffusion_model assert has_int_squareroot(num_samples), 'number of samples must have an integer square root' self.num_samples = num_samples self.save_and_sample_every = save_and_sample_every self.batch_size = train_batch_size self.gradient_accumulate_every = gradient_accumulate_every self.train_num_steps = train_num_steps self.image_size = diffusion_model.image_size # dataset and dataloader self.ds = Dataset(folder, self.image_size, augment_horizontal_flip = augment_horizontal_flip, convert_image_to = convert_image_to) dl = DataLoader(self.ds, batch_size = train_batch_size, shuffle = True, pin_memory = True, num_workers = cpu_count()) dl = self.accelerator.prepare(dl) self.dl = cycle(dl) # optimizer self.opt = Adam(diffusion_model.parameters(), lr = train_lr, betas = adam_betas) # for logging results in a folder periodically if self.accelerator.is_main_process: self.ema = EMA(diffusion_model, beta = ema_decay, update_every = ema_update_every) self.results_folder = Path(results_folder) self.results_folder.mkdir(exist_ok = True) # step counter state self.step = 0 # prepare model, dataloader, optimizer with accelerator self.model, self.opt = self.accelerator.prepare(self.model, self.opt) def save(self, milestone): if not self.accelerator.is_local_main_process: return data = { 'step': self.step, 'model': self.accelerator.get_state_dict(self.model), 'opt': self.opt.state_dict(), 'ema': self.ema.state_dict(), 'scaler': self.accelerator.scaler.state_dict() if exists(self.accelerator.scaler) else None } torch.save(data, str(self.results_folder / f'model-{milestone}.pt')) def load(self, milestone): data = torch.load(str(self.results_folder / f'model-{milestone}.pt')) model = self.accelerator.unwrap_model(self.model) model.load_state_dict(data['model']) self.step = data['step'] self.opt.load_state_dict(data['opt']) self.ema.load_state_dict(data['ema']) if exists(self.accelerator.scaler) and exists(data['scaler']): self.accelerator.scaler.load_state_dict(data['scaler']) def train(self): accelerator = self.accelerator device = accelerator.device with tqdm(initial = self.step, total = self.train_num_steps, disable = not accelerator.is_main_process) as pbar: while self.step < self.train_num_steps: total_loss = 0. for _ in range(self.gradient_accumulate_every): data = next(self.dl).to(device) with self.accelerator.autocast(): loss = self.model(data) loss = loss / self.gradient_accumulate_every total_loss += loss.item() self.accelerator.backward(loss) pbar.set_description(f'loss: {total_loss:.4f}') accelerator.wait_for_everyone() self.opt.step() self.opt.zero_grad() accelerator.wait_for_everyone() if accelerator.is_main_process: self.ema.to(device) self.ema.update() if self.step != 0 and self.step % self.save_and_sample_every == 0: self.ema.ema_model.eval() with torch.no_grad(): milestone = self.step // self.save_and_sample_every batches = num_to_groups(self.num_samples, self.batch_size) all_images_list = list(map(lambda n: self.ema.ema_model.sample(batch_size=n), batches)) all_images = torch.cat(all_images_list, dim = 0) utils.save_image(all_images, str(self.results_folder / f'sample-{milestone}.png'), nrow = int(math.sqrt(self.num_samples))) self.save(milestone) self.step += 1 pbar.update(1) accelerator.print('training complete')
terminate = False def signal_handling(signum,frame): global terminate terminate = True num_attempts = 4 for attempt in range(num_attempts): dream = Imagine( text = "an armchair in the form of pikachu\\an armchair imitating pikachu\\abstract", text_min = "blur\\zoom", lr = 7e-2, image_size = 512, gradient_accumulate_every = 1, save_every = 50, epochs = 5, iterations = 50, save_progress = False, bilinear = False, open_folder = False, seed = None, torch_deterministic = False, max_classes = 20, class_temperature = 2., save_date_time = False, save_best = True, experimental_resample = True, ema_decay = 0.99 ) dream() shutil.copy(dream.textpath + ".best.png", f"{attempt}.png") try: time.sleep(2) del dream time.sleep(2) torch.cuda.empty_cache() except Exception: torch.cuda.empty_cache()
__version__ = '0.9.1'
"""Good differentiable image resampling for PyTorch.""" def sinc(x): return torch.where(x != 0, torch.sin(math.pi * x) / (math.pi * x), x.new_ones([])) def lanczos(x, a): cond = torch.logical_and(-a < x, x < a) out = torch.where(cond, sinc(x) * sinc(x/a), x.new_zeros([])) return out / out.sum() def ramp(ratio, width): n = math.ceil(width / ratio + 1) out = torch.empty([n]) cur = 0 for i in range(out.shape[0]): out[i] = cur cur += ratio return torch.cat([-out[1:].flip([0]), out])[1:-1] def odd(fn): return update_wrapper(lambda x: torch.sign(x) * fn(abs(x)), fn) def _to_linear_srgb(input): cond = input <= 0.04045 a = input / 12.92 b = ((input + 0.055) / 1.055)**2.4 return torch.where(cond, a, b) def _to_nonlinear_srgb(input): cond = input <= 0.0031308 a = 12.92 * input b = 1.055 * input**(1/2.4) - 0.055 return torch.where(cond, a, b) to_linear_srgb = odd(_to_linear_srgb) to_nonlinear_srgb = odd(_to_nonlinear_srgb) def resample(input, size, align_corners=True, is_srgb=False): n, c, h, w = input.shape dh, dw = size if is_srgb: input = to_linear_srgb(input) input = input.view([n * c, 1, h, w]) if dh < h: kernel_h = lanczos(ramp(dh / h, 3), 3).to(input.device, input.dtype) pad_h = (kernel_h.shape[0] - 1) // 2 input = F.pad(input, (0, 0, pad_h, pad_h), 'reflect') input = F.conv2d(input, kernel_h[None, None, :, None]) if dw < w: kernel_w = lanczos(ramp(dw / w, 3), 3).to(input.device, input.dtype) pad_w = (kernel_w.shape[0] - 1) // 2 input = F.pad(input, (pad_w, pad_w, 0, 0), 'reflect') input = F.conv2d(input, kernel_w[None, None, None, :]) input = input.view([n, c, h, w]) input = F.interpolate(input, size, mode='bicubic', align_corners=align_corners) if is_srgb: input = to_nonlinear_srgb(input) return input
# Exponential Moving Average (from https://gist.github.com/crowsonkb/76b94d5238272722290734bf4725d204) """Exponential moving average for PyTorch. Adapted from https://www.zijianhu.com/post/pytorch/ema/ by crowsonkb """ class EMA(nn.Module): def __init__(self, model, decay): super().__init__() self.model = model self.decay = decay self.register_buffer('accum', torch.tensor(1.)) self._biased = deepcopy(self.model) self.average = deepcopy(self.model) for param in self._biased.parameters(): param.detach_().zero_() for param in self.average.parameters(): param.detach_().zero_() self.update() @torch.no_grad() def update(self): assert self.training, 'Update should only be called during training' self.accum *= self.decay model_params = dict(self.model.named_parameters()) biased_params = dict(self._biased.named_parameters()) average_params = dict(self.average.named_parameters()) assert model_params.keys() == biased_params.keys() == average_params.keys(), f'Model parameter keys incompatible with EMA stored parameter keys' for name, param in model_params.items(): biased_params[name].mul_(self.decay) biased_params[name].add_((1 - self.decay) * param) average_params[name].copy_(biased_params[name]) average_params[name].div_(1 - self.accum) model_buffers = dict(self.model.named_buffers()) biased_buffers = dict(self._biased.named_buffers()) average_buffers = dict(self.average.named_buffers()) assert model_buffers.keys() == biased_buffers.keys() == average_buffers.keys() for name, buffer in model_buffers.items(): biased_buffers[name].copy_(buffer) average_buffers[name].copy_(buffer) def forward(self, *args, **kwargs): if self.training: return self.model(*args, **kwargs) return self.average(*args, **kwargs)
# this code is a copy from huggingface # with some minor modifications try: from urllib.parse import urlparse except ImportError: from urlparse import urlparse try: from pathlib import Path PYTORCH_PRETRAINED_BIGGAN_CACHE = Path(os.getenv('PYTORCH_PRETRAINED_BIGGAN_CACHE', Path.home() / '.pytorch_pretrained_biggan')) except (AttributeError, ImportError): PYTORCH_PRETRAINED_BIGGAN_CACHE = os.getenv('PYTORCH_PRETRAINED_BIGGAN_CACHE', os.path.join(os.path.expanduser("~"), '.pytorch_pretrained_biggan')) logger = logging.getLogger(__name__) # pylint: disable=invalid-name PRETRAINED_MODEL_ARCHIVE_MAP = { 'biggan-deep-128': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-128-pytorch_model.bin", 'biggan-deep-256': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-256-pytorch_model.bin", 'biggan-deep-512': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-512-pytorch_model.bin", } PRETRAINED_CONFIG_ARCHIVE_MAP = { 'biggan-deep-128': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-128-config.json", 'biggan-deep-256': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-256-config.json", 'biggan-deep-512': "https://s3.amazonaws.com/models.huggingface.co/biggan/biggan-deep-512-config.json", } WEIGHTS_NAME = 'pytorch_model.bin' CONFIG_NAME = 'config.json' def url_to_filename(url, etag=None): """ Convert `url` into a hashed filename in a repeatable way. If `etag` is specified, append its hash to the url's, delimited by a period. """ url_bytes = url.encode('utf-8') url_hash = sha256(url_bytes) filename = url_hash.hexdigest() if etag: etag_bytes = etag.encode('utf-8') etag_hash = sha256(etag_bytes) filename += '.' + etag_hash.hexdigest() return filename def filename_to_url(filename, cache_dir=None): """ Return the url and etag (which may be ``None``) stored for `filename`. Raise ``EnvironmentError`` if `filename` or its stored metadata do not exist. """ if cache_dir is None: cache_dir = PYTORCH_PRETRAINED_BIGGAN_CACHE if sys.version_info[0] == 3 and isinstance(cache_dir, Path): cache_dir = str(cache_dir) cache_path = os.path.join(cache_dir, filename) if not os.path.exists(cache_path): raise EnvironmentError("file {} not found".format(cache_path)) meta_path = cache_path + '.json' if not os.path.exists(meta_path): raise EnvironmentError("file {} not found".format(meta_path)) with open(meta_path, encoding="utf-8") as meta_file: metadata = json.load(meta_file) url = metadata['url'] etag = metadata['etag'] return url, etag def cached_path(url_or_filename, cache_dir=None): """ Given something that might be a URL (or might be a local path), determine which. If it's a URL, download the file and cache it, and return the path to the cached file. If it's already a local path, make sure the file exists and then return the path. """ if cache_dir is None: cache_dir = PYTORCH_PRETRAINED_BIGGAN_CACHE if sys.version_info[0] == 3 and isinstance(url_or_filename, Path): url_or_filename = str(url_or_filename) if sys.version_info[0] == 3 and isinstance(cache_dir, Path): cache_dir = str(cache_dir) parsed = urlparse(url_or_filename) if parsed.scheme in ('http', 'https', 's3'): # URL, so get it from the cache (downloading if necessary) return get_from_cache(url_or_filename, cache_dir) elif os.path.exists(url_or_filename): # File, and it exists. return url_or_filename elif parsed.scheme == '': # File, but it doesn't exist. raise EnvironmentError("file {} not found".format(url_or_filename)) else: # Something unknown raise ValueError("unable to parse {} as a URL or as a local path".format(url_or_filename)) def split_s3_path(url): """Split a full s3 path into the bucket name and path.""" parsed = urlparse(url) if not parsed.netloc or not parsed.path: raise ValueError("bad s3 path {}".format(url)) bucket_name = parsed.netloc s3_path = parsed.path # Remove '/' at beginning of path. if s3_path.startswith("/"): s3_path = s3_path[1:] return bucket_name, s3_path def s3_request(func): """ Wrapper function for s3 requests in order to create more helpful error messages. """ @wraps(func) def wrapper(url, *args, **kwargs): try: return func(url, *args, **kwargs) except ClientError as exc: if int(exc.response["Error"]["Code"]) == 404: raise EnvironmentError("file {} not found".format(url)) else: raise return wrapper @s3_request def s3_etag(url): """Check ETag on S3 object.""" s3_resource = boto3.resource("s3") bucket_name, s3_path = split_s3_path(url) s3_object = s3_resource.Object(bucket_name, s3_path) return s3_object.e_tag @s3_request def s3_get(url, temp_file): """Pull a file directly from S3.""" s3_resource = boto3.resource("s3") bucket_name, s3_path = split_s3_path(url) s3_resource.Bucket(bucket_name).download_fileobj(s3_path, temp_file) def http_get(url, temp_file): req = requests.get(url, stream=True) content_length = req.headers.get('Content-Length') total = int(content_length) if content_length is not None else None progress = tqdm(unit="B", total=total) for chunk in req.iter_content(chunk_size=1024): if chunk: # filter out keep-alive new chunks progress.update(len(chunk)) temp_file.write(chunk) progress.close() def get_from_cache(url, cache_dir=None): """ Given a URL, look for the corresponding dataset in the local cache. If it's not there, download it. Then return the path to the cached file. """ if cache_dir is None: cache_dir = PYTORCH_PRETRAINED_BIGGAN_CACHE if sys.version_info[0] == 3 and isinstance(cache_dir, Path): cache_dir = str(cache_dir) if not os.path.exists(cache_dir): os.makedirs(cache_dir) # Get eTag to add to filename, if it exists. if url.startswith("s3://"): etag = s3_etag(url) else: response = requests.head(url, allow_redirects=True) if response.status_code != 200: raise IOError("HEAD request failed for url {} with status code {}" .format(url, response.status_code)) etag = response.headers.get("ETag") filename = url_to_filename(url, etag) # get cache path to put the file cache_path = os.path.join(cache_dir, filename) if not os.path.exists(cache_path): # Download to temporary file, then copy to cache dir once finished. # Otherwise you get corrupt cache entries if the download gets interrupted. with tempfile.NamedTemporaryFile() as temp_file: logger.info("%s not found in cache, downloading to %s", url, temp_file.name) # GET file object if url.startswith("s3://"): s3_get(url, temp_file) else: http_get(url, temp_file) # we are copying the file before closing it, so flush to avoid truncation temp_file.flush() # shutil.copyfileobj() starts at the current position, so go to the start temp_file.seek(0) logger.info("copying %s to cache at %s", temp_file.name, cache_path) with open(cache_path, 'wb') as cache_file: shutil.copyfileobj(temp_file, cache_file) logger.info("creating metadata file for %s", cache_path) meta = {'url': url, 'etag': etag} meta_path = cache_path + '.json' with open(meta_path, 'w', encoding="utf-8") as meta_file: json.dump(meta, meta_file) logger.info("removing temp file %s", temp_file.name) return cache_path def read_set_from_file(filename): ''' Extract a de-duped collection (set) of text from a file. Expected file format is one item per line. ''' collection = set() with open(filename, 'r', encoding='utf-8') as file_: for line in file_: collection.add(line.rstrip()) return collection def get_file_extension(path, dot=True, lower=True): ext = os.path.splitext(path)[1] ext = ext if dot else ext[1:] return ext.lower() if lower else ext class BigGANConfig(object): """ Configuration class to store the configuration of a `BigGAN`. Defaults are for the 128x128 model. layers tuple are (up-sample in the layer ?, input channels, output channels) """ def __init__(self, output_dim=128, z_dim=128, class_embed_dim=128, channel_width=128, num_classes=1000, layers=[(False, 16, 16), (True, 16, 16), (False, 16, 16), (True, 16, 8), (False, 8, 8), (True, 8, 4), (False, 4, 4), (True, 4, 2), (False, 2, 2), (True, 2, 1)], attention_layer_position=8, eps=1e-4, n_stats=51): """Constructs BigGANConfig. """ self.output_dim = output_dim self.z_dim = z_dim self.class_embed_dim = class_embed_dim self.channel_width = channel_width self.num_classes = num_classes self.layers = layers self.attention_layer_position = attention_layer_position self.eps = eps self.n_stats = n_stats @classmethod def from_dict(cls, json_object): """Constructs a `BigGANConfig` from a Python dictionary of parameters.""" config = BigGANConfig() for key, value in json_object.items(): config.__dict__[key] = value return config @classmethod def from_json_file(cls, json_file): """Constructs a `BigGANConfig` from a json file of parameters.""" with open(json_file, "r", encoding='utf-8') as reader: text = reader.read() return cls.from_dict(json.loads(text)) def __repr__(self): return str(self.to_json_string()) def to_dict(self): """Serializes this instance to a Python dictionary.""" output = copy.deepcopy(self.__dict__) return output def to_json_string(self): """Serializes this instance to a JSON string.""" return json.dumps(self.to_dict(), indent=2, sort_keys=True) + "\n" def snconv2d(eps=1e-12, **kwargs): return nn.utils.spectral_norm(nn.Conv2d(**kwargs), eps=eps) def snlinear(eps=1e-12, **kwargs): return nn.utils.spectral_norm(nn.Linear(**kwargs), eps=eps) def sn_embedding(eps=1e-12, **kwargs): return nn.utils.spectral_norm(nn.Embedding(**kwargs), eps=eps) class SelfAttn(nn.Module): """ Self attention Layer""" def __init__(self, in_channels, eps=1e-12): super(SelfAttn, self).__init__() self.in_channels = in_channels self.snconv1x1_theta = snconv2d(in_channels=in_channels, out_channels=in_channels//8, kernel_size=1, bias=False, eps=eps) self.snconv1x1_phi = snconv2d(in_channels=in_channels, out_channels=in_channels//8, kernel_size=1, bias=False, eps=eps) self.snconv1x1_g = snconv2d(in_channels=in_channels, out_channels=in_channels//2, kernel_size=1, bias=False, eps=eps) self.snconv1x1_o_conv = snconv2d(in_channels=in_channels//2, out_channels=in_channels, kernel_size=1, bias=False, eps=eps) self.maxpool = nn.MaxPool2d(2, stride=2, padding=0) self.softmax = nn.Softmax(dim=-1) self.gamma = nn.Parameter(torch.zeros(1)) def forward(self, x): _, ch, h, w = x.size() # Theta path theta = self.snconv1x1_theta(x) theta = theta.view(-1, ch//8, h*w) # Phi path phi = self.snconv1x1_phi(x) phi = self.maxpool(phi) phi = phi.view(-1, ch//8, h*w//4) # Attn map attn = torch.bmm(theta.permute(0, 2, 1), phi) attn = self.softmax(attn) # g path g = self.snconv1x1_g(x) g = self.maxpool(g) g = g.view(-1, ch//2, h*w//4) # Attn_g - o_conv attn_g = torch.bmm(g, attn.permute(0, 2, 1)) attn_g = attn_g.view(-1, ch//2, h, w) attn_g = self.snconv1x1_o_conv(attn_g) # Out out = x + self.gamma*attn_g return out class BigGANBatchNorm(nn.Module): """ This is a batch norm module that can handle conditional input and can be provided with pre-computed activation means and variances for various truncation parameters. We cannot just rely on torch.batch_norm since it cannot handle batched weights (pytorch 1.0.1). We computate batch_norm our-self without updating running means and variances. If you want to train this model you should add running means and variance computation logic. """ def __init__(self, num_features, condition_vector_dim=None, n_stats=51, eps=1e-4, conditional=True): super(BigGANBatchNorm, self).__init__() self.num_features = num_features self.eps = eps self.conditional = conditional # We use pre-computed statistics for n_stats values of truncation between 0 and 1 self.register_buffer('running_means', torch.zeros(n_stats, num_features)) self.register_buffer('running_vars', torch.ones(n_stats, num_features)) self.step_size = 1.0 / (n_stats - 1) if conditional: assert condition_vector_dim is not None self.scale = snlinear(in_features=condition_vector_dim, out_features=num_features, bias=False, eps=eps) self.offset = snlinear(in_features=condition_vector_dim, out_features=num_features, bias=False, eps=eps) else: self.weight = torch.nn.Parameter(torch.Tensor(num_features)) self.bias = torch.nn.Parameter(torch.Tensor(num_features)) def forward(self, x, truncation, condition_vector=None): # Retreive pre-computed statistics associated to this truncation coef, start_idx = math.modf(truncation / self.step_size) start_idx = int(start_idx) if coef != 0.0: # Interpolate running_mean = self.running_means[start_idx] * coef + self.running_means[start_idx + 1] * (1 - coef) running_var = self.running_vars[start_idx] * coef + self.running_vars[start_idx + 1] * (1 - coef) else: running_mean = self.running_means[start_idx] running_var = self.running_vars[start_idx] if self.conditional: running_mean = running_mean.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) running_var = running_var.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) weight = 1 + self.scale(condition_vector).unsqueeze(-1).unsqueeze(-1) bias = self.offset(condition_vector).unsqueeze(-1).unsqueeze(-1) out = (x - running_mean) / torch.sqrt(running_var + self.eps) * weight + bias else: out = F.batch_norm(x, running_mean, running_var, self.weight, self.bias, training=False, momentum=0.0, eps=self.eps) return out class GenBlock(nn.Module): def __init__(self, in_size, out_size, condition_vector_dim, reduction_factor=4, up_sample=False, n_stats=51, eps=1e-12): super(GenBlock, self).__init__() self.up_sample = up_sample self.drop_channels = (in_size != out_size) middle_size = in_size // reduction_factor self.bn_0 = BigGANBatchNorm(in_size, condition_vector_dim, n_stats=n_stats, eps=eps, conditional=True) self.conv_0 = snconv2d(in_channels=in_size, out_channels=middle_size, kernel_size=1, eps=eps) self.bn_1 = BigGANBatchNorm(middle_size, condition_vector_dim, n_stats=n_stats, eps=eps, conditional=True) self.conv_1 = snconv2d(in_channels=middle_size, out_channels=middle_size, kernel_size=3, padding=1, eps=eps) self.bn_2 = BigGANBatchNorm(middle_size, condition_vector_dim, n_stats=n_stats, eps=eps, conditional=True) self.conv_2 = snconv2d(in_channels=middle_size, out_channels=middle_size, kernel_size=3, padding=1, eps=eps) self.bn_3 = BigGANBatchNorm(middle_size, condition_vector_dim, n_stats=n_stats, eps=eps, conditional=True) self.conv_3 = snconv2d(in_channels=middle_size, out_channels=out_size, kernel_size=1, eps=eps) self.relu = nn.ReLU() def forward(self, x, cond_vector, truncation): x0 = x x = self.bn_0(x, truncation, cond_vector) x = self.relu(x) x = self.conv_0(x) x = self.bn_1(x, truncation, cond_vector) x = self.relu(x) if self.up_sample: x = F.interpolate(x, scale_factor=2, mode='nearest') x = self.conv_1(x) x = self.bn_2(x, truncation, cond_vector) x = self.relu(x) x = self.conv_2(x) x = self.bn_3(x, truncation, cond_vector) x = self.relu(x) x = self.conv_3(x) if self.drop_channels: new_channels = x0.shape[1] // 2 x0 = x0[:, :new_channels, ...] if self.up_sample: x0 = F.interpolate(x0, scale_factor=2, mode='nearest') out = x + x0 return out class Generator(nn.Module): def __init__(self, config): super(Generator, self).__init__() self.config = config ch = config.channel_width condition_vector_dim = config.z_dim * 2 self.gen_z = snlinear(in_features=condition_vector_dim, out_features=4 * 4 * 16 * ch, eps=config.eps) layers = [] for i, layer in enumerate(config.layers): if i == config.attention_layer_position: layers.append(SelfAttn(ch*layer[1], eps=config.eps)) layers.append(GenBlock(ch*layer[1], ch*layer[2], condition_vector_dim, up_sample=layer[0], n_stats=config.n_stats, eps=config.eps)) self.layers = nn.ModuleList(layers) self.bn = BigGANBatchNorm(ch, n_stats=config.n_stats, eps=config.eps, conditional=False) self.relu = nn.ReLU() self.conv_to_rgb = snconv2d(in_channels=ch, out_channels=ch, kernel_size=3, padding=1, eps=config.eps) self.tanh = nn.Tanh() def forward(self, cond_vector, truncation): z = self.gen_z(cond_vector[0].unsqueeze(0)) # We use this conversion step to be able to use TF weights: # TF convention on shape is [batch, height, width, channels] # PT convention on shape is [batch, channels, height, width] z = z.view(-1, 4, 4, 16 * self.config.channel_width) z = z.permute(0, 3, 1, 2).contiguous() next_available_latent_index = 1 for layer in self.layers: if isinstance(layer, GenBlock): z = layer(z, cond_vector[next_available_latent_index].unsqueeze(0), truncation) next_available_latent_index += 1 else: z = layer(z) z = self.bn(z, truncation) z = self.relu(z) z = self.conv_to_rgb(z) z = z[:, :3, ...] z = self.tanh(z) return z class BigGAN(nn.Module): """BigGAN Generator.""" @classmethod def from_pretrained(cls, pretrained_model_name_or_path, cache_dir=None, *inputs, **kwargs): if pretrained_model_name_or_path in PRETRAINED_MODEL_ARCHIVE_MAP: model_file = PRETRAINED_MODEL_ARCHIVE_MAP[pretrained_model_name_or_path] config_file = PRETRAINED_CONFIG_ARCHIVE_MAP[pretrained_model_name_or_path] else: model_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_NAME) config_file = os.path.join(pretrained_model_name_or_path, CONFIG_NAME) try: resolved_model_file = cached_path(model_file, cache_dir=cache_dir) resolved_config_file = cached_path(config_file, cache_dir=cache_dir) except EnvironmentError: logger.error("Wrong model name, should be a valid path to a folder containing " "a {} file and a {} file or a model name in {}".format( WEIGHTS_NAME, CONFIG_NAME, PRETRAINED_MODEL_ARCHIVE_MAP.keys())) raise logger.info("loading model {} from cache at {}".format(pretrained_model_name_or_path, resolved_model_file)) # Load config config = BigGANConfig.from_json_file(resolved_config_file) logger.info("Model config {}".format(config)) # Instantiate model. model = cls(config, *inputs, **kwargs) state_dict = torch.load(resolved_model_file, map_location='cpu' if not torch.cuda.is_available() else None) model.load_state_dict(state_dict, strict=False) return model def __init__(self, config): super(BigGAN, self).__init__() self.config = config self.embeddings = nn.Linear(config.num_classes, config.z_dim, bias=False) self.generator = Generator(config) def forward(self, z, class_label, truncation): assert 0 < truncation <= 1 embed = self.embeddings(class_label) cond_vector = torch.cat((z, embed), dim=1) z = self.generator(cond_vector, truncation) return z
def train( text=None, img=None, text_min="", lr = .07, image_size = 512, gradient_accumulate_every = 1, epochs = 20, iterations = 1050, save_every = 50, overwrite = False, save_progress = False, save_date_time = False, bilinear = False, open_folder = True, seed = 0, append_seed = False, random = False, torch_deterministic = False, max_classes = None, class_temperature = 2., save_best = False, experimental_resample = False, ema_decay = 0.5, num_cutouts = 128, center_bias = False, larger_model = False ): print(f'Starting up... v{__version__}') if random: seed = rnd.randint(0, 1e6) imagine = Imagine( text=text, img=img, text_min=text_min, lr = lr, image_size = image_size, gradient_accumulate_every = gradient_accumulate_every, epochs = epochs, iterations = iterations, save_every = save_every, save_progress = save_progress, bilinear = bilinear, seed = seed, append_seed = append_seed, torch_deterministic = torch_deterministic, open_folder = open_folder, max_classes = max_classes, class_temperature = class_temperature, save_date_time = save_date_time, save_best = save_best, experimental_resample = experimental_resample, ema_decay = ema_decay, num_cutouts = num_cutouts, center_bias = center_bias, larger_clip = larger_model ) if not overwrite and imagine.filename.exists(): answer = input('Imagined image already exists, do you want to overwrite? (y/n) ').lower() if answer not in ('yes', 'y'): exit() imagine() def main(): fire.Fire(train)
assert torch.cuda.is_available(), 'CUDA must be available in order to use Big Sleep' # graceful keyboard interrupt terminate = False def signal_handling(signum,frame): print('detecting keyboard interrupt, gracefully exiting') global terminate terminate = True signal.signal(signal.SIGINT,signal_handling) # helpers def exists(val): return val is not None def open_folder(path): if os.path.isfile(path): path = os.path.dirname(path) if not os.path.isdir(path): return cmd_list = None if sys.platform == 'darwin': cmd_list = ['open', '--', path] elif sys.platform == 'linux2' or sys.platform == 'linux': cmd_list = ['xdg-open', path] elif sys.platform in ['win32', 'win64']: cmd_list = ['explorer', path.replace('/','\\')] if cmd_list == None: return try: subprocess.check_call(cmd_list) except subprocess.CalledProcessError: pass except OSError: pass def create_text_path(text=None, img=None, encoding=None): input_name = "" if text is not None: input_name += text if img is not None: if isinstance(img, str): img_name = "".join(img.split(".")[:-1]) # replace spaces by underscores, remove img extension img_name = img_name.split("/")[-1] # only take img name, not path else: img_name = "PIL_img" input_name += "_" + img_name if encoding is not None: input_name = "your_encoding" return input_name.replace("-", "_").replace(",", "").replace(" ", "_").replace("|", "--").strip('-_')[:255] # tensor helpers def differentiable_topk(x, k, temperature=1.): n, dim = x.shape topk_tensors = [] for i in range(k): is_last = i == (k - 1) values, indices = (x / temperature).softmax(dim=-1).topk(1, dim=-1) topks = torch.zeros_like(x).scatter_(-1, indices, values) topk_tensors.append(topks) if not is_last: x = x.scatter(-1, indices, float('-inf')) topks = torch.cat(topk_tensors, dim=-1) return topks.reshape(n, k, dim).sum(dim = 1) def create_clip_img_transform(image_width): clip_mean = [0.48145466, 0.4578275, 0.40821073] clip_std = [0.26862954, 0.26130258, 0.27577711] transform = T.Compose([ #T.ToPILImage(), T.Resize(image_width), T.CenterCrop((image_width, image_width)), T.ToTensor(), T.Normalize(mean=clip_mean, std=clip_std) ]) return transform def rand_cutout(image, size, center_bias=False, center_focus=2): width = image.shape[-1] min_offset = 0 max_offset = width - size if center_bias: # sample around image center center = max_offset / 2 std = center / center_focus offset_x = int(random.gauss(mu=center, sigma=std)) offset_y = int(random.gauss(mu=center, sigma=std)) # resample uniformly if over boundaries offset_x = random.randint(min_offset, max_offset) if (offset_x > max_offset or offset_x < min_offset) else offset_x offset_y = random.randint(min_offset, max_offset) if (offset_y > max_offset or offset_y < min_offset) else offset_y else: offset_x = random.randint(min_offset, max_offset) offset_y = random.randint(min_offset, max_offset) cutout = image[:, :, offset_x:offset_x + size, offset_y:offset_y + size] return cutout # load biggan class Latents(torch.nn.Module): def __init__( self, num_latents = 15, num_classes = 1000, z_dim = 128, max_classes = None, class_temperature = 2. ): super().__init__() self.normu = torch.nn.Parameter(torch.zeros(num_latents, z_dim).normal_(std = 1)) self.cls = torch.nn.Parameter(torch.zeros(num_latents, num_classes).normal_(mean = -3.9, std = .3)) self.register_buffer('thresh_lat', torch.tensor(1)) assert not exists(max_classes) or max_classes > 0 and max_classes <= num_classes, f'max_classes must be between 0 and {num_classes}' self.max_classes = max_classes self.class_temperature = class_temperature def forward(self): if exists(self.max_classes): classes = differentiable_topk(self.cls, self.max_classes, temperature = self.class_temperature) else: classes = torch.sigmoid(self.cls) return self.normu, classes class Model(nn.Module): def __init__( self, image_size, max_classes = None, class_temperature = 2., ema_decay = 0.99 ): super().__init__() assert image_size in (128, 256, 512), 'image size must be one of 128, 256, or 512' self.biggan = BigGAN.from_pretrained(f'biggan-deep-{image_size}') self.max_classes = max_classes self.class_temperature = class_temperature self.ema_decay\ = ema_decay self.init_latents() def init_latents(self): latents = Latents( num_latents = len(self.biggan.config.layers) + 1, num_classes = self.biggan.config.num_classes, z_dim = self.biggan.config.z_dim, max_classes = self.max_classes, class_temperature = self.class_temperature ) self.latents = EMA(latents, self.ema_decay) def forward(self): self.biggan.eval() out = self.biggan(*self.latents(), 1) return (out + 1) / 2 class BigSleep(nn.Module): def __init__( self, num_cutouts = 128, loss_coef = 100, image_size = 512, bilinear = False, max_classes = None, class_temperature = 2., experimental_resample = False, ema_decay = 0.99, center_bias = False, larger_clip = False ): super().__init__() self.loss_coef = loss_coef self.image_size = image_size self.num_cutouts = num_cutouts self.experimental_resample = experimental_resample self.center_bias = center_bias self.interpolation_settings = {'mode': 'bilinear', 'align_corners': False} if bilinear else {'mode': 'nearest'} model_name = 'ViT-B/32' if not larger_clip else 'ViT-L/14' self.perceptor, self.normalize_image = load(model_name, jit = False) self.model = Model( image_size = image_size, max_classes = max_classes, class_temperature = class_temperature, ema_decay = ema_decay ) def reset(self): self.model.init_latents() def sim_txt_to_img(self, text_embed, img_embed, text_type="max"): sign = -1 if text_type == "min": sign = 1 return sign * self.loss_coef * torch.cosine_similarity(text_embed, img_embed, dim = -1).mean() def forward(self, text_embeds, text_min_embeds=[], return_loss = True): width, num_cutouts = self.image_size, self.num_cutouts out = self.model() if not return_loss: return out pieces = [] for ch in range(num_cutouts): # sample cutout size size = int(width * torch.zeros(1,).normal_(mean=.8, std=.3).clip(.5, .95)) # get cutout apper = rand_cutout(out, size, center_bias=self.center_bias) if (self.experimental_resample): apper = resample(apper, (224, 224)) else: apper = F.interpolate(apper, (224, 224), **self.interpolation_settings) pieces.append(apper) into = torch.cat(pieces) into = self.normalize_image(into) image_embed = self.perceptor.encode_image(into) latents, soft_one_hot_classes = self.model.latents() num_latents = latents.shape[0] latent_thres = self.model.latents.model.thresh_lat lat_loss = torch.abs(1 - torch.std(latents, dim=1)).mean() + \ torch.abs(torch.mean(latents, dim = 1)).mean() + \ 4 * torch.max(torch.square(latents).mean(), latent_thres) for array in latents: mean = torch.mean(array) diffs = array - mean var = torch.mean(torch.pow(diffs, 2.0)) std = torch.pow(var, 0.5) zscores = diffs / std skews = torch.mean(torch.pow(zscores, 3.0)) kurtoses = torch.mean(torch.pow(zscores, 4.0)) - 3.0 lat_loss = lat_loss + torch.abs(kurtoses) / num_latents + torch.abs(skews) / num_latents cls_loss = ((50 * torch.topk(soft_one_hot_classes, largest = False, dim = 1, k = 999)[0]) ** 2).mean() results = [] for txt_embed in text_embeds: results.append(self.sim_txt_to_img(txt_embed, image_embed)) for txt_min_embed in text_min_embeds: results.append(self.sim_txt_to_img(txt_min_embed, image_embed, "min")) sim_loss = sum(results).mean() return out, (lat_loss, cls_loss, sim_loss) class Imagine(nn.Module): def __init__( self, *, text=None, img=None, encoding=None, text_min = "", lr = .07, image_size = 512, gradient_accumulate_every = 1, save_every = 50, epochs = 20, iterations = 1050, save_progress = False, bilinear = False, open_folder = True, seed = None, append_seed = False, torch_deterministic = False, max_classes = None, class_temperature = 2., save_date_time = False, save_best = False, experimental_resample = False, ema_decay = 0.99, num_cutouts = 128, center_bias = False, larger_clip = False ): super().__init__() if torch_deterministic: assert not bilinear, 'the deterministic (seeded) operation does not work with interpolation (PyTorch 1.7.1)' torch.set_deterministic(True) self.seed = seed self.append_seed = append_seed if exists(seed): print(f'setting seed of {seed}') if seed == 0: print('you can override this with --seed argument in the command line, or --random for a randomly chosen one') torch.manual_seed(seed) self.epochs = epochs self.iterations = iterations model = BigSleep( image_size = image_size, bilinear = bilinear, max_classes = max_classes, class_temperature = class_temperature, experimental_resample = experimental_resample, ema_decay = ema_decay, num_cutouts = num_cutouts, center_bias = center_bias, larger_clip = larger_clip ).cuda() self.model = model self.lr = lr self.optimizer = Adam(model.model.latents.model.parameters(), lr) self.gradient_accumulate_every = gradient_accumulate_every self.save_every = save_every self.save_progress = save_progress self.save_date_time = save_date_time self.save_best = save_best self.current_best_score = 0 self.open_folder = open_folder self.total_image_updates = (self.epochs * self.iterations) / self.save_every self.encoded_texts = { "max": [], "min": [] } # create img transform self.clip_transform = create_clip_img_transform(224) # create starting encoding self.set_clip_encoding(text=text, img=img, encoding=encoding, text_min=text_min) @property def seed_suffix(self): return f'.{self.seed}' if self.append_seed and exists(self.seed) else '' def set_text(self, text): self.set_clip_encoding(text = text) def create_clip_encoding(self, text=None, img=None, encoding=None): self.text = text self.img = img if encoding is not None: encoding = encoding.cuda() #elif self.create_story: # encoding = self.update_story_encoding(epoch=0, iteration=1) elif text is not None and img is not None: encoding = (self.create_text_encoding(text) + self.create_img_encoding(img)) / 2 elif text is not None: encoding = self.create_text_encoding(text) elif img is not None: encoding = self.create_img_encoding(img) return encoding def create_text_encoding(self, text): tokenized_text = tokenize(text).cuda() with torch.no_grad(): text_encoding = self.model.perceptor.encode_text(tokenized_text).detach() return text_encoding def create_img_encoding(self, img): if isinstance(img, str): img = Image.open(img) normed_img = self.clip_transform(img).unsqueeze(0).cuda() with torch.no_grad(): img_encoding = self.model.perceptor.encode_image(normed_img).detach() return img_encoding def encode_multiple_phrases(self, text, img=None, encoding=None, text_type="max"): if text is not None and "|" in text: self.encoded_texts[text_type] = [self.create_clip_encoding(text=prompt_min, img=img, encoding=encoding) for prompt_min in text.split("|")] else: self.encoded_texts[text_type] = [self.create_clip_encoding(text=text, img=img, encoding=encoding)] def encode_max_and_min(self, text, img=None, encoding=None, text_min=""): self.encode_multiple_phrases(text, img=img, encoding=encoding) if text_min is not None and text_min != "": self.encode_multiple_phrases(text_min, img=img, encoding=encoding, text_type="min") def set_clip_encoding(self, text=None, img=None, encoding=None, text_min=""): self.current_best_score = 0 self.text = text self.text_min = text_min if len(text_min) > 0: text = text + "_wout_" + text_min[:255] if text is not None else "wout_" + text_min[:255] text_path = create_text_path(text=text, img=img, encoding=encoding) if self.save_date_time: text_path = datetime.now().strftime("%y%m%d-%H%M%S-") + text_path self.text_path = text_path self.filename = Path(f'./{text_path}{self.seed_suffix}.png') self.encode_max_and_min(text, img=img, encoding=encoding, text_min=text_min) # Tokenize and encode each prompt def reset(self): self.model.reset() self.model = self.model.cuda() self.optimizer = Adam(self.model.model.latents.parameters(), self.lr) def train_step(self, epoch, i, pbar=None): total_loss = 0 for _ in range(self.gradient_accumulate_every): out, losses = self.model(self.encoded_texts["max"], self.encoded_texts["min"]) loss = sum(losses) / self.gradient_accumulate_every total_loss += loss loss.backward() self.optimizer.step() self.model.model.latents.update() self.optimizer.zero_grad() if (i + 1) % self.save_every == 0: with torch.no_grad(): self.model.model.latents.eval() out, losses = self.model(self.encoded_texts["max"], self.encoded_texts["min"]) top_score, best = torch.topk(losses[2], k=1, largest=False) image = self.model.model()[best].cpu() self.model.model.latents.train() save_image(image, str(self.filename)) if pbar is not None: pbar.update(1) else: print(f'image updated at "./{str(self.filename)}"') if self.save_progress: total_iterations = epoch * self.iterations + i num = total_iterations // self.save_every save_image(image, Path(f'./{self.text_path}.{num}{self.seed_suffix}.png')) if self.save_best and top_score.item() < self.current_best_score: self.current_best_score = top_score.item() save_image(image, Path(f'./{self.text_path}{self.seed_suffix}.best.png')) return out, total_loss def forward(self): penalizing = "" if len(self.text_min) > 0: penalizing = f'penalizing "{self.text_min}"' print(f'Imagining "{self.text_path}" {penalizing}...') with torch.no_grad(): self.model(self.encoded_texts["max"][0]) # one warmup step due to issue with CLIP and CUDA if self.open_folder: open_folder('./') self.open_folder = False image_pbar = tqdm(total=self.total_image_updates, desc='image update', position=2, leave=True) epoch_pbar = trange(self.epochs, desc = ' epochs', position=0, leave=True) for epoch in (ep for ep in epoch_pbar if not terminate): pbar = trange(self.iterations, desc=' iteration', position=1, leave=True) image_pbar.update(0) for i in (it for it in pbar if not terminate): out, loss = self.train_step(epoch, i, image_pbar) pbar.set_description(f'loss: {loss.item():04.2f}')
_MODELS = { "RN50": "https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt", "RN101": "https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt", "RN50x4": "https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt", "ViT-B/32": "https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt", "ViT-L/14": "https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt" } def _download(url: str, root: str = os.path.expanduser("~/.cache/clip")): os.makedirs(root, exist_ok=True) filename = os.path.basename(url) expected_sha256 = url.split("/")[-2] download_target = os.path.join(root, filename) if os.path.exists(download_target) and not os.path.isfile(download_target): raise RuntimeError(f"{download_target} exists and is not a regular file") if os.path.isfile(download_target): if hashlib.sha256(open(download_target, "rb").read()).hexdigest() == expected_sha256: return download_target else: warnings.warn(f"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file") with urllib.request.urlopen(url) as source, open(download_target, "wb") as output: with tqdm(total=int(source.info().get("Content-Length")), ncols=80, unit='iB', unit_scale=True) as loop: while True: buffer = source.read(8192) if not buffer: break output.write(buffer) loop.update(len(buffer)) if hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256: raise RuntimeError(f"Model has been downloaded but the SHA256 checksum does not not match") return download_target def _transform(): return Compose([ Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)), ]) def available_models() -> List[str]: """Returns the names of available CLIP models""" return list(_MODELS.keys()) def load(name: str, device: Union[str, torch.device] = "cuda" if torch.cuda.is_available() else "cpu", jit=True): """Load a CLIP model Parameters ---------- name : str A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict device : Union[str, torch.device] The device to put the loaded model jit : bool Whether to load the optimized JIT model (default) or more hackable non-JIT model. Returns ------- model : torch.nn.Module The CLIP model preprocess : Callable[[PIL.Image], torch.Tensor] A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input """ if name in _MODELS: model_path = _download(_MODELS[name]) elif os.path.isfile(name): model_path = name else: raise RuntimeError(f"Model {name} not found; available models = {available_models()}") try: # loading JIT archive model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval() state_dict = None except RuntimeError: # loading saved state dict if jit: warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead") jit = False state_dict = torch.load(model_path, map_location="cpu") if not jit: model = build_model(state_dict or model.state_dict()).to(device) if str(device) == "cpu": model.float() return model, _transform() # patch the device names device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]) device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1] def patch_device(module): graphs = [module.graph] if hasattr(module, "graph") else [] if hasattr(module, "forward1"): graphs.append(module.forward1.graph) for graph in graphs: for node in graph.findAllNodes("prim::Constant"): if "value" in node.attributeNames() and str(node["value"]).startswith("cuda"): node.copyAttributes(device_node) model.apply(patch_device) patch_device(model.encode_image) patch_device(model.encode_text) # patch dtype to float32 on CPU if str(device) == "cpu": float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[]) float_input = list(float_holder.graph.findNode("aten::to").inputs())[1] float_node = float_input.node() def patch_float(module): graphs = [module.graph] if hasattr(module, "graph") else [] if hasattr(module, "forward1"): graphs.append(module.forward1.graph) for graph in graphs: for node in graph.findAllNodes("aten::to"): inputs = list(node.inputs()) for i in [1, 2]: # dtype can be the second or third argument to aten::to() if inputs[i].node()["value"] == 5: inputs[i].node().copyAttributes(float_node) model.apply(patch_float) patch_float(model.encode_image) patch_float(model.encode_text) model.float() return model, _transform() def tokenize(texts: Union[str, List[str]], context_length: int = 77) -> torch.LongTensor: """ Returns the tokenized representation of given input string(s) Parameters ---------- texts : Union[str, List[str]] An input string or a list of input strings to tokenize context_length : int The context length to use; all CLIP models use 77 as the context length Returns ------- A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length] """ if isinstance(texts, str): texts = [texts] sot_token = _tokenizer.encoder["<|startoftext|>"] eot_token = _tokenizer.encoder["<|endoftext|>"] all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts] result = torch.zeros(len(all_tokens), context_length, dtype=torch.long) for i, tokens in enumerate(all_tokens): if len(tokens) > context_length: raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}") result[i, :len(tokens)] = torch.tensor(tokens) return result class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1): super().__init__() # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1 self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity() self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = None self.stride = stride if stride > 1 or inplanes != planes * Bottleneck.expansion: # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1 self.downsample = nn.Sequential(OrderedDict([ ("-1", nn.AvgPool2d(stride)), ("0", nn.Conv2d(inplanes, planes * self.expansion, 1, stride=1, bias=False)), ("1", nn.BatchNorm2d(planes * self.expansion)) ])) def forward(self, x: torch.Tensor): identity = x out = self.relu(self.bn1(self.conv1(x))) out = self.relu(self.bn2(self.conv2(out))) out = self.avgpool(out) out = self.bn3(self.conv3(out)) if self.downsample is not None: identity = self.downsample(x) out += identity out = self.relu(out) return out class AttentionPool2d(nn.Module): def __init__(self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None): super().__init__() self.positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2 + 1, embed_dim) / embed_dim ** 0.5) self.k_proj = nn.Linear(embed_dim, embed_dim) self.q_proj = nn.Linear(embed_dim, embed_dim) self.v_proj = nn.Linear(embed_dim, embed_dim) self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim) self.num_heads = num_heads def forward(self, x): x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1) # NCHW -> (HW)NC x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC x, _ = F.multi_head_attention_forward( query=x, key=x, value=x, embed_dim_to_check=x.shape[-1], num_heads=self.num_heads, q_proj_weight=self.q_proj.weight, k_proj_weight=self.k_proj.weight, v_proj_weight=self.v_proj.weight, in_proj_weight=None, in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]), bias_k=None, bias_v=None, add_zero_attn=False, dropout_p=0, out_proj_weight=self.c_proj.weight, out_proj_bias=self.c_proj.bias, use_separate_proj_weight=True, training=self.training, need_weights=False ) return x[0] class ModifiedResNet(nn.Module): """ A ResNet class that is similar to torchvision's but contains the following changes: - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool. - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1 - The final pooling layer is a QKV attention instead of an average pool """ def __init__(self, layers, output_dim, heads, input_resolution=224, width=64): super().__init__() self.output_dim = output_dim self.input_resolution = input_resolution # the 3-layer stem self.conv1 = nn.Conv2d(3, width // 2, kernel_size=3, stride=2, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(width // 2) self.conv2 = nn.Conv2d(width // 2, width // 2, kernel_size=3, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(width // 2) self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False) self.bn3 = nn.BatchNorm2d(width) self.avgpool = nn.AvgPool2d(2) self.relu = nn.ReLU(inplace=True) # residual layers self._inplanes = width # this is a *mutable* variable used during construction self.layer1 = self._make_layer(width, layers[0]) self.layer2 = self._make_layer(width * 2, layers[1], stride=2) self.layer3 = self._make_layer(width * 4, layers[2], stride=2) self.layer4 = self._make_layer(width * 8, layers[3], stride=2) embed_dim = width * 32 # the ResNet feature dimension self.attnpool = AttentionPool2d(input_resolution // 32, embed_dim, heads, output_dim) def _make_layer(self, planes, blocks, stride=1): layers = [Bottleneck(self._inplanes, planes, stride)] self._inplanes = planes * Bottleneck.expansion for _ in range(1, blocks): layers.append(Bottleneck(self._inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): def stem(x): for conv, bn in [(self.conv1, self.bn1), (self.conv2, self.bn2), (self.conv3, self.bn3)]: x = self.relu(bn(conv(x))) x = self.avgpool(x) return x x = x.type(self.conv1.weight.dtype) x = stem(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.attnpool(x) return x class LayerNorm(nn.LayerNorm): """Subclass torch's LayerNorm to handle fp16.""" def forward(self, x: torch.Tensor): orig_type = x.dtype ret = super().forward(x.type(torch.float32)) return ret.type(orig_type) class QuickGELU(nn.Module): def forward(self, x: torch.Tensor): return x * torch.sigmoid(1.702 * x) class ResidualAttentionBlock(nn.Module): def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None): super().__init__() self.attn = nn.MultiheadAttention(d_model, n_head) self.ln_1 = LayerNorm(d_model) self.mlp = nn.Sequential(OrderedDict([ ("c_fc", nn.Linear(d_model, d_model * 4)), ("gelu", QuickGELU()), ("c_proj", nn.Linear(d_model * 4, d_model)) ])) self.ln_2 = LayerNorm(d_model) self.attn_mask = attn_mask def attention(self, x: torch.Tensor): self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0] def forward(self, x: torch.Tensor): x = x + self.attention(self.ln_1(x)) x = x + self.mlp(self.ln_2(x)) return x class Transformer(nn.Module): def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None): super().__init__() self.width = width self.layers = layers self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)]) def forward(self, x: torch.Tensor): return self.resblocks(x) class VisualTransformer(nn.Module): def __init__(self, input_resolution: int, patch_size: int, width: int, layers: int, heads: int, output_dim: int): super().__init__() self.input_resolution = input_resolution self.output_dim = output_dim self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False) scale = width ** -0.5 self.class_embedding = nn.Parameter(scale * torch.randn(width)) self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width)) self.ln_pre = LayerNorm(width) self.transformer = Transformer(width, layers, heads) self.ln_post = LayerNorm(width) self.proj = nn.Parameter(scale * torch.randn(width, output_dim)) def forward(self, x: torch.Tensor): x = self.conv1(x) # shape = [*, width, grid, grid] x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2] x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width] x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1) # shape = [*, grid ** 2 + 1, width] x = x + self.positional_embedding.to(x.dtype) x = self.ln_pre(x) x = x.permute(1, 0, 2) # NLD -> LND x = self.transformer(x) x = x.permute(1, 0, 2) # LND -> NLD x = self.ln_post(x[:, 0, :]) if self.proj is not None: x = x @ self.proj return x class CLIP(nn.Module): def __init__(self, embed_dim: int, # vision image_resolution: int, vision_layers: Union[Tuple[int, int, int, int], int], vision_width: int, vision_patch_size: int, # text context_length: int, vocab_size: int, transformer_width: int, transformer_heads: int, transformer_layers: int ): super().__init__() self.context_length = context_length if isinstance(vision_layers, (tuple, list)): vision_heads = vision_width * 32 // 64 self.visual = ModifiedResNet( layers=vision_layers, output_dim=embed_dim, heads=vision_heads, input_resolution=image_resolution, width=vision_width ) else: vision_heads = vision_width // 64 self.visual = VisualTransformer( input_resolution=image_resolution, patch_size=vision_patch_size, width=vision_width, layers=vision_layers, heads=vision_heads, output_dim=embed_dim ) self.transformer = Transformer( width=transformer_width, layers=transformer_layers, heads=transformer_heads, attn_mask=self.build_attention_mask() ) self.vocab_size = vocab_size self.token_embedding = nn.Embedding(vocab_size, transformer_width) self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width)) self.ln_final = LayerNorm(transformer_width) self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim)) self.logit_scale = nn.Parameter(torch.ones([])) self.initialize_parameters() def initialize_parameters(self): nn.init.normal_(self.token_embedding.weight, std=0.02) nn.init.normal_(self.positional_embedding, std=0.01) if isinstance(self.visual, ModifiedResNet): if self.visual.attnpool is not None: std = self.visual.attnpool.c_proj.in_features ** -0.5 nn.init.normal_(self.visual.attnpool.q_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.k_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.v_proj.weight, std=std) nn.init.normal_(self.visual.attnpool.c_proj.weight, std=std) for resnet_block in [self.visual.layer1, self.visual.layer2, self.visual.layer3, self.visual.layer4]: for name, param in resnet_block.named_parameters(): if name.endswith("bn3.weight"): nn.init.zeros_(param) proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5) attn_std = self.transformer.width ** -0.5 fc_std = (2 * self.transformer.width) ** -0.5 for block in self.transformer.resblocks: nn.init.normal_(block.attn.in_proj_weight, std=attn_std) nn.init.normal_(block.attn.out_proj.weight, std=proj_std) nn.init.normal_(block.mlp.c_fc.weight, std=fc_std) nn.init.normal_(block.mlp.c_proj.weight, std=proj_std) if self.text_projection is not None: nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5) def build_attention_mask(self): # lazily create causal attention mask, with full attention between the vision tokens # pytorch uses additive attention mask; fill with -inf mask = torch.empty(self.context_length, self.context_length) mask.fill_(float("-inf")) mask.triu_(1) # zero out the lower diagonal return mask @property def dtype(self): return self.visual.conv1.weight.dtype def encode_image(self, image): return self.visual(image.type(self.dtype)) def encode_text(self, text): x = self.token_embedding(text).type(self.dtype) # [batch_size, n_ctx, d_model] x = x + self.positional_embedding.type(self.dtype) x = x.permute(1, 0, 2) # NLD -> LND x = self.transformer(x) x = x.permute(1, 0, 2) # LND -> NLD x = self.ln_final(x).type(self.dtype) # x.shape = [batch_size, n_ctx, transformer.width] # take features from the eot embedding (eot_token is the highest number in each sequence) x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection return x def forward(self, image, text): image_features = self.encode_image(image) text_features = self.encode_text(text) # normalized features image_features = image_features / image_features.norm(dim=-1, keepdim=True) text_features = text_features / text_features.norm(dim=-1, keepdim=True) # cosine similarity as logits logit_scale = self.logit_scale.exp() logits_per_image = logit_scale * image_features @ text_features.t() logits_per_text = logit_scale * text_features @ image_features.t() # shape = [global_batch_size, global_batch_size] return logits_per_image, logits_per_text def convert_weights(model: nn.Module): """Convert applicable model parameters to fp16""" def _convert_weights_to_fp16(l): if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)): l.weight.data = l.weight.data.half() if l.bias is not None: l.bias.data = l.bias.data.half() if isinstance(l, nn.MultiheadAttention): for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]: tensor = getattr(l, attr) if tensor is not None: tensor.data = tensor.data.half() for name in ["text_projection", "proj"]: if hasattr(l, name): attr = getattr(l, name) if attr is not None: attr.data = attr.data.half() model.apply(_convert_weights_to_fp16) def build_model(state_dict: dict): vit = "visual.proj" in state_dict if vit: vision_width = state_dict["visual.conv1.weight"].shape[0] vision_layers = len([k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")]) vision_patch_size = state_dict["visual.conv1.weight"].shape[-1] grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5) image_resolution = vision_patch_size * grid_size else: counts: list = [len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]] vision_layers = tuple(counts) vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0] output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5) vision_patch_size = None assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0] image_resolution = output_width * 32 embed_dim = state_dict["text_projection"].shape[1] context_length = state_dict["positional_embedding"].shape[0] vocab_size = state_dict["token_embedding.weight"].shape[0] transformer_width = state_dict["ln_final.weight"].shape[0] transformer_heads = transformer_width // 64 transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks"))) model = CLIP( embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size, context_length, vocab_size, transformer_width, transformer_heads, transformer_layers ) for key in ["input_resolution", "context_length", "vocab_size"]: if key in state_dict: del state_dict[key] convert_weights(model) model.load_state_dict(state_dict) return model.eval() @lru_cache() def default_bpe(): return os.path.join(os.path.dirname(os.path.abspath(__file__)), "data/bpe_simple_vocab_16e6.txt") @lru_cache() def bytes_to_unicode(): """ Returns list of utf-8 byte and a corresponding list of unicode strings. The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. This is a signficant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup tables between utf-8 bytes and unicode strings. And avoids mapping to whitespace/control characters the bpe code barfs on. """ bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) cs = bs[:] n = 0 for b in range(2**8): if b not in bs: bs.append(b) cs.append(2**8+n) n += 1 cs = [chr(n) for n in cs] return dict(zip(bs, cs)) def get_pairs(word): """Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings). """ pairs = set() prev_char = word[0] for char in word[1:]: pairs.add((prev_char, char)) prev_char = char return pairs def basic_clean(text): text = ftfy.fix_text(text) text = html.unescape(html.unescape(text)) return text.strip() def whitespace_clean(text): text = re.sub(r'\s+', ' ', text) text = text.strip() return text class SimpleTokenizer(object): def __init__(self, bpe_path: str = default_bpe()): self.byte_encoder = bytes_to_unicode() self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} merges = Path(bpe_path).read_text(encoding='utf8').split('\n') merges = merges[1:49152-256-2+1] merges = [tuple(merge.split()) for merge in merges] vocab = list(bytes_to_unicode().values()) vocab = vocab + [v+'</w>' for v in vocab] for merge in merges: vocab.append(''.join(merge)) vocab.extend(['<|startoftext|>', '<|endoftext|>']) self.encoder = dict(zip(vocab, range(len(vocab)))) self.decoder = {v: k for k, v in self.encoder.items()} self.bpe_ranks = dict(zip(merges, range(len(merges)))) self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'} self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE) def bpe(self, token): if token in self.cache: return self.cache[token] word = tuple(token[:-1]) + ( token[-1] + '</w>',) pairs = get_pairs(word) if not pairs: return token+'</w>' while True: bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf'))) if bigram not in self.bpe_ranks: break first, second = bigram new_word = [] i = 0 while i < len(word): try: j = word.index(first, i) new_word.extend(word[i:j]) i = j except: new_word.extend(word[i:]) break if word[i] == first and i < len(word)-1 and word[i+1] == second: new_word.append(first+second) i += 2 else: new_word.append(word[i]) i += 1 new_word = tuple(new_word) word = new_word if len(word) == 1: break else: pairs = get_pairs(word) word = ' '.join(word) self.cache[token] = word return word def encode(self, text): bpe_tokens = [] text = whitespace_clean(basic_clean(text)).lower() for token in re.findall(self.pat, text): token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8')) bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' ')) return bpe_tokens def decode(self, tokens): text = ''.join([self.decoder[token] for token in tokens]) text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ') return text import gzip _tokenizer = SimpleTokenizer()
class AxialPositionalEmbedding(nn.Module): def __init__(self, dim, axial_shape, axial_dims = None): super().__init__() self.dim = dim self.shape = axial_shape self.max_seq_len = reduce(mul, axial_shape, 1) self.summed = axial_dims is None axial_dims = ((dim,) * len(axial_shape)) if self.summed else axial_dims assert len(self.shape) == len(axial_dims), 'number of axial dimensions must equal the number of dimensions in the shape' assert self.summed or not self.summed and sum(axial_dims) == dim, f'axial dimensions must sum up to the target dimension {dim}' self.weights = ParameterList(self, 'weights', len(axial_shape)) for ind, (shape, axial_dim) in enumerate(zip(self.shape, axial_dims)): ax_shape = [1] * len(self.shape) ax_shape[ind] = shape ax_shape = (1, *ax_shape, axial_dim) ax_emb = nn.Parameter(torch.zeros(ax_shape).normal_(0, 1)) self.weights.append(ax_emb) def forward(self, x): b, t, e = x.shape assert (t <= self.max_seq_len), f'Sequence length ({t}) must be less than the maximum sequence length allowed ({self.max_seq_len})' embs = [] for ax_emb in self.weights.to_list(): axial_dim = ax_emb.shape[-1] expand_shape = (b, *self.shape, axial_dim) emb = ax_emb.expand(expand_shape).reshape(b, self.max_seq_len, axial_dim) embs.append(emb) pos_emb = sum(embs) if self.summed else torch.cat(embs, dim=-1) return pos_emb[:, :t].to(x) # a mock parameter list object until below issue is resolved # https://github.com/pytorch/pytorch/issues/36035 class ParameterList(object): def __init__(self, kls, prefix, length): self.ind = 0 self.kls = kls self.prefix = prefix self.length = length def _keyname(self, prefix, ind): return f'{prefix}_{ind}' def append(self, x): setattr(self.kls, self._keyname(self.prefix, self.ind), x) self.ind += 1 def to_list(self): return [getattr(self.kls, self._keyname(self.prefix, i)) for i in range(self.length)] # Axial Positional Embedding for Images class AxialPositionalEmbeddingImage(nn.Module): def __init__(self, dim, axial_shape, axial_dims = None): super().__init__() assert len(axial_shape) == 2, 'Axial shape must have 2 dimensions for images' self.pos_emb = AxialPositionalEmbedding(dim, axial_shape, axial_dims) def forward(self, img): b, c, h, w = img.shape img = img.permute(0, 2, 3, 1).reshape(b, h * w, c) pos_emb = self.pos_emb(img) return pos_emb.reshape(b, h, w, c).permute(0, 3, 1, 2)
# constants DEVICE = None # defaults to cuda if available, else cpu NUM_BATCHES = int(1e5) GRADIENT_ACCUMULATE_EVERY = 16 LEARNING_RATE = 3e-4 IGNORE_INDEX = -100 THRESHOLD_LENGTH = 250 # set device DISTOGRAM_BUCKETS = constants.DISTOGRAM_BUCKETS DEVICE = constants.DEVICE # helpers def cycle(loader, cond = lambda x: True): while True: for data in loader: if not cond(data): continue yield data # get data data = scn.load( casp_version = 12, thinning = 30, with_pytorch = 'dataloaders', batch_size = 1, dynamic_batching = False ) data = iter(data['train']) data_cond = lambda t: t[1].shape[1] < THRESHOLD_LENGTH dl = cycle(data, data_cond) # model model = Alphafold2( dim = 256, depth = 1, heads = 8, dim_head = 64 ).to(DEVICE) # optimizer optim = Adam(model.parameters(), lr = LEARNING_RATE) # training loop for _ in range(NUM_BATCHES): for _ in range(GRADIENT_ACCUMULATE_EVERY): batch = next(dl) seq, coords, mask = batch.seqs, batch.crds, batch.msks b, l, _ = seq.shape # prepare mask, labels seq, coords, mask = seq.argmax(dim = -1).to(DEVICE), coords.to(DEVICE), mask.to(DEVICE).bool() coords = rearrange(coords, 'b (l c) d -> b l c d', l = l) discretized_distances = get_bucketed_distance_matrix(coords[:, :, 1], mask, DISTOGRAM_BUCKETS, IGNORE_INDEX) # predict distogram = model(seq, mask = mask) distogram = rearrange(distogram, 'b i j c -> b c i j') # loss loss = F.cross_entropy( distogram, discretized_distances, ignore_index = IGNORE_INDEX ) loss.backward() print('loss:', loss.item()) optim.step() optim.zero_grad()